A Governed Multi-Node Computational Discovery Pipeline: Pre-Specified Gates, Bounded Claims, and Honest Branch Closure in Phytochemical Polypharmacology

A Governed Multi-Node Computational Discovery Pipeline: Pre-Specified Gates, Bounded Claims, and Honest Branch Closure in Phytochemical Polypharmacology | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Short Report A Governed Multi-Node Computational Discovery Pipeline: Pre-Specified Gates, Bounded Claims, and Honest Branch Closure in Phytochemical Polypharmacology faycal ferhat This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9587405/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Computational multi-target profiling is technically accessible but methodologically ungoverned: thresholds set post-hoc, claims exceeding the data, failed branches omitted, and ΔG values untraceable to source files. We describe a platform that addresses these failure modes by construction. The platform enforces pre-specified gates (frozen in named decision sheets before any simulation runs), outcome-keyed ceiling templates (format_verdict.py, NemoClaw constitutional review), a 46-check regression oracle, and a SHA256 provenance chain linking every published number to the ΔTOTAL line of the source FINAL_RESULTS_DECOMP.dat file. All closed cycles feed a retrieval-first knowledge graph. Applied to nine fig and olive phytochemicals across two targets: (Node 1) nine compounds versus COX-2 (5KIR), N = 4 replicates each, ALL_PASS; oleocanthal ΔG = − 61.8 kcal/mol, CV = 2.4%, ΔΔG = − 22.9 kcal/mol versus celecoxib. (Node 2) luteolin versus IKKβ (4KIK): initial CV = 16.1% (FAIL); structural diagnosis and pre-registered deep-pocket reseed produced ALL_PASS (mean ΔG = − 31.457 kcal/mol, CV = 6.85%, N = 3). (Node 3) quercetin versus IKKβ: generalization, ALL_PASS (mean ΔG = − 24.603 kcal/mol, CV = 3.04%, N = 3). (Node 4) quercetin versus p65 (1VKX): POSE_UNSTABLE, branch closed; platform claim bounded to two-node model. All results are EXPLORATORY DDT tier: computationally reproducible under the stated protocol, not biologically validated. Appropriate next steps are enzymatic binding assays, kinase selectivity profiling, and safety characterization. Frozen artifact set (28 SHA256-verified files): https://doi.org/10.5281/zenodo.19076569 . Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction: The Problem with Ungoverned Computational Polypharmacology Computational screening of phytochemicals against multiple molecular targets has become technically accessible. GPU-accelerated molecular dynamics, open force fields (CHARMM36 [ 1 ], GAFF2 [ 2 ]), and publicly available structural data (PDB [ 3 ], AlphaFold [ 4 ]) allow a single workstation to run multi-target binding free energy estimates in days. This technical accessibility has outpaced the methodological infrastructure needed to interpret the results. The rational-polypharmacology framing has been articulated previously [ 5 , 6 ], but the governance layer needed to make multi-target computational claims auditable has lagged behind the raw compute capacity. The most common failure mode is not technical. It is interpretive: a computation finishes, produces a ΔG value, and the ΔG value is reported as evidence of binding. The claim is calibrated not to the evidence but to the desired conclusion. Five specific patterns recur in the literature of computational polypharmacology and are worth naming directly. First, the threshold problem: gate criteria (ΔG cutoff, RMSD stability threshold, replicate CV limit) are stated in the methods section but set after the results are seen. A CV of 16.1% is described as “modest variability” when the result is favorable; the same value would be described as a failure if a different threshold had been locked before the run. The threshold is not a gate — it is a post-hoc frame. Second, the ceiling problem: computational results support specific claims about computational observables. “Luteolin shows a mean MM-GBSA ΔG of − 31 kcal/mol under the stated protocol” is a claim the data support. “Luteolin inhibits IKKβ” is not. The distance between those two statements is the distance between what the experiment measures and what the claim asserts. That distance is routinely collapsed. The underlying limitations of end-point MM-GBSA scoring are well-characterised [ 7 ]. Third, the branch-closure problem: computational screens generate both positive and negative results. Positive results appear in papers. Negative results are silently absent. A paper that reports two confirmed nodes and omits the two failed branches presents a biased picture of the evidence. A reader cannot assess the false discovery rate of the pipeline from the published results alone. Fourth, the provenance problem: a ΔG value in a manuscript should be traceable to the raw output file it came from. In practice, values are often transferred through summary spreadsheets, re-keyed, rounded, or averaged by rules not stated in the methods. The number in the paper may not match the number in the output file. Fifth, the single-trajectory problem: a single 5 ns MD run produces a single ΔG estimate. That estimate has a variance that is not visible from the run itself. Without independent velocity-seed replicates and a pre-specified convergence criterion, the estimate cannot be called reproducible — it can only be called observed. The value of independent replicates for obtaining consistent binding-affinity predictions has been demonstrated directly in the alchemical-free-energy literature [ 22 ]. This paper describes a platform that addresses all five problems. It does not address them by stricter statistical methods alone, but by building governance into the experimental pipeline: pre-specified gates, outcome-keyed ceiling templates, explicit branch closure, SHA256 provenance chains, and a retrieval-first knowledge graph that requires every claim to be grounded in a queryable evidence record. The platform was applied to nine quantified fig and olive phytochemicals (eleven docked; two excluded by force-field failure before MD) across two molecular targets: the COX-2 cyclooxygenase (Node 1) [ 9 ] and the IKKβ kinase ATP pocket (Node 2, with a generalization test at Node 3) [ 11 ]. Four experimental branches were run at the MD stage. One failed on first attempt and was rescued by a pre-registered deep-pocket reseed (Node 2). One generalized the Node 2 result to a second flavonoid (Node 3). One failed and was closed without any upward claim revision (Node 4). These outcomes are reported in the order they occurred, with the negative results held to the same evidential standard as the positive ones. The scientific contribution of this work is the platform, not the compound ranking. The compound results are evidence that the platform operates correctly — including when it records failure. 2. Platform Architecture 2.1 Overview The platform runs four sequential stages for each experimental branch: (1) structure preparation, docking, and MD simulation; (2) MM-GBSA binding free energy estimation; (3) gate evaluation against pre-specified criteria; and (4) verdict formatting, ceiling enforcement, and knowledge graph ingestion. Each stage produces auditable output. No stage consumes data from the preceding stage without a documented provenance link. The overall pipeline is shown in Fig. 1 . 2.2 Pre-specified gates For each experimental branch, a decision sheet is created and frozen before any simulation is run. The decision sheet specifies four gate criteria: ligand RMSD stability ( 80% of the equilibrated trajectory in ≥ 3/N seeds), pocket contact occupancy (≥ 5/9 pre-specified pocket residues at ≥ 50% occupancy in ≥ 3/N seeds), MM-GBSA direction (mean ΔG < − 10 kcal/mol), and MM-GBSA reproducibility (coefficient of variation < 15% across N ≥ 3 converged replicates). The threshold values and the outcome key for each possible gate combination are written into the sheet before results arrive. The gate evaluator (gate_evaluate.py) reads raw contact occupancy files and FINAL_RESULTS_DECOMP.dat output from gmx_MMPBSA [ 8 ], applies the pre-specified criteria, and emits an outcome_key from a fixed vocabulary: ALL_PASS, CV_FAIL_ONLY, or POSE_UNSTABLE. The outcome_key is written to gate_verdict.json and is the sole input to the verdict formatter. The gate evaluator contains no logic that adjusts criteria based on results; it reads the pre-specified thresholds from the decision sheet verbatim. A complete summary of gate outcomes across all four experimental branches is presented in Table 1 . Table 1 Gate evaluation summary across all four experimental branches. Gate criteria pre-specified in named decision sheets before any simulation ran. Thresholds: Gate 1 LIG RMSD 80% stable in ≥ 2/3 (or ≥ 3/4) seeds; Gate 2 ≥ 5/9 pocket residues ≥ 50% occupancy in ≥ 2/3 seeds; Gate 3 mean ΔG < − 10 kcal/mol; Gate 4 CV < 15%. Node / Run Gate 1 (RMSD) Gate 2 (Contacts) Gate 3 (ΔG dir.) Gate 4 (CV) outcome_key Node 1 — COX-2 panel (N = 4) PASS (all) PASS (all) PASS (all < − 10) PASS (all < 15%) ALL_PASS Node 2 initial — Luteolin IKKβ (N = 4) PASS (3/4) PASS (4/4) PASS (− 28.18) FAIL (16.1%) CV_FAIL_ONLY Node 2 reseed — Luteolin IKKβ dp (N = 3) PASS (3/3) PASS (3/3) PASS (− 31.457) PASS (6.85%) ALL_PASS Node 3 — Quercetin IKKβ (N = 3) PASS (2/3)ᵃ PASS (3/3) PASS (− 24.603) PASS (3.04%) ALL_PASS Node 4 — Quercetin p65 (N = 3) FAIL (0/3) FAIL (0/3) PASS (− 12.357) FAIL (55.85%) POSE_UNSTABLE ᵃ s2222 borderline (34.4% stable); gate passes at 2/3 threshold. 2.3 Claim ceiling enforcement Each outcome_key maps to a verdict template in verdict_templates.json (and per-node variants for Nodes 3–4). The template specifies the exact language permitted for the corresponding outcome: numerical slots filled from gate_verdict.json, ceiling phrase drawn from a fixed list, and a NOT SAFE list of forbidden phrases. The formatter (format_verdict.py) populates the template deterministically. It cannot paraphrase ceiling language, round values, or omit required fields. Each template enforces a four-column structure: what was observed (ΔG, CV, gate outcome), what may be claimed (the pre-specified ceiling phrase), what may not be claimed (the NOT SAFE list), and what condition would permit an upgrade (the unlock gate). This table is the operative output of the pipeline — not the ΔG value alone. A constitutional review (NemoClaw) scans every formatted verdict for language on the forbidden list before the verdict is filed. The forbidden list includes phrases such as “confirms binding,” “validates binding,” “strong binder,” and “Boltz confirms,” which assert conclusions the computational data do not support. If a forbidden phrase is detected, the verdict is rejected and the error is logged before any archiving occurs. The ceiling templates and the forbidden-phrase list are frozen artifacts (chmod 444, SHA256-locked in FREEZE_MANIFEST.json). They cannot be modified by the formatter, the knowledge graph ingestion pipeline, or the self-improving training loop without a deliberate human action on the file system. 2.4 Oracle and SHA256 provenance A regression oracle (verify_node1.py, 46 checks) locks the Node 1 result set. The oracle verifies that every ΔG value in the paper matches the ΔTOTAL line in the corresponding FINAL_RESULTS_DECOMP.dat file, that every SHA256 hash in the training pool matches the file on disk, and that no governance file has been modified since its freeze date. The oracle is run before any claim is published and after any change to the artifact set. Failure on any of the 46 checks halts the pipeline. The SHA256 chain runs from raw output file to published claim. For each training cycle, verify_training_pool.py extracts ΔTOTAL values directly from FINAL_RESULTS_DECOMP.dat using a regex that matches the Unicode Δ character in the output format, recomputes the mean and CV from scratch, and compares the recomputed values to those stored in gate_verdict.json. Discrepancies fail the cycle-admission check. No value enters the training corpus or the knowledge graph without passing this extraction-and-recompute verification step. 2.5 Retrieval-first knowledge graph All compound results, gate verdicts, claim ceilings, and literature evidence are ingested as structured triples into a knowledge graph (kg_data/kg_export.json). The agent operating the pipeline follows a retrieval-first behavioral rule: it must query the knowledge graph before answering any question about a compound, target, or ΔG value. Memory-only responses are not permitted for questions that can be grounded in the graph. This rule prevents the agent from reporting a value from a prior conversation that may have since been corrected by a raw-data audit. The knowledge graph is cross-referenced against the primary literature via PubMed. Ten claims about IKKβ flavonoid binding were verified against PubMed search results (crosscheck_report.json); all ten were supported by at least one citable source [ 16 , 19 ]. Literature evidence for each compound–target pair is stored as a corpus document (kg_data/corpus/) and cited in the relevant decision sheet. 2.6 Self-improving verdict formatter The platform includes a training loop for the verdict formatter component. Closed experimental cycles contribute (gate_verdict.json, verdict_section3.md) instruction pairs to a training pool. Cycle admission requires SHA256-verified raw data, NemoClaw constitutional review of the ground-truth output, and confirmation that the output does not contain language from the forbidden-phrase list. The training pool currently holds three verified cycles. A baseline evaluation of the current production model (qwen3-coder-next via Ollama) against the three-cycle corpus showed 0/3 cycles pass the ceiling compliance and numerical accuracy checks. A LoRA fine-tune [ 24 ] (Qwen/Qwen2.5-7B [ 25 ], rank = 16, 3 epochs, MAX_SEQ_LEN = 4096) was run against the corpus. The adapter produces structurally correct outputs — correct per-seed ΔG values, no fabricated numbers, reproduced SHA256 provenance — but does not yet emit the required ceiling phrase or markdown template format, reflecting the insufficient size of the current training corpus (three examples cannot anchor a specific phrase pattern in a 7B model). A human promotion gate prevents any adapter from replacing the production formatter until eval_report.json shows overall_pass = True across all cycles. The training loop is an architectural component, not a deployed production system. Its inclusion in this paper documents the design pattern, not a completed training result. 3. Results 3.1 Node 1: COX-2 Binding Free Energy Panel 3.1.1 Protocol and replicate structure Nine fig and olive phytochemicals were screened against the COX-2 crystal structure (PDB: 5KIR; rofecoxib-bound complex [ 9 , 10 ]) using GROMACS 2026 [ 2 ] with the CHARMM36 protein force field [ 1 ] and GAFF2 ligand parameters [ 2 ]. Each compound was run as N = 4 independent velocity-seed replicates (seeds: main, s1234, s42, s9999), each for 5 ns production MD following energy minimization, 100 ps NVT, and 100 ps NPT equilibration. MM-GBSA binding free energies were computed using gmx_MMPBSA [ 8 ] with igb = 2 and saltcon = 0.150 M. Celecoxib [ 10 ] was included as a clinical COX-2 inhibitor reference, docked and evaluated under identical protocol; it is not the co-crystallized ligand of the 5KIR template. Two compounds were excluded prior to analysis: oleuropein (+ 310 kcal/mol) and rutin (+ 576 kcal/mol) both produced unphysical positive ΔG values attributable to GAFF2 c6 cross-term failure for glycosidic linkages. This force-field limitation is documented; the compounds are carried as explicit exclusions rather than quietly omitted. 3.1.2 Results All ΔG values are mean ± SD across N = 4 replicates, sourced from the ΔTOTAL line of FINAL_RESULTS_DECOMP.dat for each run (SHA256-verified, archived in md_runs/PAPER_RESULTS_FINAL.txt). The full ranked panel is presented in Table 2 and visualized in Fig. 2. Oleocanthal: ΔG = − 61.765 ± 1.477 kcal/mol (CV = 2.4%, N = 4; source: md_runs/Oleocanthal_COX2_v2/FINAL_RESULTS_DECOMP.dat and three replicates). Against the celecoxib reference (ΔG = − 38.875 ± 1.199 kcal/mol), the differential is ΔΔG = − 22.890 kcal/mol (Welch t = − 24.1, p < < 0.001). A CV of 2.4% is consistent with convergence across independent velocity seeds under this protocol. The oleocanthal ΔG is driven primarily by van der Waals interactions (ΔVDW = − 40.2 kcal/mol, ΔEEL = − 14.6 kcal/mol, N = 4 aggregate from FINAL_RESULTS_DECOMP.dat), consistent with deep insertion of the aliphatic chain and dialdehyde moiety into the hydrophobic channel, consistent with prior reports of oleocanthal anti-inflammatory activity and olive-polyphenol COX-2 engagement [ 12 , 13 , 14 , 15 ]. The aggregate electrostatic contribution (ΔEEL = − 14.6 kcal/mol) is consistent with known COX-2 channel contacts; per-residue decomposition was not performed; however, the electrostatic contribution (ΔEEL = − 14.6 kcal/mol) is moderate relative to the dominant van der Waals term (ΔVDW = − 40.2 kcal/mol), arguing against a purely electrostatic artifact. MM-GBSA absolute values are known to overestimate binding strength for some chemotypes [ 7 ]; the oleocanthal figure should be interpreted as a within-study comparative rank rather than a physical free energy equivalent to experimental IC₅₀. Figure 2. MM-GBSA binding free energies for the nine-compound COX-2 panel (PDB 5KIR, rofecoxib-bound complex) plus celecoxib reference. Values are mean ± SD across N = 4 independent velocity-seed replicates; all sourced from the ΔTOTAL line of FINAL_RESULTS_DECOMP.dat (SHA256-verified, oracle PASS, verify_node1.py, 2026-03-24). Celecoxib was evaluated as an external reference compound under identical protocol; dashed line = celecoxib reference (− 38.875 kcal/mol). Oleuropein and rutin excluded (GAFF2 c6 cross-term failure; documented). Source file: manuscript/figures/Fig. 2_mmpbsa_barplot.png. [Figure 2 is supplied separately by the author and placed here.] Ligstroside D2, the second olive compound, produced ΔG = − 32.420 ± 2.082 kcal/mol (N = 4, ΔΔG = + 6.455 kcal/mol relative to celecoxib). It falls below the celecoxib reference but ranks second in the full panel, ahead of all fig-panel compounds. The seven fig compounds span approximately 10 kcal/mol among themselves (bergapten − 20.320 ± 0.972 kcal/mol to hydroxytyrosol − 10.342 ± 0.564 kcal/mol), with psoralen − 18.945 ± 1.163, luteolin − 16.535 ± 0.876, elenolic acid − 14.875 ± 1.021, and quercetin − 14.357 ± 1.532 kcal/mol (N = 4, with s9999 outlier disclosed in the companion paper) in between. All nine quantified compounds show negative ΔG values at the identified COX-2 pocket; none except oleocanthal exceeds the celecoxib reference. Bergapten and psoralen, despite their favorable binding energies, carry a structural caveat: both are furanocoumarins with documented CYP450 inhibitory and phototoxic properties [ 23 ]. The computational result does not address these safety liabilities; they are noted here to prevent any unqualified advancement claim. 3.1.3 Ceiling and two-paper disposition Gate evaluation (gate_evaluate.py, thresholds from NODE1_DECISION_SHEET.md): mean ΔG < − 10 kcal/mol — all nine quantified compounds pass; CV < 15% — all nine pass; ligand RMSD 80% of equilibrated trajectory — all pass; pocket contact occupancy ≥ 5/9 residues at ≥ 50% — all pass. outcome_key = ALL_PASS for the full panel (oracle verify_node1.py, 46 checks, PASS, last run 2026-03-24). Per-compound values are tabulated in Table 2 . The authoritative ceiling language for Node 1: “Oleocanthal demonstrates the strongest computed COX-2 binding affinity among screened fig and olive phytochemicals (ΔG = − 61.8 kcal/mol, N = 4, CV = 2.4%), substantially exceeding celecoxib (ΔΔG = − 22.9 kcal/mol, Welch t = − 24.1, p < < 0.001). Results are computationally reproducible under the stated protocol. Wet-lab and safety characterization required before any therapeutic claim.” The Node 1 results were split into two companion papers at submission: Paper 1 (OLIVE) covers oleocanthal and ligstroside D2; Paper 2 (FIG) covers the seven remaining fig compounds. The split reflects the natural compound origin and does not alter the gate outcome. Node 1 is locked; the result set is protected by a 46-check regression oracle (verify_node1.py) that verifies every ΔG value against the raw FINAL_RESULTS_DECOMP.dat files and confirms no governance artifact has been modified since its freeze date. Table 2 MM-GBSA binding free energies for the COX-2 panel (PDB 5KIR, rofecoxib-bound complex; GROMACS 2026, CHARMM36/GAFF2, igb = 2, saltcon = 0.150 M). Celecoxib included as external reference compound (not co-crystallized in 5KIR template). All values from ΔTOTAL line of FINAL_RESULTS_DECOMP.dat; SHA256-verified, oracle PASS (verify_node1.py, 2026-03-24). ΔΔG = compound − celecoxib. CV = SD / |mean| × 100. Compound Mean ΔG (kcal/mol) SD N CV (%) ΔΔG vs celecoxib Note Oleocanthal −61.765 1.477 4 2.4 −22.890 OLIVE paper Celecoxib −38.875 1.199 4 3.1 reference Clinical comparator Ligstroside D2 −32.420 2.082 4 6.4 + 6.455 OLIVE paper Bergapten −20.320 0.972 4 4.8 + 18.555 FIG paper; furanocoumarin caveat Psoralen −18.945 1.163 4 6.1 + 19.930 FIG paper; furanocoumarin caveat Luteolin −16.535 0.876 4 5.3 + 22.340 FIG paper Elenolic acid −14.875 1.021 4 6.9 + 24.000 FIG paper Quercetin −14.357 1.532 4 10.7 + 24.518 FIG paper; s9999 outlier disclosed Hydroxytyrosol −10.342 0.564 4 5.5 + 28.533 FIG paper Oleuropein EXCLUDED — — — — GAFF2 failure (+ 310 kcal/mol) Rutin EXCLUDED — — — — GAFF2 failure (+ 576 kcal/mol) 3.2 Node 2: IKKβ × Luteolin 3.2.1 Initial run and gate evaluation Four independent production simulations of the luteolin–IKKβ complex (PDB 4KIK [ 3 ]) were run from distinct velocity seeds (s1234, s42, s9999, s_rerun), each for 5 ns. One additional trajectory (prod_main) was excluded from all analysis due to a concurrent-write file corruption detected at t = 60 ps; its exclusion is documented in the Node 2 receipt trail and did not affect the four-replicate evidence base. Gate evaluation proceeded against four pre-specified criteria frozen in the decision sheet before any result was seen (NODE2_DECISION_SHEET.md, frozen 2026-03-21). Three of four gates passed. Gate 1 (ligand RMSD 80% of t = 500–5000 ps): 3/4 runs pass (s1234 99.9% of frames stable, s9999 98.2%, s_rerun 100%; s42 fails at 43.1% stable). Gate 2 (≥ 5/9 ATP-pocket residues at ≥ 50% contact occupancy): 4/4 pass (s1234 8/9, s42 5/9, s9999 8/9, s_rerun 6/9). Gate 3 (mean ΔG < − 10 kcal/mol): pass, mean ΔG = − 28.18 kcal/mol across N = 4. Gate 4 (CV < 15%): fail, CV = 16.1% (SD = 4.55 kcal/mol over the range − 23.30 to − 32.41 kcal/mol). The per-seed distribution is shown in the left panel of Fig. 3 . The platform emitted outcome_key = CV_FAIL_ONLY and applied the pre-specified ceiling for that outcome class: “Luteolin remains a computationally credible but not yet reproducible IKKβ hypothesis.” No upward revision was permitted. The safe claim was written, filed, and the node was held at “computationally credible.” 3.2.2 Diagnosis The CV failure prompted structural investigation rather than quiet retry. The N = 4 ΔG distribution is visibly bimodal: two replicates (s1234 − 23.30, s42 − 25.32 kcal/mol) cluster near − 24 kcal/mol, while two (s9999 − 32.41, s_rerun − 31.67 kcal/mol) cluster near − 32 kcal/mol. Contact occupancy data pointed to the mechanism: s42 undergoes a pose transition at approximately t = 1700 ps, after which LYS44, GLU61, and MET96 occupancies fall below 50%. The replicate does not leave the pocket but shifts from deep-pocket engagement to a shallower orientation. s1234, despite maintaining contacts throughout, shows weaker MM-GBSA values than s9999/s_rerun, suggesting the two sub-populations represent distinct binding modes within the same pocket region. To test whether structural clustering could resolve the CV issue — separating the two apparent modes and re-evaluating each independently — a KMeans analysis (k = 2) was run on 363 frames of ligand heavy-atom coordinates after backbone Cα superposition. The silhouette score of 0.53 confirmed two structurally distinguishable populations (Cluster 0: GLU61/MET96 dominant, 37% of frames; Cluster 1: GLY102/ASP166 dominant, 63% of frames). MM-GBSA was re-run within each cluster. Result: Cluster 0 CV = 21.5%, Cluster 1 CV = 21.3%. Both values exceed the 15% gate threshold; both exceed the original N = 4 CV of 16.1%. Structural stratification did not reduce CV. The data are consistent with energy variance driven by trajectory phase — the timing of the s42 pose transition — rather than by structural mode. The two binding mode families are not energetically separable by this cluster partition. This does not prove the pose transition is the sole source of variance; it establishes that the cluster separation does not resolve it. This diagnosis was documented in the decision sheet and filed before the reseed decision was made. The gate verdict remained unchanged: CV_FAIL_ONLY. The ceiling was not revised. 3.2.3 Pre-specified upgrade gate A deep-pocket reseed was motivated by the diagnosis: if the shallow-pocket geometry allows pose drift (as observed in s42), constraining the starting geometry to the deep-pocket binding family may eliminate the bimodal distribution and resolve the CV failure. Before any reseed trajectory was run, a four-criterion upgrade gate was locked in the decision sheet (NODE2_UPGRADE_RECORD.md). All four criteria were required simultaneously; any single failure would leave the ceiling unchanged: LIG RMSD: ≥3/N seeds 80% of t = 500–5000 ps Contact occupancy: ≥3/N seeds ≥ 5/9 ATP-pocket residues at ≥ 50% occupancy MM-GBSA direction: mean ΔG < − 10 kcal/mol MM-GBSA spread: CV < 15% across N ≥ 3 converged runs The upgrade gate follows the same four-criterion structure as the original gate. It is not a relaxed standard — it is the same standard applied to a new experiment. This distinction matters: the reseed is not a do-over of the original run with lower expectations; it is a new, pre-registered experiment testing a specific structural hypothesis. 3.2.4 Deep-pocket reseed (IKKb_LUT_deepPocket_v1) Three independent seeds (s_dp1, s_dp2, s_dp3) were run from the deep-pocket starting geometry for 5 ns each. All four upgrade gate criteria passed. Gate 1 (LIG RMSD): 3/3 seeds stable (s_dp1: 1.56 Å mean, s_dp2: 1.66 Å, s_dp3: 1.56 Å; all > 80% frames < 3 Å). Gate 2 (contact occupancy): 3/3 seeds ≥ 5/9 residues (s_dp1 7/9, s_dp2 5/9, s_dp3 6/9). Gate 3 (ΔG direction): mean ΔG = − 31.457 kcal/mol. Gate 4 (CV): 2.156/31.457 = 6.85%, well within the 15% threshold. Individual replicate values (from SHA256-verified FINAL_RESULTS_DECOMP.dat files): s_dp1 = − 29.05, s_dp2 = − 32.11, s_dp3 = − 33.21 kcal/mol. All three seeds maintain deep-cavity contact (Mode B orientation) throughout, with no pose transitions observed. The bimodal distribution from the original N = 4 run is absent. The contrast between the initial run and the reseed is shown in Fig. 3 . The upgrade gate was met in full. Per the pre-specified rule, the ceiling was upgraded to “computationally reproducible.” The authoritative ceiling language: Luteolin demonstrates computationally reproducible binding to the IKKβ deep-cavity ATP pocket (mean ΔG = − 31.457 kcal/mol, CV = 6.85%, N = 3, ALL_PASS). Evidence tier: EXPLORATORY DDT. Safety characterization required before advancing. The upgrade is narrow in scope: it applies to the deep-pocket geometry and Mode B binding family. It does not assert that luteolin inhibits IKKβ in vitro or in vivo [ 16 , 17 ], and it does not apply to alternative binding orientations that were not tested. 3.2.5 What the Node 2 narrative demonstrates Node 2 illustrates the full platform cycle in a single experiment: initial failure captured correctly (CV_FAIL_ONLY), ceiling applied without revision, structural diagnosis performed, negative result of the cluster analysis filed before any further decision was made, upgrade gate locked, reseed run, and ceiling upgraded only after all criteria met. Every step is auditable through the decision sheet, receipt trail, and SHA256-verified raw data. No intermediate result was discarded. 3.3 Node 3: IKKβ × Quercetin (generalization test) 3.3.1 Rationale and design The deep-pocket reseed result for luteolin (Node 2) established a specific claim: luteolin binds the IKKβ deep-cavity ATP pocket reproducibly in the Mode B orientation. The natural next question was whether this result was ligand-specific or whether the pocket supports a class of structurally related flavonoids [ 17 , 18 , 19 ]. Quercetin is the structurally closest compound in the fig panel to luteolin, differing by a single hydroxyl group at the 3-position of the C-ring. If the IKKβ deep-cavity pocket is a generalizable feature, quercetin should reproduce under the same starting geometry and gate criteria. If quercetin fails, the luteolin result stands as a single-compound observation and the generalization claim is not supported. Node 3 was run as a pre-specified generalization test — not an exploratory screen — with the same four gate criteria applied, the same starting geometry (deep-cavity), and N = 3 independent seeds (s1111, s2222, s3333) for 5 ns each. 3.3.2 Results Individual replicate ΔG values (ΔTOTAL, SHA256-verified FINAL_RESULTS_DECOMP.dat): s1111 = − 24.47 kcal/mol, s2222 = − 23.93 kcal/mol, s3333 = − 25.41 kcal/mol. Mean ± SD = − 24.603 ± 0.749 kcal/mol, CV = 3.04%. Gate evaluation: Gate 1 (ligand RMSD) — s1111: 1.893 Å mean, 99.4% stable (PASS); s2222: 3.191 Å mean, 34.4% stable (FAIL, borderline); s3333: 2.016 Å mean, 93.2% stable (PASS). Gate passes at 2/3 threshold; s2222 is a disclosed borderline case that a fourth seed could affect. Gate 2 (contact occupancy): s1111 6/9 PASS, s2222 5/9 PASS, s3333 7/9 PASS; 3/3 pass at the ≥ 5/9 threshold. Gate 3 (ΔG direction): mean − 24.603 kcal/mol, PASS ( < − 10 kcal/mol). Gate 4 (CV): 3.04%, PASS (< 15%). outcome_key = ALL_PASS. Binding mode distribution (from contact occupancy timeline, analysis_scripts/contact_analysis.py): all three quercetin seeds sample both Mode A (hinge-region contact) and Mode B (deep-cavity contact) in a 2:1 B:A ratio over the 5 ns trajectories. The per-mode mean ΔG difference is 1.2 kcal/mol, less than RT at physiological temperature (≈ 0.6 kcal/mol at 310 K). The two modes are energetically degenerate under this protocol. Luteolin, in contrast, maintained Mode B in all three deep-pocket seeds with no observed transitions. The two-flavonoid comparison is presented in Fig. 5 and tabulated in Table 3 . The structural basis for the mode difference between luteolin and quercetin was not determined by this analysis. The additional 3-OH group on quercetin's C-ring is one candidate explanation; no structural simulation specifically testing this hypothesis was run. The observation is descriptive: quercetin samples two modes, luteolin samples one. 3.3.3 Boltz-2 orthogonal check An orthogonal structure prediction was run for quercetin at the IKKβ pocket using Boltz-2. The pre-specified proximity gate required the predicted ligand centroid to fall within 8 Å of the ATP-pocket anchor residue. The measured centroid distance was 10.1 Å from the pre-specified anchor (INCOMPATIBLE). In parallel, 8 of the 9 pre-specified ATP-pocket contact residues fell within 5 Å of the predicted ligand centroid (CONSISTENT). The Boltz model pLDDT was 84 (INTERPRETABLE — above the confidence threshold). A summary of the Boltz-2 evaluation is presented in Fig. 4 . The governance rule applied to Boltz-2 results is explicit: an orthogonal structural prediction cannot rescue a failed gate or upgrade a ceiling. It also cannot downgrade a passed gate. The Boltz result is recorded as an orthogonal observation: the centroid is shifted relative to the MD-derived pocket anchor, but the contact residue overlap is substantial. This is consistent with a slightly displaced orientation within the same pocket region, not with absence of pose. The MD ceiling was not altered by the Boltz result in either direction. 3.3.4 Two-flavonoid IKKβ result The combination of Node 2 (luteolin, CV = 6.85%, Mode B exclusively) and Node 3 (quercetin, CV = 3.04%, Modes A + B degenerate) supports the following bounded claim, which is the highest claim the data support: “Both fig-side flavonoids tested reproduce computationally at the IKKβ deep-cavity ATP pocket under identical starting conditions. Mode selection differs: luteolin locks Mode B; quercetin samples both modes with energetic degeneracy. The IKKβ deep-cavity pocket is computationally viable for flavonoid binding. Starting-pose dependence is confirmed; alternative starting geometries may yield different mode selections.” Table 3 IKKβ two-flavonoid comparison (PDB 4KIK, deep-cavity starting geometry). Node 2 = luteolin (IKKb_LUT_deepPocket_v1); Node 3 = quercetin (Quercetin_IKKb_v1). Per-seed ΔG from SHA256-verified FINAL_RESULTS_DECOMP.dat (training_pool/cycle_1–2). Parameter Node 2: Luteolin Node 3: Quercetin Experiment ID IKKb_LUT_deepPocket_v1 Quercetin_IKKb_v1 N seeds 3 (s_dp1, s_dp2, s_dp3) 3 (s1111, s2222, s3333) Per-seed ΔG (kcal/mol) −29.05, − 32.11, − 33.21 −24.47, − 23.93, − 25.41 Mean ΔG ± SD −31.457 ± 2.154 −24.603 ± 0.749 CV (%) 6.85 (PASS < 15%) 3.04 (PASS < 15%) Gate 1 (RMSD) 3/3 PASS 2/3 PASS (s2222 borderline) Gate 2 (contacts) 3/3 PASS (7/9, 5/9, 6/9) 3/3 PASS (6/9, 5/9, 7/9) Gate 3 (ΔG direction) PASS ( < − 10 kcal/mol) PASS ( < − 10 kcal/mol) Gate 4 (CV) PASS (6.85%) PASS (3.04%) outcome_key ALL_PASS ALL_PASS Binding mode Mode B exclusively (3/3) Modes A + B, 2:1 B:A (degenerate) Boltz-2 centroid (Å) Stalled — no output 10.1 Å (INCOMPATIBLE) Boltz-2 contacts N/A 8/9 within 5 Å (CONSISTENT) Boltz-2 pLDDT N/A 84 (INTERPRETABLE) Evidence tier EXPLORATORY DDT EXPLORATORY DDT Safety tier CONCERN (ER agonist, IKK clinical fail) [ 19 ] CAUTION (hERG and CYP-mediated drug-drug interaction risk reported; see [ 19 ] for flavonoid class review) This claim does not assert IKKβ inhibition by either compound, does not extend to in vitro or in vivo contexts, and does not assert flavonoid class-level activity. Both nodes are EXPLORATORY DDT tier. 3.4 Node 4: p65 × Quercetin (honest branch closure) 3.4.1 Rationale and hypothesis With quercetin showing computationally reproducible binding at IKKβ, an exploratory hypothesis was formed: if quercetin engages the IKKβ node of the NF-κB pathway [ 20 , 21 ], does it also engage the pathway downstream at the NF-κB p65 transcription factor? The p65 DNA-binding interface (PDB: 1VKX [ 11 ]) was selected as the candidate target. The 1VKX structure is a DNA-bound complex; no validated small-molecule binding pocket is established for this conformation, and no prior crystallographic evidence of small-molecule engagement at this interface exists. This selection was therefore explicitly exploratory — a structural adjacency test, not a docking target with known druggability. This is explicitly an exploratory branch, not a validated DDT node — the target is mechanistically adjacent but structurally unrelated to IKKβ, and the hypothesis was formed after, not before, the IKKβ result. The decision sheet for Node 4 pre-specified the same four gate criteria as Nodes 2 and 3. The pre-specified outcome envelope for POSE_UNSTABLE was written before any simulation was run (NODE4_VERDICT_ENVELOPES.md): “Quercetin does not show reproducible binding at the NF-κB p65 DNA-binding interface under the current gate framework. POSE UNSTABLE — branch CLOSED.” The platform claim update was also pre-specified: if Node 4 fails, the two-node claim is preserved and no Node 5 routing is permitted. 3.4.2 Results and closure Three independent seeds (s1111, s2222, s3333) were run from the docked starting geometry for 5 ns each. Gate 1 (ligand RMSD): s1111 3.103 Å mean, 51.2% stable (FAIL); s2222 11.967 Å, 0.0% stable (FAIL); s3333 3.977 Å, 40.4% stable (FAIL); 0/3 pass. Gate 2 (contact occupancy): 0/3 seeds reach ≥ 5/9 interface residues; maximum observed was 1/9 per seed. Gate 3 (ΔG direction): mean = − 12.357 kcal/mol; PASS ( < − 10 kcal/mol). Gate 4 (CV): CV = 55.85%; FAIL. Failing gates: 1, 2, 4. outcome_key = POSE_UNSTABLE. The pre-specified ceiling was applied without modification: “Quercetin does not show reproducible binding at the NF-κB p65 DNA-binding interface under the current gate framework. POSE UNSTABLE — pose drift observed across replicates, insufficient for credibility claim. p65 branch CLOSED. No Node 5 routing permitted.” The platform claim was updated accordingly: “Two-node model (COX-2 + IKKβ). p65 branch closed. No cascade evidence.” This is a bounded update — the two-node model is not weakened by the Node 4 closure; it is correctly scoped by it. Gate 3 passing does not alter the verdict: a favorable ΔG direction mean without pose stability or pocket contact is insufficient evidence for a credibility claim under the pre-specified gate structure. 3.4.3 What honest closure demonstrates The value of Node 4 is not its scientific content — pose instability at p65 does not advance the mechanistic understanding of quercetin. Its value is architectural: it demonstrates that the platform closes a branch under the same gate framework that produces positive results in other branches. A platform that closes branches honestly is one whose positive results carry evidential weight. If Node 2 and Node 3 had both passed while Node 4 failed but was never documented, a reader of the final publication would have no way to assess the false discovery rate of the pipeline. By documenting Node 4 — with the pre-specified envelope, the gate failures, and the explicit update to the platform claim — the pipeline provides the information needed to evaluate the positive results in context. The Node 4 receipt trail and verdict are archived as frozen artifacts in the same artifact set as the positive nodes. Node 4 is not a footnote; it is a first-class result of the wrong kind. A cross-node summary of all four experimental branches is presented in Table 4 . Table 4 Cross-node summary: all four experimental branches. All values from SHA256-verified FINAL_RESULTS_DECOMP.dat or gate_verdict.json. Claim ceilings locked in verdict_templates.json before verdicts were written. Node / Run Target Compound Mean ΔG (kcal/mol) CV (%) N Gate outcome Branch decision Node 1 COX-2 (5KIR) Oleocanthal (lead) −61.765 ± 1.477 2.4 4 ALL_PASS PROMOTE (published) Node 2 initial IKKβ (4KIK) Luteolin −28.18 ± 4.55 16.1 4 CV_FAIL_ONLY ARCHIVE → reseed Node 2 reseed IKKβ (4KIK) Luteolin −31.457 ± 2.154 6.85 3 ALL_PASS ADEQUATE — exploratory DDT Node 3 IKKβ (4KIK) Quercetin −24.603 ± 0.749 3.04 3 ALL_PASS ADEQUATE — exploratory DDT Node 4 p65 (1VKX) Quercetin −12.357 ± 6.906 55.85 3 POSE_UNSTABLE KILL — branch closed 4. Discussion 4.1 Governed multi-node discovery versus compound ranking A standard computational polypharmacology screen [ 5 , 6 ] produces a ranked list of compounds by binding affinity. This list has interpretive value, but it lacks governance: the ranking does not specify what claims the ΔG values support, does not document which branches failed, and does not constrain the language available to describe the results. The ranked list is a computation; it is not by itself an evidence-governed claim. The platform described here produces a ranked list and enforces governance on every entry. For each compound–target pair, the operative unit is the four-column ceiling table described in Section 2.3 : what was observed, what may be claimed, what may not be claimed, and what condition would permit an upgrade. This table is not commentary on the results; it is the output of the pipeline, generated deterministically from raw data and pre-specified templates. The ceiling table constrains what can be claimed from any given ΔG value: the claim boundary is set before results arrive and cannot be revised based on what the numbers turned out to be. It does not constrain the ΔG values themselves; it constrains the language used to describe them. 4.2 Failure as a first-class result Node 2's initial CV failure (CV = 16.1%, gate threshold < 15%; gate_verdict.json, NODE2_DECISION_SHEET.md) produced three subsequent analysis steps: a structural diagnosis of the bimodal ΔG distribution, a KMeans cluster analysis (result: CV worsened to 21.5% and 21.3% by cluster — variance is not mode-separable), and a pre-specified upgrade gate locked before the reseed was executed. If CV = 16.1% had been described as “modest variability” — as it could have been, and as similar values routinely are in the computational polypharmacology literature — none of that analysis would have been triggered. The diagnosis, the cluster analysis, the upgrade gate, and the Mode B characterization all follow from treating the CV failure as a failure rather than a success with a softened label. Node 4 is the same principle applied to branch closure. The p65 result is negative. Its documentation prevents the platform from claiming a mechanistic cascade it has not established, and it provides the denominator needed to assess the positive result rate of the pipeline. A system that reports only successes has no interpretable success rate. 4.3 The rescue pattern as a reusable protocol contribution The Node 2 rescue sequence follows a three-step structure that can be applied to any node producing outcome_key = CV_FAIL_ONLY with a structural diagnostic candidate: (1) diagnose the source of CV variance using structural clustering or contact timeline analysis; (2) formulate a specific, falsifiable structural hypothesis about what change to the starting conditions would eliminate the variance; (3) lock an upgrade gate — with the same four criteria as the original gate, in a named file, before executing the reseed. Step 3 is the one that prevents the reseed from becoming a do-over with softened expectations. Step 2 is the one that prevents step 3 from being justified post-hoc. And step 1 requires the original failure to be documented — which is why a protocol that encourages quiet retry cannot support this pattern. The Node 2 cluster analysis that made the CV worse is not a failed experiment; it is the evidentiary foundation that makes the upgrade gate legitimate. 4.4 Boltz-2 as bounded orthogonal evidence The Boltz-2 predictions for quercetin represent a methodological choice about how to handle orthogonal evidence. The governance rule applied is explicit: structure prediction cannot rescue a failed MD gate or upgrade an MD ceiling, and it cannot downgrade a passed gate. This rule is necessary because the two methods measure different things under different assumptions; allowing one to override the other removes the interpretive boundary between complementary observations and post-hoc rationalization. Within this constraint, Boltz-2 functions as a consistency check. For quercetin, the centroid distance (10.1 Å from the pre-specified anchor) and the contact residue overlap (8/9 within 5 Å) together describe a prediction shifted from the MD-derived pocket center but substantially overlapping with the MD contact set (Fig. 4 ). Two independent methods — one dynamics-based, one structure-prediction-based — both locate quercetin in the vicinity of the IKKβ ATP-pocket contact residues. Neither method is upgraded or downgraded by that observation; both are reported at their respective evidence tiers. The luteolin Boltz-2 prediction stalled due to a CUDA/PyTorch incompatibility and produced no scientific output. This outcome is documented in the artifact set alongside the positive Boltz result. The absence of a Boltz prediction for luteolin does not affect the MD ceiling. 4.5 Self-improving formatter: architecture versus deployment The self-improving verdict formatter is included here as an architectural contribution, not as a deployed system. The first complete training run — Qwen/Qwen2.5-7B [ 25 ], LoRA [ 24 ] rank = 16, three verified cycles — produced an adapter that outputs correct per-seed ΔG values, includes SHA256 provenance, and generates no fabricated numbers. It does not yet produce the required ceiling phrase or markdown template format: three training examples are insufficient to anchor a specific phrase pattern in a 7B model. The adapter produces evidence of learning without meeting the promotion gate (eval_report.json overall_pass = True). It is not promoted. The production formatter remains the rule-based system (format_verdict.py) until the corpus reaches sufficient size for ceiling template anchoring — estimated 5–10 cycles, including at least one CV_FAIL_ONLY example. The training loop accumulates each new closed experimental cycle; the next fine-tune trigger will be data-driven, not calendar-driven. The architecture is described here so that the pattern is available to other groups building governance-compliant computational pipelines. The current training state is described as it is, so that no reader mistakes the architectural description for a claim about a production-ready system. 4.6 Rational polypharmacology context The experimental design — two independent molecular targets, multiple compounds tested at each — reflects the 2026 rational polypharmacology paradigm [ 5 , 6 ]: multi-target engagement as a feature to be characterized and governed, not incidentally discovered. The COX-2 and IKKβ nodes represent independent anti-inflammatory mechanisms (prostaglandin synthesis [ 12 ] and NF-κB activation [ 20 , 21 ], respectively), and the compound panel includes structures — luteolin, quercetin — with prior literature support for activity at both target classes [ 16 , 19 ]. The IKKβ ATP-binding pocket targeted here corresponds to the conserved kinase hinge region and catalytic cleft, which are established binding sites for small-molecule inhibitors [ 3 ] and have prior literature support for flavonoid engagement [ 16 , 19 ]. The platform's contribution to this context is not the specific compound results but the methodology for generating bounded, auditable, reproducible results at multiple nodes. The highest supported claim of this work is that a computationally reproducible two-node profile exists for select fig and olive phytochemicals at COX-2 and IKKβ, documented under governance-constrained conditions. The appropriate next step is wet-lab validation at each node independently before any cross-node inference is drawn. 5. Limitations 5.1 Computational evidence only All binding free energy values in this paper are MM-GBSA estimates from classical force-field molecular dynamics simulations. MM-GBSA is an end-point free energy method with known limitations [ 7 ]: it does not account for full conformational entropy of the protein, it uses implicit solvent, and its absolute values are sensitive to the GB model and salt concentration parameters. The values reported here are internally consistent within the stated protocol (igb = 2, saltcon = 0.150 M) and reproducible across independent replicates, but they are not experimental binding affinities. MM-GBSA ΔG values reported here are not directly comparable to experimental binding free energies (ΔG_exp, −RT·ln(K_i), or IC₅₀ equivalents); rank-ordering within this study under a fixed protocol is valid, but cross-study numerical comparison of absolute values is not. No IC₅₀, K_i, K_d, or surface plasmon resonance data were collected for any compound in this paper. All results are classified at the EXPLORATORY DDT (Drug Discovery Tool) tier: appropriate for target selection and hypothesis prioritization, not for therapeutic claims. Advancement beyond this tier requires experimental validation. 5.2 Force-field coverage gaps The GAFF2 force field [ 2 ] failed for glycosidic linkages in oleuropein and rutin, producing unphysical positive ΔG values (+ 310 and + 576 kcal/mol, respectively). These compounds were excluded with documentation rather than silently dropped. The exclusions do not affect the ranked results for the nine quantified compounds, but they represent a genuine coverage gap: glycoside-containing phytochemicals cannot be reliably evaluated under the current protocol without a force-field extension or alternative parameterization. No instability or unphysical energy terms were observed for oleocanthal or other non-glycosidic compounds, suggesting the GAFF2 limitation is specific to glycosidic linkages rather than a global parameterization issue. Bergapten and psoralen are included in the ranked results with a standing caveat: both are furanocoumarins with documented CYP450 inhibitory and phototoxic liabilities [ 23 ]. The computational binding result at COX-2 is accurate under the stated protocol, but the safety signal is not addressed by any analysis in this paper. Advancement of either compound requires explicit safety characterization prior to any in vitro testing. 5.3 Starting-geometry dependence The Node 2 rescue demonstrated directly that MD results at IKKβ are starting-geometry dependent: the initial run from a shallow-pocket geometry produced a bimodal ΔG distribution (CV = 16.1%), while the deep-pocket reseed from a constrained geometry produced a convergent single-mode result (CV = 6.85%). The quercetin Node 3 result — run from the same deep-pocket starting geometry — sampled both Mode A and Mode B, confirming that mode selection is not fully determined by the starting pose. Results at IKKβ are therefore conditional on the starting geometry. The two-flavonoid result described in Section 3.3.4 is valid within the stated starting conditions; it does not characterize the full conformational landscape of either compound at the IKKβ pocket, and alternative starting geometries may produce different mode distributions. 5.4 Fixed crystal structure and single-conformation constraint Each node in this platform tests a single molecular target with a fixed crystal structure (5KIR for COX-2 (rofecoxib-bound complex) [ 9 , 10 ], 4KIK for IKKβ [ 3 ], 1VKX for p65 [ 11 ]; all entries retrieved from the PDB [ 3 ]). Crystal structures represent a single conformational snapshot; induced fit, allosteric effects, and protein flexibility not captured by the starting structure are not modeled. The gate criteria and ceiling templates are calibrated to the specific structure used; results may not transfer directly to alternative structures of the same target. For COX-2, alternative crystal structures (e.g., celecoxib-bound 3LN1 or arachidonic-acid-bound 1DIY) were not explored; all COX-2 results are conditional on the 5KIR conformational state. For IKKβ in particular, the 4KIK structure (655-residue kinase domain, chain A) was used with a residue offset of − 8 between PDB numbering and the GRO system used in analysis. This offset is documented in the decision sheet and applied consistently; it is a potential source of numbering confusion when cross-referencing literature that uses PDB-canonical residue numbers. 5.5 Training pool size The self-improving verdict formatter has been trained on three experimental cycles. Three cycles are insufficient to anchor the required ceiling phrase pattern in a 7B model; the gap is characterized in Section 4.5 and represents a genuine limitation on the current state of the formatter component. The formatter in production use is the rule-based system (format_verdict.py), not the adapter; the gap does not affect any result reported in this paper, but the self-improving loop has not yet reached the state where it can substitute for the rule-based formatter without human review. 5.6 Absence of wet-lab validation No compounds in this paper have been tested in vitro or in vivo. There is no biochemical binding data, no cell-based NF-κB reporter assay, no kinase selectivity panel, and no safety characterization against the LTKB or EDKB toxicity databases. All claims are bounded to computationally reproducible under the stated protocol. This is the correct claim for the evidence tier. It is not a temporary limitation that will be removed by additional computation; it is a ceiling that can only be raised by experimental work. 6. Conclusions This paper describes a governed multi-node computational discovery platform and reports its gated outcomes across four experimental branches: a nine-compound COX-2 panel (Node 1, ALL_PASS; Table 2 ), luteolin at IKKβ following a pre-registered rescue from an initial CV failure (Node 2, CV_FAIL_ONLY → ALL_PASS; Fig. 3 ), quercetin at IKKβ as a generalization test (Node 3, ALL_PASS; Table 3 , Fig. 5 ), and quercetin at p65 as a pre-specified branch closure (Node 4, POSE_UNSTABLE; Table 4 ). The compound results are bounded to the evidence tier at which they were obtained. They are illustrations of the platform producing correct gated outcomes — including failure. The platform contribution comprises seven interlocking components: pre-specified gates frozen before any simulation runs; outcome-keyed ceiling templates that remove interpretive latitude from the result-to-claim step; a 46-check regression oracle that verifies every published number against the raw output file it came from; a SHA256 provenance chain from raw data to published claim; a retrieval-first knowledge graph that prevents memory-only responses to questions that can be grounded in the evidence record; a NemoClaw constitutional review layer that scans every formatted verdict for forbidden language before archiving; and a self-improving formatter that accumulates each closed experimental cycle as a training example under hash-verified admission criteria. The integrated pipeline is summarized in Fig. 1 and the full gate outcomes are in Table 1 . Each component targets one of the five failure modes named in Section 1 : pre-specified gates address post-hoc threshold setting; ceiling templates address claim ceiling collapse; receipt trails and branch closure records address silent omission; SHA256 oracle addresses provenance breaks; N ≥ 3 velocity-seed replicates with pre-specified CV gate address single-trajectory reporting. None of these failure modes is addressed by better statistics alone; all require governance built into the pipeline structure. The scientific output is a bounded two-node model: oleocanthal at COX-2 (ΔG = − 61.8 kcal/mol, CV = 2.4%, N = 4) and both luteolin (ΔG = − 31.5 kcal/mol, CV = 6.85%, N = 3) and quercetin (ΔG = − 24.6 kcal/mol, CV = 3.04%, N = 3) at the IKKβ deep-cavity ATP pocket. These results are computationally reproducible under the stated protocol and are EXPLORATORY DDT tier. The appropriate next steps are enzymatic binding assays at COX-2 and IKKβ, a kinase selectivity panel for the IKKβ hits, and safety characterization (CYP450 profiling for furanocoumarins [ 23 ], LTKB/EDKB screening for all confirmed nodes) before any advancement claim is made. The full frozen artifact set — 28 SHA256-verified files including decision sheets, receipt trails, verdict templates, the regression oracle, and the KG export — is archived at https://doi.org/10.5281/zenodo.19076569 and is sufficient to audit and verify all reported results (a representative subset is enumerated in Table 5 ). The governance layer is not optional; it is structurally embedded in every stage of the experimental cycle. The pipeline implementation is proprietary and not publicly released; the methodology, protocols, and gate framework are described in full in this paper so that the approach can be independently implemented by other groups. Table 5 Representative frozen artifacts (FREEZE_MANIFEST.json, 2026-03-24). Full set: 28 files. All chmod 444 at freeze. SHA256 hashes in FREEZE_MANIFEST.json (Zenodo: https://doi.org/10.5281/zenodo.19076569 ). Artifact Path Role Master outcomes MASTER_OUTCOMES_FROZEN.md Authoritative ΔG, gate verdicts, ceilings (all nodes) Node 2 decision sheet node2_IKKb/NODE2_DECISION_SHEET.md Pre-specified gates; upgrade record; diagnosis trail Node 3 decision sheet node2_IKKb/NODE3_DECISION_SHEET.md Quercetin IKKβ gates and verdict Node 4 decision sheet node4_p65/NODE4_DECISION_SHEET.md p65 branch gates and closure record Verdict templates pipeline/templates/verdict_templates.json Ceiling phrases; NOT SAFE lists Claim ceiling rules pipeline/manifests/claim_ceiling_rules.json Governance rules for ceiling assignment Regression oracle pipeline/oracle/verify_node1.py 46-check Node 1 oracle (PASS 2026-03-24) Node 1 oracle manifest pipeline/oracle/node1_manifest_v1.json Expected values for all 46 oracle checks Raw results md_runs/PAPER_RESULTS_FINAL.txt All Node 1 ΔG values, individual replicates Node 2 receipt trail node2_IKKb/RECEIPT_TRAIL.md Full audit trail Node 2 (590 lines) Cycle 1 gate verdict pipeline/training_pool/cycle_1/inputs/gate_verdict.json SHA256-verified verdict luteolin dp reseed Cycle 2 gate verdict pipeline/training_pool/cycle_2/inputs/gate_verdict.json SHA256-verified verdict quercetin IKKβ Boltz quercetin sheet node2_IKKb/BOLTZ_EVIDENCE_SHEET_QUERCETIN.md Centroid 10.1 Å; 8/9 contacts; pLDDT = 84 Freeze manifest FREEZE_MANIFEST.json SHA256 of all 28 critical artifacts What this paper does not claim: that any compound inhibits its target in vitro or in vivo; that the two-node profile constitutes multi-target synergy; that the training adapter is production-ready; that the computationally reproducible result is transferable to alternative structures of the same target; or that any result in this paper is sufficient to advance a compound beyond the EXPLORATORY DDT tier without experimental validation. Declarations Funding Declaration This work received no specific grant, contract, or sponsorship from any funding agency in the public, commercial, or not-for-profit sectors. The research was conducted independently by the author as an Independent Researcher, using self-funded computational resources. No external financial support, institutional backing, or industry sponsorship contributed to the design, execution, analysis, or writing of this study. Conflicts of Interest The author declares no conflicts of interest. The pipeline implementation is proprietary (developed solely by the author) but is not commercialized; no financial or professional relationships exist that could be perceived to influence the work reported. Data and Code Availability All frozen artifacts supporting the reported results — decision sheets, receipt trails, gate verdicts, verdict templates, the regression oracle, and the knowledge-graph export (28 SHA256-verified files) — are archived at https://doi.org/10.5281/zenodo.19076569. A representative subset is enumerated in Table 5. The pipeline implementation source code is proprietary and not publicly released; the methodology, protocols, and gate framework are described in sufficient detail in Sections 2 and 3 to permit independent re-implementation. AI Use Statement During manuscript preparation, generative AI tools were used to assist in drafting, revising, and language editing of human-directed text. All underlying scientific analyses, numerical results, interpretations, and conclusions were independently reviewed and verified by the author, who takes full responsibility for the final content. AI tools were not used to generate results, perform trajectory analysis, evaluate gate criteria, or make scientific decisions; all such work was performed by the author using the computational tools and protocols described in the Methods section. No AI tool is listed as an author. References Huang J, MacKerell AD Jr. (2013) CHARMM36 all-atom additive protein force field: validation based on comparison to NMR data. J Comput Chem. ;34(25):2135–2145. 10.1002/jcc.23354 . PMID: 23832629 Abraham MJ, Murtola T, Schulz R, Páll S, Smith JC, Hess B, Lindahl E (2015) GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 1–2:19–25 (Tool paper; not indexed in PubMed.) Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE (2000) The Protein Data Bank. Nucleic Acids Res. ;28(1):235–242. 10.1093/nar/28.1.235 . PMID: 10592235 Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O et al (2021) Highly accurate protein structure prediction with AlphaFold. Nature. ;596(7873):583–589. 10.1038/s41586-021-03819-2 . PMID: 34265844 Hopkins AL (2008) Network pharmacology: the next paradigm in drug discovery. Nat Chem Biol. ;4(11):682–690. 10.1038/nchembio.118 . PMID: 18936753 Anighoro A, Bajorath J, Rastelli G (2014) Polypharmacology: challenges and opportunities in drug discovery. J Med Chem. ;57(19):7874–7887. 10.1021/jm5006463 . PMID: 24946140 Genheden S, Ryde U (2015) The MM/PBSA and MM/GBSA methods to estimate ligand-binding affinities. Expert Opin Drug Discov. ;10(5):449–461. 10.1517/17460441.2015.1032936 . PMID: 25835573 Valdés-Tresanco MS, Valdés-Tresanco ME, Valiente PA, Moreno E, gmx_MMPBSA: (2021) A New Tool to Perform End-State Free Energy Calculations with GROMACS. J Chem Theory Comput. ;17(10):6281–6291. 10.1021/acs.jctc.1c00645 . PMID: 34586825 Orlando BJ, Malkowski MG (2016) Crystal structure of rofecoxib bound to human cyclooxygenase-2. Acta Crystallogr F Struct Biol Commun 72(Pt 10):772–776. 10.1107/S2053230X16014230 PMID: 27710942. [PDB 5KIR] Kurumbail RG, Stevens AM, Gierse JK, McDonald JJ, Stegeman RA, Pak JY et al (1996) Structural basis for selective inhibition of cyclooxygenase-2 by anti-inflammatory agents. Nature. ;384(6610):644–648. 10.1038/384644a0 . PMID: 8967954 Chen YQ, Ghosh S, Ghosh G (1998) A novel DNA recognition mode by the NF-κB p65 homodimer. Nat Struct Biol 5(1):67–73. 10.1038/nsb0198-67 PMID: 9437432. [PDB 1VKX] Beauchamp GK, Keast RSJ, Morel D, Lin J, Pika J, Han Q, Lee CH, Smith AB III, Breslin PAS (2005) Phytochemistry: ibuprofen-like activity in extra-virgin olive oil. Nature. ;437(7055):45–46. 10.1038/437045a . PMID: 16136122 Montoya T, Castejón ML, Sánchez-Hidalgo M, González-Benjumea A, Fernández-Bolaños JG, Alarcón-de-la-Lastra C (2019) Oleocanthal modulates LPS-induced murine peritoneal macrophages activation via regulation of inflammasome, Nrf-2/HO-1, and MAPKs signaling pathways. J Agric Food Chem 67(19):5552–5559. 10.1021/acs.jafc.9b00771. PMID: 31042377 Liang J, Bonvino NP, Hung A, Karagiannis TC (2020) In silico characterisation of olive phenolic compounds as potential cyclooxygenase modulators. Part 1. J Mol Graph Model. ;101:107719. 10.1016/j.jmgm.2020.107719 . PMID: 32898836 Karagiannis TC, Ververis K, Liang JJ, Pitsillou E, Kagarakis EA, Yi DTZ, Xu V, Hung A, El-Osta A (2024) Investigation of the anti-inflammatory properties of bioactive compounds from olive mill waste: in silico evaluation of cyclooxygenase enzyme inhibition and pharmacokinetic profiling. Molecules. ;29(15):3502. 10.3390/molecules29153502 . PMID: 39124908 Kim JS, Jobin C (2005) The flavonoid luteolin prevents lipopolysaccharide-induced NF-κB signalling and gene expression by blocking IκB kinase activity in intestinal epithelial cells and bone-marrow derived dendritic cells. Immunology. ;115(3):375–387. 10.1111/j.1365-2567.2005.02156.x . PMID: 15946255 Chen CY, Peng WH, Tsai KD, Hsu SL (2007) Luteolin suppresses inflammation-associated gene expression by blocking NF-κB and AP-1 activation pathway in mouse alveolar macrophages. Life Sci. ;81(23–24):1602–1614. 10.1016/j.lfs.2007.09.028 . PMID: 17977562 Sun ZJ, Chen G, Hu X, Zhang W, Liu Y, Zhu LX, Zhou Q, Zhao YF (2010) Activation of PI3K/Akt/IKK-α/NF-κB signaling pathway is required for apoptosis-evasion in human salivary adenoid cystic carcinoma: its inhibition by quercetin. Apoptosis. ;15(7):850–863. 10.1007/s10495-010-0497-5 . PMID: 20386985 Endale M, Park SC, Kim S, Kim SH, Yang Y, Cho JY, Rhee MH (2013) Quercetin disrupts tyrosine-phosphorylated phosphatidylinositol 3-kinase and myeloid differentiation factor-88 association, and inhibits MAPK/AP-1 and IKK/NF-κB-induced inflammatory mediators production in RAW 264.7 cells. Immunobiology. ;218(12):1452–1467. 10.1016/j.imbio.2013.04.019 . PMID: 23735482 Baldwin AS Jr (1996) The NF-κB and IκB proteins: new discoveries and insights. Annu Rev Immunol. ;14:649–683. 10.1146/annurev.immunol.14.1.649 . PMID: 8717528 Liu F, Xia Y, Parker AS, Verma IM (2012) IKK biology. Immunol Rev. ;246(1):239–253. 10.1111/j.1600-065X.2012.01107.x . PMID: 22435559 Bastys T, Gapsys V, Doncheva NT, Kaiser R, de Groot BL, Kalinina OV (2018) Consistent prediction of mutation effect on drug binding in HIV-1 protease using alchemical calculations. J Chem Theory Comput. ;14(7):3397–3408. 10.1021/acs.jctc.7b01109 . PMID: 29847122 Messer A, Raquet N, Lohr C, Schrenk D (2012) Major furocoumarins in grapefruit juice II: phototoxicity, photogenotoxicity, and inhibitory potency vs. cytochrome P450 3A4 activity. Food Chem Toxicol. ;50(3–4):756–760. doi:10.1016/j.fct.2011.11.023. PMID: 22155270 Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, Wang L, Chen W, LoRA (2022) Low-rank adaptation of large language models. In: Proceedings of the International Conference on Learning Representations (ICLR). arXiv:2106.09685. (Preprint; not indexed in PubMed.) Qwen Team Qwen2.5 Technical Report. arXiv:2412.15115 [cs.CL]. 2024. (Preprint; not indexed in PubMed.) Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9587405","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":641538030,"identity":"642086b1-2b75-4cae-89f4-122dbb6c8c4d","order_by":0,"name":"faycal ferhat","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyUlEQVRIiWNgGAWjYBAC9uYDIMpCDkQeeECMFp5jCSBKwhisJYEULYkNIIo4LWzMzyR+1Eikzw87/BBoi52cbgNBLWxmkj3HJHI33k4zAGpJNjY7QECLvXyDmTQDG1DL7ASQlgOJ2whp4WFj/ybN8E8i3XB2+gditfCYSTO2SSTIS+cQbQtPsWVvn4ThBumcggMJBkT4BeiwjTd+fLORl5+dvvnDhwo7OYJagIBFAkQagFUaEFYOAswfQKR8A3GqR8EoGAWjYAQCADpAP+6StDg/AAAAAElFTkSuQmCC","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"faycal","middleName":"","lastName":"ferhat","suffix":""}],"badges":[],"createdAt":"2026-05-01 15:24:34","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9587405/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9587405/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109456273,"identity":"882ea18e-cd0f-46b4-95fb-6dac5b2eb571","added_by":"auto","created_at":"2026-05-18 10:04:58","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1040882,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePlatform pipeline schematic. Four sequential stages per experimental branch: (1) structure preparation, docking, and MD simulation; (2) MM-GBSA binding free energy estimation via gmx_MMPBSA [8]; (3) gate evaluation against pre-specified criteria frozen in the decision sheet; (4) verdict formatting, ceiling enforcement via NemoClaw constitutional review, and knowledge graph ingestion. SHA256 provenance links every stage output to its successor. Arrows indicate mandatory data flow; no stage may be bypassed.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"figure1pipelineschematic.png","url":"https://assets-eu.researchsquare.com/files/rs-9587405/v1/f336238c044459a5004d2414.png"},{"id":109760387,"identity":"8c560ced-4e8b-4c7e-be6c-d7f1b92a0f37","added_by":"auto","created_at":"2026-05-22 07:28:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":811248,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eMM-GBSA binding free energies for the nine-compound COX-2 panel (PDB 5KIR, rofecoxib-bound complex) plus celecoxib reference. Values are mean ± SD across N=4 independent velocity-seed replicates; all sourced from the ΔTOTAL line of FINAL_RESULTS_DECOMP.dat (SHA256-verified, oracle PASS, verify_node1.py, 2026-03-24). Celecoxib was evaluated as an external reference compound under identical protocol; dashed line = celecoxib reference (−38.875 kcal/mol). Oleuropein and rutin excluded (GAFF2 c6 cross-term failure; documented). Source file: manuscript/figures/fig2_mmpbsa_barplot.png. [Figure 2 is supplied separately by the author and placed here.]\u003c/em\u003e\u003c/p\u003e","description":"","filename":"figure2mmpbsabarplot.png","url":"https://assets-eu.researchsquare.com/files/rs-9587405/v1/c741783583d8ccd4b0c535eb.png"},{"id":109456275,"identity":"ed0e7538-a2b5-430c-8689-fe58bbef1005","added_by":"auto","created_at":"2026-05-18 10:04:58","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":840180,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eNode 2 IKKβ × luteolin: initial run versus deep-pocket reseed. Left panel: initial N=4 run, CV = 16.1% (CV_FAIL_ONLY; per-seed ΔG: s1234 = −23.30, s42 = −25.32, s9999 = −32.41, s_rerun = −31.67 kcal/mol). Right panel: deep-pocket reseed (IKKb_LUT_deepPocket_v1), N=3, CV = 6.85% (ALL_PASS; s_dp1 = −29.05, s_dp2 = −32.11, s_dp3 = −33.21 kcal/mol). All values from SHA256-verified FINAL_RESULTS_DECOMP.dat. Bimodal distribution absent in reseed; all seeds maintain Mode B throughout.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"figure3node2initialvsreseed.png","url":"https://assets-eu.researchsquare.com/files/rs-9587405/v1/f79b2e97c59cf8509b231251.png"},{"id":109456276,"identity":"e2d9c7de-362e-4ffe-a006-109c44eddf80","added_by":"auto","created_at":"2026-05-18 10:04:58","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":741415,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eBoltz-2 orthogonal structure prediction for quercetin at the IKKβ ATP pocket (PDB 4KIK). Predicted ligand centroid distance to pre-specified anchor residue: 10.1 Å (INCOMPATIBLE with proximity gate of 8 Å). Contact residue overlap: 8/9 pre-specified ATP-pocket residues within 5 Å of predicted centroid (CONSISTENT). Model pLDDT = 84 (INTERPRETABLE). Governance rule applied: orthogonal structural prediction cannot rescue a failed gate or upgrade a ceiling. Recorded as orthogonal observation only. Source: node2_IKKb/BOLTZ_EVIDENCE_SHEET_QUERCETIN.md.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"figure4boltzquercetin.png","url":"https://assets-eu.researchsquare.com/files/rs-9587405/v1/e601bd69df35eb1b7b6f8128.png"},{"id":109456277,"identity":"d86431f1-d8f7-4fd3-aa6f-80380211d42f","added_by":"auto","created_at":"2026-05-18 10:04:58","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":220813,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eNode 3 IKKβ × quercetin (generalization test) versus Node 2 luteolin. Per-seed ΔG values (ΔTOTAL, SHA256-verified FINAL_RESULTS_DECOMP.dat): quercetin s1111 = −24.47, s2222 = −23.93, s3333 = −25.41 kcal/mol (mean = −24.603 ± 0.749 kcal/mol, CV = 3.04%, ALL_PASS); luteolin reseed s_dp1 = −29.05, s_dp2 = −32.11, s_dp3 = −33.21 kcal/mol (mean = −31.457 ± 2.154 kcal/mol, CV = 6.85%, ALL_PASS). Mode distribution: luteolin 3/3 seeds Mode B exclusively; quercetin 2:1 B:A ratio (energetically degenerate, ΔΔG = 1.2 kcal/mol \u0026lt; RT).\u003c/em\u003e\u003c/p\u003e","description":"","filename":"figure5twoflavonoidIKKb1.png","url":"https://assets-eu.researchsquare.com/files/rs-9587405/v1/6f7eede083ee7555943a3036.png"},{"id":109839212,"identity":"3fa63ee0-13a9-4555-85f4-339bdb4073a9","added_by":"auto","created_at":"2026-05-23 13:10:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4024646,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9587405/v1/ff2fd461-cd8d-4cb8-a0d3-c32cc0088409.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Governed Multi-Node Computational Discovery Pipeline: Pre-Specified Gates, Bounded Claims, and Honest Branch Closure in Phytochemical Polypharmacology","fulltext":[{"header":"1. Introduction: The Problem with Ungoverned Computational Polypharmacology","content":"\u003cp\u003eComputational screening of phytochemicals against multiple molecular targets has become technically accessible. GPU-accelerated molecular dynamics, open force fields (CHARMM36 [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], GAFF2 [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]), and publicly available structural data (PDB [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], AlphaFold [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]) allow a single workstation to run multi-target binding free energy estimates in days. This technical accessibility has outpaced the methodological infrastructure needed to interpret the results. The rational-polypharmacology framing has been articulated previously [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], but the governance layer needed to make multi-target computational claims auditable has lagged behind the raw compute capacity.\u003c/p\u003e \u003cp\u003eThe most common failure mode is not technical. It is interpretive: a computation finishes, produces a ΔG value, and the ΔG value is reported as evidence of binding. The claim is calibrated not to the evidence but to the desired conclusion. Five specific patterns recur in the literature of computational polypharmacology and are worth naming directly.\u003c/p\u003e \u003cp\u003eFirst, the threshold problem: gate criteria (ΔG cutoff, RMSD stability threshold, replicate CV limit) are stated in the methods section but set after the results are seen. A CV of 16.1% is described as \u0026ldquo;modest variability\u0026rdquo; when the result is favorable; the same value would be described as a failure if a different threshold had been locked before the run. The threshold is not a gate \u0026mdash; it is a post-hoc frame.\u003c/p\u003e \u003cp\u003eSecond, the ceiling problem: computational results support specific claims about computational observables. \u0026ldquo;Luteolin shows a mean MM-GBSA ΔG of \u0026minus;\u0026thinsp;31 kcal/mol under the stated protocol\u0026rdquo; is a claim the data support. \u0026ldquo;Luteolin inhibits IKKβ\u0026rdquo; is not. The distance between those two statements is the distance between what the experiment measures and what the claim asserts. That distance is routinely collapsed. The underlying limitations of end-point MM-GBSA scoring are well-characterised [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThird, the branch-closure problem: computational screens generate both positive and negative results. Positive results appear in papers. Negative results are silently absent. A paper that reports two confirmed nodes and omits the two failed branches presents a biased picture of the evidence. A reader cannot assess the false discovery rate of the pipeline from the published results alone.\u003c/p\u003e \u003cp\u003eFourth, the provenance problem: a ΔG value in a manuscript should be traceable to the raw output file it came from. In practice, values are often transferred through summary spreadsheets, re-keyed, rounded, or averaged by rules not stated in the methods. The number in the paper may not match the number in the output file.\u003c/p\u003e \u003cp\u003eFifth, the single-trajectory problem: a single 5 ns MD run produces a single ΔG estimate. That estimate has a variance that is not visible from the run itself. Without independent velocity-seed replicates and a pre-specified convergence criterion, the estimate cannot be called reproducible \u0026mdash; it can only be called observed. The value of independent replicates for obtaining consistent binding-affinity predictions has been demonstrated directly in the alchemical-free-energy literature [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis paper describes a platform that addresses all five problems. It does not address them by stricter statistical methods alone, but by building governance into the experimental pipeline: pre-specified gates, outcome-keyed ceiling templates, explicit branch closure, SHA256 provenance chains, and a retrieval-first knowledge graph that requires every claim to be grounded in a queryable evidence record.\u003c/p\u003e \u003cp\u003eThe platform was applied to nine quantified fig and olive phytochemicals (eleven docked; two excluded by force-field failure before MD) across two molecular targets: the COX-2 cyclooxygenase (Node 1) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and the IKKβ kinase ATP pocket (Node 2, with a generalization test at Node 3) [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Four experimental branches were run at the MD stage. One failed on first attempt and was rescued by a pre-registered deep-pocket reseed (Node 2). One generalized the Node 2 result to a second flavonoid (Node 3). One failed and was closed without any upward claim revision (Node 4). These outcomes are reported in the order they occurred, with the negative results held to the same evidential standard as the positive ones.\u003c/p\u003e \u003cp\u003eThe scientific contribution of this work is the platform, not the compound ranking. The compound results are evidence that the platform operates correctly \u0026mdash; including when it records failure.\u003c/p\u003e"},{"header":"2. Platform Architecture","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Overview\u003c/h2\u003e \u003cp\u003eThe platform runs four sequential stages for each experimental branch: (1) structure preparation, docking, and MD simulation; (2) MM-GBSA binding free energy estimation; (3) gate evaluation against pre-specified criteria; and (4) verdict formatting, ceiling enforcement, and knowledge graph ingestion. Each stage produces auditable output. No stage consumes data from the preceding stage without a documented provenance link. The overall pipeline is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Pre-specified gates\u003c/h2\u003e \u003cp\u003eFor each experimental branch, a decision sheet is created and frozen before any simulation is run. The decision sheet specifies four gate criteria: ligand RMSD stability (\u0026lt;\u0026thinsp;3 \u0026Aring; over \u0026gt;\u0026thinsp;80% of the equilibrated trajectory in \u0026ge;\u0026thinsp;3/N seeds), pocket contact occupancy (\u0026ge;\u0026thinsp;5/9 pre-specified pocket residues at \u0026ge;\u0026thinsp;50% occupancy in \u0026ge;\u0026thinsp;3/N seeds), MM-GBSA direction (mean ΔG\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol), and MM-GBSA reproducibility (coefficient of variation\u0026thinsp;\u0026lt;\u0026thinsp;15% across N\u0026thinsp;\u0026ge;\u0026thinsp;3 converged replicates). The threshold values and the outcome key for each possible gate combination are written into the sheet before results arrive.\u003c/p\u003e \u003cp\u003eThe gate evaluator (gate_evaluate.py) reads raw contact occupancy files and FINAL_RESULTS_DECOMP.dat output from gmx_MMPBSA [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], applies the pre-specified criteria, and emits an outcome_key from a fixed vocabulary: ALL_PASS, CV_FAIL_ONLY, or POSE_UNSTABLE. The outcome_key is written to gate_verdict.json and is the sole input to the verdict formatter. The gate evaluator contains no logic that adjusts criteria based on results; it reads the pre-specified thresholds from the decision sheet verbatim. A complete summary of gate outcomes across all four experimental branches is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eGate evaluation summary across all four experimental branches. Gate criteria pre-specified in named decision sheets before any simulation ran. Thresholds: Gate 1 LIG RMSD\u0026thinsp;\u0026lt;\u0026thinsp;3 \u0026Aring; \u0026gt;80% stable in \u0026ge;\u0026thinsp;2/3 (or \u0026ge;\u0026thinsp;3/4) seeds; Gate 2\u0026thinsp;\u0026ge;\u0026thinsp;5/9 pocket residues\u0026thinsp;\u0026ge;\u0026thinsp;50% occupancy in \u0026ge;\u0026thinsp;2/3 seeds; Gate 3 mean ΔG\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol; Gate 4 CV\u0026thinsp;\u0026lt;\u0026thinsp;15%.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode / Run\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGate 1 (RMSD)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGate 2 (Contacts)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGate 3 (ΔG dir.)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eGate 4 (CV)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eoutcome_key\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 1 \u0026mdash; COX-2 panel (N\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePASS (all)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePASS (all)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePASS (all \u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePASS (all \u0026lt;\u0026thinsp;15%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 2 initial \u0026mdash; Luteolin IKKβ (N\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePASS (3/4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePASS (4/4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePASS (\u0026minus;\u0026thinsp;28.18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFAIL (16.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCV_FAIL_ONLY\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 2 reseed \u0026mdash; Luteolin IKKβ dp (N\u0026thinsp;=\u0026thinsp;3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePASS (3/3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePASS (3/3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePASS (\u0026minus;\u0026thinsp;31.457)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePASS (6.85%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 3 \u0026mdash; Quercetin IKKβ (N\u0026thinsp;=\u0026thinsp;3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePASS (2/3)ᵃ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePASS (3/3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePASS (\u0026minus;\u0026thinsp;24.603)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePASS (3.04%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 4 \u0026mdash; Quercetin p65 (N\u0026thinsp;=\u0026thinsp;3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFAIL (0/3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFAIL (0/3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePASS (\u0026minus;\u0026thinsp;12.357)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFAIL (55.85%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePOSE_UNSTABLE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eᵃ s2222 borderline (34.4% stable); gate passes at 2/3 threshold.\u003c/em\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Claim ceiling enforcement\u003c/h2\u003e \u003cp\u003eEach outcome_key maps to a verdict template in verdict_templates.json (and per-node variants for Nodes 3\u0026ndash;4). The template specifies the exact language permitted for the corresponding outcome: numerical slots filled from gate_verdict.json, ceiling phrase drawn from a fixed list, and a NOT SAFE list of forbidden phrases. The formatter (format_verdict.py) populates the template deterministically. It cannot paraphrase ceiling language, round values, or omit required fields.\u003c/p\u003e \u003cp\u003eEach template enforces a four-column structure: what was observed (ΔG, CV, gate outcome), what may be claimed (the pre-specified ceiling phrase), what may not be claimed (the NOT SAFE list), and what condition would permit an upgrade (the unlock gate). This table is the operative output of the pipeline \u0026mdash; not the ΔG value alone.\u003c/p\u003e \u003cp\u003eA constitutional review (NemoClaw) scans every formatted verdict for language on the forbidden list before the verdict is filed. The forbidden list includes phrases such as \u0026ldquo;confirms binding,\u0026rdquo; \u0026ldquo;validates binding,\u0026rdquo; \u0026ldquo;strong binder,\u0026rdquo; and \u0026ldquo;Boltz confirms,\u0026rdquo; which assert conclusions the computational data do not support. If a forbidden phrase is detected, the verdict is rejected and the error is logged before any archiving occurs.\u003c/p\u003e \u003cp\u003eThe ceiling templates and the forbidden-phrase list are frozen artifacts (chmod 444, SHA256-locked in FREEZE_MANIFEST.json). They cannot be modified by the formatter, the knowledge graph ingestion pipeline, or the self-improving training loop without a deliberate human action on the file system.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Oracle and SHA256 provenance\u003c/h2\u003e \u003cp\u003eA regression oracle (verify_node1.py, 46 checks) locks the Node 1 result set. The oracle verifies that every ΔG value in the paper matches the ΔTOTAL line in the corresponding FINAL_RESULTS_DECOMP.dat file, that every SHA256 hash in the training pool matches the file on disk, and that no governance file has been modified since its freeze date. The oracle is run before any claim is published and after any change to the artifact set. Failure on any of the 46 checks halts the pipeline.\u003c/p\u003e \u003cp\u003eThe SHA256 chain runs from raw output file to published claim. For each training cycle, verify_training_pool.py extracts ΔTOTAL values directly from FINAL_RESULTS_DECOMP.dat using a regex that matches the Unicode Δ character in the output format, recomputes the mean and CV from scratch, and compares the recomputed values to those stored in gate_verdict.json. Discrepancies fail the cycle-admission check. No value enters the training corpus or the knowledge graph without passing this extraction-and-recompute verification step.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Retrieval-first knowledge graph\u003c/h2\u003e \u003cp\u003eAll compound results, gate verdicts, claim ceilings, and literature evidence are ingested as structured triples into a knowledge graph (kg_data/kg_export.json). The agent operating the pipeline follows a retrieval-first behavioral rule: it must query the knowledge graph before answering any question about a compound, target, or ΔG value. Memory-only responses are not permitted for questions that can be grounded in the graph. This rule prevents the agent from reporting a value from a prior conversation that may have since been corrected by a raw-data audit.\u003c/p\u003e \u003cp\u003eThe knowledge graph is cross-referenced against the primary literature via PubMed. Ten claims about IKKβ flavonoid binding were verified against PubMed search results (crosscheck_report.json); all ten were supported by at least one citable source [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Literature evidence for each compound\u0026ndash;target pair is stored as a corpus document (kg_data/corpus/) and cited in the relevant decision sheet.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Self-improving verdict formatter\u003c/h2\u003e \u003cp\u003eThe platform includes a training loop for the verdict formatter component. Closed experimental cycles contribute (gate_verdict.json, verdict_section3.md) instruction pairs to a training pool. Cycle admission requires SHA256-verified raw data, NemoClaw constitutional review of the ground-truth output, and confirmation that the output does not contain language from the forbidden-phrase list. The training pool currently holds three verified cycles.\u003c/p\u003e \u003cp\u003eA baseline evaluation of the current production model (qwen3-coder-next via Ollama) against the three-cycle corpus showed 0/3 cycles pass the ceiling compliance and numerical accuracy checks. A LoRA fine-tune [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] (Qwen/Qwen2.5-7B [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], rank\u0026thinsp;=\u0026thinsp;16, 3 epochs, MAX_SEQ_LEN\u0026thinsp;=\u0026thinsp;4096) was run against the corpus. The adapter produces structurally correct outputs \u0026mdash; correct per-seed ΔG values, no fabricated numbers, reproduced SHA256 provenance \u0026mdash; but does not yet emit the required ceiling phrase or markdown template format, reflecting the insufficient size of the current training corpus (three examples cannot anchor a specific phrase pattern in a 7B model). A human promotion gate prevents any adapter from replacing the production formatter until eval_report.json shows overall_pass\u0026thinsp;=\u0026thinsp;True across all cycles.\u003c/p\u003e \u003cp\u003eThe training loop is an architectural component, not a deployed production system. Its inclusion in this paper documents the design pattern, not a completed training result.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Node 1: COX-2 Binding Free Energy Panel\u003c/h2\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 Protocol and replicate structure\u003c/h2\u003e \u003cp\u003eNine fig and olive phytochemicals were screened against the COX-2 crystal structure (PDB: 5KIR; rofecoxib-bound complex [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]) using GROMACS 2026 [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] with the CHARMM36 protein force field [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] and GAFF2 ligand parameters [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Each compound was run as N\u0026thinsp;=\u0026thinsp;4 independent velocity-seed replicates (seeds: main, s1234, s42, s9999), each for 5 ns production MD following energy minimization, 100 ps NVT, and 100 ps NPT equilibration. MM-GBSA binding free energies were computed using gmx_MMPBSA [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] with igb\u0026thinsp;=\u0026thinsp;2 and saltcon\u0026thinsp;=\u0026thinsp;0.150 M. Celecoxib [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] was included as a clinical COX-2 inhibitor reference, docked and evaluated under identical protocol; it is not the co-crystallized ligand of the 5KIR template.\u003c/p\u003e \u003cp\u003eTwo compounds were excluded prior to analysis: oleuropein (+\u0026thinsp;310 kcal/mol) and rutin (+\u0026thinsp;576 kcal/mol) both produced unphysical positive ΔG values attributable to GAFF2 c6 cross-term failure for glycosidic linkages. This force-field limitation is documented; the compounds are carried as explicit exclusions rather than quietly omitted.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Results\u003c/h2\u003e \u003cp\u003eAll ΔG values are mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD across N\u0026thinsp;=\u0026thinsp;4 replicates, sourced from the ΔTOTAL line of FINAL_RESULTS_DECOMP.dat for each run (SHA256-verified, archived in md_runs/PAPER_RESULTS_FINAL.txt). The full ranked panel is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and visualized in Fig.\u0026nbsp;2.\u003c/p\u003e \u003cp\u003eOleocanthal: ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;61.765\u0026thinsp;\u0026plusmn;\u0026thinsp;1.477 kcal/mol (CV\u0026thinsp;=\u0026thinsp;2.4%, N\u0026thinsp;=\u0026thinsp;4; source: md_runs/Oleocanthal_COX2_v2/FINAL_RESULTS_DECOMP.dat and three replicates). Against the celecoxib reference (ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;38.875\u0026thinsp;\u0026plusmn;\u0026thinsp;1.199 kcal/mol), the differential is ΔΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;22.890 kcal/mol (Welch t\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;24.1, p\u0026thinsp;\u0026lt;\u0026thinsp;\u0026lt;\u0026thinsp;0.001). A CV of 2.4% is consistent with convergence across independent velocity seeds under this protocol. The oleocanthal ΔG is driven primarily by van der Waals interactions (ΔVDW\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;40.2 kcal/mol, ΔEEL\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;14.6 kcal/mol, N\u0026thinsp;=\u0026thinsp;4 aggregate from FINAL_RESULTS_DECOMP.dat), consistent with deep insertion of the aliphatic chain and dialdehyde moiety into the hydrophobic channel, consistent with prior reports of oleocanthal anti-inflammatory activity and olive-polyphenol COX-2 engagement [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. The aggregate electrostatic contribution (ΔEEL\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;14.6 kcal/mol) is consistent with known COX-2 channel contacts; per-residue decomposition was not performed; however, the electrostatic contribution (ΔEEL\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;14.6 kcal/mol) is moderate relative to the dominant van der Waals term (ΔVDW\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;40.2 kcal/mol), arguing against a purely electrostatic artifact. MM-GBSA absolute values are known to overestimate binding strength for some chemotypes [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]; the oleocanthal figure should be interpreted as a within-study comparative rank rather than a physical free energy equivalent to experimental IC₅₀.\u003c/p\u003e \u003cp\u003e \u003cem\u003eFigure 2. MM-GBSA binding free energies for the nine-compound COX-2 panel (PDB 5KIR, rofecoxib-bound complex) plus celecoxib reference. Values are mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD across N\u0026thinsp;=\u0026thinsp;4 independent velocity-seed replicates; all sourced from the ΔTOTAL line of FINAL_RESULTS_DECOMP.dat (SHA256-verified, oracle PASS, verify_node1.py, 2026-03-24). Celecoxib was evaluated as an external reference compound under identical protocol; dashed line\u0026thinsp;=\u0026thinsp;celecoxib reference (\u0026minus;\u0026thinsp;38.875 kcal/mol). Oleuropein and rutin excluded (GAFF2 c6 cross-term failure; documented). Source file: manuscript/figures/Fig.\u0026nbsp;2_mmpbsa_barplot.png. [Figure 2 is supplied separately by the author and placed here.]\u003c/em\u003e \u003c/p\u003e \u003cp\u003eLigstroside D2, the second olive compound, produced ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;32.420\u0026thinsp;\u0026plusmn;\u0026thinsp;2.082 kcal/mol (N\u0026thinsp;=\u0026thinsp;4, ΔΔG\u0026thinsp;=\u0026thinsp;+\u0026thinsp;6.455 kcal/mol relative to celecoxib). It falls below the celecoxib reference but ranks second in the full panel, ahead of all fig-panel compounds.\u003c/p\u003e \u003cp\u003eThe seven fig compounds span approximately 10 kcal/mol among themselves (bergapten\u0026thinsp;\u0026minus;\u0026thinsp;20.320\u0026thinsp;\u0026plusmn;\u0026thinsp;0.972 kcal/mol to hydroxytyrosol\u0026thinsp;\u0026minus;\u0026thinsp;10.342\u0026thinsp;\u0026plusmn;\u0026thinsp;0.564 kcal/mol), with psoralen\u0026thinsp;\u0026minus;\u0026thinsp;18.945\u0026thinsp;\u0026plusmn;\u0026thinsp;1.163, luteolin\u0026thinsp;\u0026minus;\u0026thinsp;16.535\u0026thinsp;\u0026plusmn;\u0026thinsp;0.876, elenolic acid\u0026thinsp;\u0026minus;\u0026thinsp;14.875\u0026thinsp;\u0026plusmn;\u0026thinsp;1.021, and quercetin\u0026thinsp;\u0026minus;\u0026thinsp;14.357\u0026thinsp;\u0026plusmn;\u0026thinsp;1.532 kcal/mol (N\u0026thinsp;=\u0026thinsp;4, with s9999 outlier disclosed in the companion paper) in between. All nine quantified compounds show negative ΔG values at the identified COX-2 pocket; none except oleocanthal exceeds the celecoxib reference.\u003c/p\u003e \u003cp\u003eBergapten and psoralen, despite their favorable binding energies, carry a structural caveat: both are furanocoumarins with documented CYP450 inhibitory and phototoxic properties [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The computational result does not address these safety liabilities; they are noted here to prevent any unqualified advancement claim.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.1.3 Ceiling and two-paper disposition\u003c/h2\u003e \u003cp\u003eGate evaluation (gate_evaluate.py, thresholds from NODE1_DECISION_SHEET.md): mean ΔG\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol \u0026mdash; all nine quantified compounds pass; CV\u0026thinsp;\u0026lt;\u0026thinsp;15% \u0026mdash; all nine pass; ligand RMSD\u0026thinsp;\u0026lt;\u0026thinsp;3 \u0026Aring; over \u0026gt;\u0026thinsp;80% of equilibrated trajectory \u0026mdash; all pass; pocket contact occupancy\u0026thinsp;\u0026ge;\u0026thinsp;5/9 residues at \u0026ge;\u0026thinsp;50% \u0026mdash; all pass. outcome_key\u0026thinsp;=\u0026thinsp;ALL_PASS for the full panel (oracle verify_node1.py, 46 checks, PASS, last run 2026-03-24). Per-compound values are tabulated in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe authoritative ceiling language for Node 1:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003e\u0026ldquo;Oleocanthal demonstrates the strongest computed COX-2 binding affinity among screened fig and olive phytochemicals (ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;61.8 kcal/mol, N\u0026thinsp;=\u0026thinsp;4, CV\u0026thinsp;=\u0026thinsp;2.4%), substantially exceeding celecoxib (ΔΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;22.9 kcal/mol, Welch t\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;24.1, p\u0026thinsp;\u0026lt;\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Results are computationally reproducible under the stated protocol. Wet-lab and safety characterization required before any therapeutic claim.\u0026rdquo;\u003c/em\u003e \u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe Node 1 results were split into two companion papers at submission: Paper 1 (OLIVE) covers oleocanthal and ligstroside D2; Paper 2 (FIG) covers the seven remaining fig compounds. The split reflects the natural compound origin and does not alter the gate outcome. Node 1 is locked; the result set is protected by a 46-check regression oracle (verify_node1.py) that verifies every ΔG value against the raw FINAL_RESULTS_DECOMP.dat files and confirms no governance artifact has been modified since its freeze date.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMM-GBSA binding free energies for the COX-2 panel (PDB 5KIR, rofecoxib-bound complex; GROMACS 2026, CHARMM36/GAFF2, igb\u0026thinsp;=\u0026thinsp;2, saltcon\u0026thinsp;=\u0026thinsp;0.150 M). Celecoxib included as external reference compound (not co-crystallized in 5KIR template). All values from ΔTOTAL line of FINAL_RESULTS_DECOMP.dat; SHA256-verified, oracle PASS (verify_node1.py, 2026-03-24). ΔΔG\u0026thinsp;=\u0026thinsp;compound\u0026thinsp;\u0026minus;\u0026thinsp;celecoxib. CV\u0026thinsp;=\u0026thinsp;SD / |mean| \u0026times; 100.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCompound\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean ΔG (kcal/mol)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSD\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCV (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eΔΔG vs celecoxib\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNote\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOleocanthal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;61.765\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.477\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026minus;22.890\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eOLIVE paper\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCelecoxib\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;38.875\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.199\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ereference\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eClinical comparator\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLigstroside D2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;32.420\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.082\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;6.455\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eOLIVE paper\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBergapten\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;20.320\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.972\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;18.555\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFIG paper; furanocoumarin caveat\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePsoralen\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;18.945\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.163\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;19.930\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFIG paper; furanocoumarin caveat\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLuteolin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;16.535\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.876\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;22.340\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFIG paper\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eElenolic acid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;14.875\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;24.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFIG paper\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuercetin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;14.357\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.532\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e10.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;24.518\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFIG paper; s9999 outlier disclosed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHydroxytyrosol\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;10.342\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.564\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;28.533\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFIG paper\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOleuropein\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEXCLUDED\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eGAFF2 failure (+\u0026thinsp;310 kcal/mol)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRutin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEXCLUDED\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eGAFF2 failure (+\u0026thinsp;576 kcal/mol)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Node 2: IKKβ\u0026thinsp;\u0026times;\u0026thinsp;Luteolin\u003c/h2\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Initial run and gate evaluation\u003c/h2\u003e \u003cp\u003eFour independent production simulations of the luteolin\u0026ndash;IKKβ complex (PDB 4KIK [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]) were run from distinct velocity seeds (s1234, s42, s9999, s_rerun), each for 5 ns. One additional trajectory (prod_main) was excluded from all analysis due to a concurrent-write file corruption detected at t\u0026thinsp;=\u0026thinsp;60 ps; its exclusion is documented in the Node 2 receipt trail and did not affect the four-replicate evidence base.\u003c/p\u003e \u003cp\u003eGate evaluation proceeded against four pre-specified criteria frozen in the decision sheet before any result was seen (NODE2_DECISION_SHEET.md, frozen 2026-03-21). Three of four gates passed. Gate 1 (ligand RMSD\u0026thinsp;\u0026lt;\u0026thinsp;3 \u0026Aring; over \u0026gt;\u0026thinsp;80% of t\u0026thinsp;=\u0026thinsp;500\u0026ndash;5000 ps): 3/4 runs pass (s1234 99.9% of frames stable, s9999 98.2%, s_rerun 100%; s42 fails at 43.1% stable). Gate 2 (\u0026ge;\u0026thinsp;5/9 ATP-pocket residues at \u0026ge;\u0026thinsp;50% contact occupancy): 4/4 pass (s1234 8/9, s42 5/9, s9999 8/9, s_rerun 6/9). Gate 3 (mean ΔG\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol): pass, mean ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;28.18 kcal/mol across N\u0026thinsp;=\u0026thinsp;4. Gate 4 (CV\u0026thinsp;\u0026lt;\u0026thinsp;15%): fail, CV\u0026thinsp;=\u0026thinsp;16.1% (SD\u0026thinsp;=\u0026thinsp;4.55 kcal/mol over the range\u0026thinsp;\u0026minus;\u0026thinsp;23.30 to \u0026minus;\u0026thinsp;32.41 kcal/mol). The per-seed distribution is shown in the left panel of Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe platform emitted outcome_key\u0026thinsp;=\u0026thinsp;CV_FAIL_ONLY and applied the pre-specified ceiling for that outcome class: \u0026ldquo;Luteolin remains a computationally credible but not yet reproducible IKKβ hypothesis.\u0026rdquo; No upward revision was permitted. The safe claim was written, filed, and the node was held at \u0026ldquo;computationally credible.\u0026rdquo;\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 Diagnosis\u003c/h2\u003e \u003cp\u003eThe CV failure prompted structural investigation rather than quiet retry. The N\u0026thinsp;=\u0026thinsp;4 ΔG distribution is visibly bimodal: two replicates (s1234\u0026thinsp;\u0026minus;\u0026thinsp;23.30, s42\u0026thinsp;\u0026minus;\u0026thinsp;25.32 kcal/mol) cluster near \u0026minus;\u0026thinsp;24 kcal/mol, while two (s9999\u0026thinsp;\u0026minus;\u0026thinsp;32.41, s_rerun\u0026thinsp;\u0026minus;\u0026thinsp;31.67 kcal/mol) cluster near \u0026minus;\u0026thinsp;32 kcal/mol. Contact occupancy data pointed to the mechanism: s42 undergoes a pose transition at approximately t\u0026thinsp;=\u0026thinsp;1700 ps, after which LYS44, GLU61, and MET96 occupancies fall below 50%. The replicate does not leave the pocket but shifts from deep-pocket engagement to a shallower orientation. s1234, despite maintaining contacts throughout, shows weaker MM-GBSA values than s9999/s_rerun, suggesting the two sub-populations represent distinct binding modes within the same pocket region.\u003c/p\u003e \u003cp\u003eTo test whether structural clustering could resolve the CV issue \u0026mdash; separating the two apparent modes and re-evaluating each independently \u0026mdash; a KMeans analysis (k\u0026thinsp;=\u0026thinsp;2) was run on 363 frames of ligand heavy-atom coordinates after backbone Cα superposition. The silhouette score of 0.53 confirmed two structurally distinguishable populations (Cluster 0: GLU61/MET96 dominant, 37% of frames; Cluster 1: GLY102/ASP166 dominant, 63% of frames). MM-GBSA was re-run within each cluster.\u003c/p\u003e \u003cp\u003eResult: Cluster 0 CV\u0026thinsp;=\u0026thinsp;21.5%, Cluster 1 CV\u0026thinsp;=\u0026thinsp;21.3%. Both values exceed the 15% gate threshold; both exceed the original N\u0026thinsp;=\u0026thinsp;4 CV of 16.1%. Structural stratification did not reduce CV. The data are consistent with energy variance driven by trajectory phase \u0026mdash; the timing of the s42 pose transition \u0026mdash; rather than by structural mode. The two binding mode families are not energetically separable by this cluster partition. This does not prove the pose transition is the sole source of variance; it establishes that the cluster separation does not resolve it.\u003c/p\u003e \u003cp\u003eThis diagnosis was documented in the decision sheet and filed before the reseed decision was made. The gate verdict remained unchanged: CV_FAIL_ONLY. The ceiling was not revised.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e3.2.3 Pre-specified upgrade gate\u003c/h2\u003e \u003cp\u003eA deep-pocket reseed was motivated by the diagnosis: if the shallow-pocket geometry allows pose drift (as observed in s42), constraining the starting geometry to the deep-pocket binding family may eliminate the bimodal distribution and resolve the CV failure. Before any reseed trajectory was run, a four-criterion upgrade gate was locked in the decision sheet (NODE2_UPGRADE_RECORD.md). All four criteria were required simultaneously; any single failure would leave the ceiling unchanged:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eLIG RMSD: \u0026ge;3/N seeds\u0026thinsp;\u0026lt;\u0026thinsp;3 \u0026Aring; sustained over \u0026gt;\u0026thinsp;80% of t\u0026thinsp;=\u0026thinsp;500\u0026ndash;5000 ps\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eContact occupancy: \u0026ge;3/N seeds\u0026thinsp;\u0026ge;\u0026thinsp;5/9 ATP-pocket residues at \u0026ge;\u0026thinsp;50% occupancy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eMM-GBSA direction: mean ΔG\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eMM-GBSA spread: CV\u0026thinsp;\u0026lt;\u0026thinsp;15% across N\u0026thinsp;\u0026ge;\u0026thinsp;3 converged runs\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe upgrade gate follows the same four-criterion structure as the original gate. It is not a relaxed standard \u0026mdash; it is the same standard applied to a new experiment. This distinction matters: the reseed is not a do-over of the original run with lower expectations; it is a new, pre-registered experiment testing a specific structural hypothesis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e3.2.4 Deep-pocket reseed (IKKb_LUT_deepPocket_v1)\u003c/h2\u003e \u003cp\u003eThree independent seeds (s_dp1, s_dp2, s_dp3) were run from the deep-pocket starting geometry for 5 ns each. All four upgrade gate criteria passed.\u003c/p\u003e \u003cp\u003eGate 1 (LIG RMSD): 3/3 seeds stable (s_dp1: 1.56 \u0026Aring; mean, s_dp2: 1.66 \u0026Aring;, s_dp3: 1.56 \u0026Aring;; all \u0026gt;\u0026thinsp;80% frames\u0026thinsp;\u0026lt;\u0026thinsp;3 \u0026Aring;). Gate 2 (contact occupancy): 3/3 seeds\u0026thinsp;\u0026ge;\u0026thinsp;5/9 residues (s_dp1 7/9, s_dp2 5/9, s_dp3 6/9). Gate 3 (ΔG direction): mean ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;31.457 kcal/mol. Gate 4 (CV): 2.156/31.457\u0026thinsp;=\u0026thinsp;6.85%, well within the 15% threshold.\u003c/p\u003e \u003cp\u003eIndividual replicate values (from SHA256-verified FINAL_RESULTS_DECOMP.dat files): s_dp1\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;29.05, s_dp2\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;32.11, s_dp3\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;33.21 kcal/mol. All three seeds maintain deep-cavity contact (Mode B orientation) throughout, with no pose transitions observed. The bimodal distribution from the original N\u0026thinsp;=\u0026thinsp;4 run is absent. The contrast between the initial run and the reseed is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe upgrade gate was met in full. Per the pre-specified rule, the ceiling was upgraded to \u0026ldquo;computationally reproducible.\u0026rdquo; The authoritative ceiling language:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eLuteolin demonstrates computationally reproducible binding to the IKKβ deep-cavity ATP pocket (mean ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;31.457 kcal/mol, CV\u0026thinsp;=\u0026thinsp;6.85%, N\u0026thinsp;=\u0026thinsp;3, ALL_PASS). Evidence tier: EXPLORATORY DDT. Safety characterization required before advancing.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe upgrade is narrow in scope: it applies to the deep-pocket geometry and Mode B binding family. It does not assert that luteolin inhibits IKKβ in vitro or in vivo [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], and it does not apply to alternative binding orientations that were not tested.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e3.2.5 What the Node 2 narrative demonstrates\u003c/h2\u003e \u003cp\u003eNode 2 illustrates the full platform cycle in a single experiment: initial failure captured correctly (CV_FAIL_ONLY), ceiling applied without revision, structural diagnosis performed, negative result of the cluster analysis filed before any further decision was made, upgrade gate locked, reseed run, and ceiling upgraded only after all criteria met. Every step is auditable through the decision sheet, receipt trail, and SHA256-verified raw data. No intermediate result was discarded.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Node 3: IKKβ\u0026thinsp;\u0026times;\u0026thinsp;Quercetin (generalization test)\u003c/h2\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1 Rationale and design\u003c/h2\u003e \u003cp\u003eThe deep-pocket reseed result for luteolin (Node 2) established a specific claim: luteolin binds the IKKβ deep-cavity ATP pocket reproducibly in the Mode B orientation. The natural next question was whether this result was ligand-specific or whether the pocket supports a class of structurally related flavonoids [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Quercetin is the structurally closest compound in the fig panel to luteolin, differing by a single hydroxyl group at the 3-position of the C-ring. If the IKKβ deep-cavity pocket is a generalizable feature, quercetin should reproduce under the same starting geometry and gate criteria. If quercetin fails, the luteolin result stands as a single-compound observation and the generalization claim is not supported.\u003c/p\u003e \u003cp\u003eNode 3 was run as a pre-specified generalization test \u0026mdash; not an exploratory screen \u0026mdash; with the same four gate criteria applied, the same starting geometry (deep-cavity), and N\u0026thinsp;=\u0026thinsp;3 independent seeds (s1111, s2222, s3333) for 5 ns each.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2 Results\u003c/h2\u003e \u003cp\u003eIndividual replicate ΔG values (ΔTOTAL, SHA256-verified FINAL_RESULTS_DECOMP.dat): s1111\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;24.47 kcal/mol, s2222\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;23.93 kcal/mol, s3333\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;25.41 kcal/mol. Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;24.603\u0026thinsp;\u0026plusmn;\u0026thinsp;0.749 kcal/mol, CV\u0026thinsp;=\u0026thinsp;3.04%.\u003c/p\u003e \u003cp\u003eGate evaluation: Gate 1 (ligand RMSD) \u0026mdash; s1111: 1.893 \u0026Aring; mean, 99.4% stable (PASS); s2222: 3.191 \u0026Aring; mean, 34.4% stable (FAIL, borderline); s3333: 2.016 \u0026Aring; mean, 93.2% stable (PASS). Gate passes at 2/3 threshold; s2222 is a disclosed borderline case that a fourth seed could affect. Gate 2 (contact occupancy): s1111 6/9 PASS, s2222 5/9 PASS, s3333 7/9 PASS; 3/3 pass at the \u0026ge;\u0026thinsp;5/9 threshold. Gate 3 (ΔG direction): mean\u0026thinsp;\u0026minus;\u0026thinsp;24.603 kcal/mol, PASS (\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol). Gate 4 (CV): 3.04%, PASS (\u0026lt;\u0026thinsp;15%). outcome_key\u0026thinsp;=\u0026thinsp;ALL_PASS.\u003c/p\u003e \u003cp\u003eBinding mode distribution (from contact occupancy timeline, analysis_scripts/contact_analysis.py): all three quercetin seeds sample both Mode A (hinge-region contact) and Mode B (deep-cavity contact) in a 2:1 B:A ratio over the 5 ns trajectories. The per-mode mean ΔG difference is 1.2 kcal/mol, less than RT at physiological temperature (\u0026asymp;\u0026thinsp;0.6 kcal/mol at 310 K). The two modes are energetically degenerate under this protocol. Luteolin, in contrast, maintained Mode B in all three deep-pocket seeds with no observed transitions. The two-flavonoid comparison is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e and tabulated in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe structural basis for the mode difference between luteolin and quercetin was not determined by this analysis. The additional 3-OH group on quercetin's C-ring is one candidate explanation; no structural simulation specifically testing this hypothesis was run. The observation is descriptive: quercetin samples two modes, luteolin samples one.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e3.3.3 Boltz-2 orthogonal check\u003c/h2\u003e \u003cp\u003eAn orthogonal structure prediction was run for quercetin at the IKKβ pocket using Boltz-2. The pre-specified proximity gate required the predicted ligand centroid to fall within 8 \u0026Aring; of the ATP-pocket anchor residue. The measured centroid distance was 10.1 \u0026Aring; from the pre-specified anchor (INCOMPATIBLE). In parallel, 8 of the 9 pre-specified ATP-pocket contact residues fell within 5 \u0026Aring; of the predicted ligand centroid (CONSISTENT). The Boltz model pLDDT was 84 (INTERPRETABLE \u0026mdash; above the confidence threshold). A summary of the Boltz-2 evaluation is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe governance rule applied to Boltz-2 results is explicit: an orthogonal structural prediction cannot rescue a failed gate or upgrade a ceiling. It also cannot downgrade a passed gate. The Boltz result is recorded as an orthogonal observation: the centroid is shifted relative to the MD-derived pocket anchor, but the contact residue overlap is substantial. This is consistent with a slightly displaced orientation within the same pocket region, not with absence of pose. The MD ceiling was not altered by the Boltz result in either direction.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e3.3.4 Two-flavonoid IKKβ result\u003c/h2\u003e \u003cp\u003eThe combination of Node 2 (luteolin, CV\u0026thinsp;=\u0026thinsp;6.85%, Mode B exclusively) and Node 3 (quercetin, CV\u0026thinsp;=\u0026thinsp;3.04%, Modes A\u0026thinsp;+\u0026thinsp;B degenerate) supports the following bounded claim, which is the highest claim the data support:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003e\u0026ldquo;Both fig-side flavonoids tested reproduce computationally at the IKKβ deep-cavity ATP pocket under identical starting conditions. Mode selection differs: luteolin locks Mode B; quercetin samples both modes with energetic degeneracy. The IKKβ deep-cavity pocket is computationally viable for flavonoid binding. Starting-pose dependence is confirmed; alternative starting geometries may yield different mode selections.\u0026rdquo;\u003c/em\u003e \u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eIKKβ two-flavonoid comparison (PDB 4KIK, deep-cavity starting geometry). Node 2\u0026thinsp;=\u0026thinsp;luteolin (IKKb_LUT_deepPocket_v1); Node 3\u0026thinsp;=\u0026thinsp;quercetin (Quercetin_IKKb_v1). Per-seed ΔG from SHA256-verified FINAL_RESULTS_DECOMP.dat (training_pool/cycle_1\u0026ndash;2).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNode 2: Luteolin\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNode 3: Quercetin\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExperiment ID\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIKKb_LUT_deepPocket_v1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQuercetin_IKKb_v1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN seeds\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (s_dp1, s_dp2, s_dp3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3 (s1111, s2222, s3333)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePer-seed ΔG (kcal/mol)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;29.05, \u0026minus;\u0026thinsp;32.11, \u0026minus;\u0026thinsp;33.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026minus;24.47, \u0026minus;\u0026thinsp;23.93, \u0026minus;\u0026thinsp;25.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean ΔG\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026minus;31.457\u0026thinsp;\u0026plusmn;\u0026thinsp;2.154\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026minus;24.603\u0026thinsp;\u0026plusmn;\u0026thinsp;0.749\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCV (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6.85 (PASS\u0026thinsp;\u0026lt;\u0026thinsp;15%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.04 (PASS\u0026thinsp;\u0026lt;\u0026thinsp;15%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGate 1 (RMSD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3/3 PASS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2/3 PASS (s2222 borderline)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGate 2 (contacts)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3/3 PASS (7/9, 5/9, 6/9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3/3 PASS (6/9, 5/9, 7/9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGate 3 (ΔG direction)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePASS (\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePASS (\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGate 4 (CV)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePASS (6.85%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePASS (3.04%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eoutcome_key\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBinding mode\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMode B exclusively (3/3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModes A\u0026thinsp;+\u0026thinsp;B, 2:1 B:A (degenerate)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBoltz-2 centroid (\u0026Aring;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStalled \u0026mdash; no output\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10.1 \u0026Aring; (INCOMPATIBLE)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBoltz-2 contacts\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8/9 within 5 \u0026Aring; (CONSISTENT)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBoltz-2 pLDDT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84 (INTERPRETABLE)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvidence tier\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEXPLORATORY DDT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEXPLORATORY DDT\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSafety tier\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCONCERN (ER agonist, IKK clinical fail) [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCAUTION (hERG and CYP-mediated drug-drug interaction risk reported; see [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] for flavonoid class review)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThis claim does not assert IKKβ inhibition by either compound, does not extend to in vitro or in vivo contexts, and does not assert flavonoid class-level activity. Both nodes are EXPLORATORY DDT tier.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Node 4: p65 \u0026times; Quercetin (honest branch closure)\u003c/h2\u003e \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e \u003ch2\u003e3.4.1 Rationale and hypothesis\u003c/h2\u003e \u003cp\u003eWith quercetin showing computationally reproducible binding at IKKβ, an exploratory hypothesis was formed: if quercetin engages the IKKβ node of the NF-κB pathway [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], does it also engage the pathway downstream at the NF-κB p65 transcription factor? The p65 DNA-binding interface (PDB: 1VKX [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]) was selected as the candidate target. The 1VKX structure is a DNA-bound complex; no validated small-molecule binding pocket is established for this conformation, and no prior crystallographic evidence of small-molecule engagement at this interface exists. This selection was therefore explicitly exploratory \u0026mdash; a structural adjacency test, not a docking target with known druggability. This is explicitly an exploratory branch, not a validated DDT node \u0026mdash; the target is mechanistically adjacent but structurally unrelated to IKKβ, and the hypothesis was formed after, not before, the IKKβ result.\u003c/p\u003e \u003cp\u003eThe decision sheet for Node 4 pre-specified the same four gate criteria as Nodes 2 and 3. The pre-specified outcome envelope for POSE_UNSTABLE was written before any simulation was run (NODE4_VERDICT_ENVELOPES.md): \u0026ldquo;Quercetin does not show reproducible binding at the NF-κB p65 DNA-binding interface under the current gate framework. POSE UNSTABLE \u0026mdash; branch CLOSED.\u0026rdquo; The platform claim update was also pre-specified: if Node 4 fails, the two-node claim is preserved and no Node 5 routing is permitted.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e \u003ch2\u003e3.4.2 Results and closure\u003c/h2\u003e \u003cp\u003eThree independent seeds (s1111, s2222, s3333) were run from the docked starting geometry for 5 ns each. Gate 1 (ligand RMSD): s1111 3.103 \u0026Aring; mean, 51.2% stable (FAIL); s2222 11.967 \u0026Aring;, 0.0% stable (FAIL); s3333 3.977 \u0026Aring;, 40.4% stable (FAIL); 0/3 pass. Gate 2 (contact occupancy): 0/3 seeds reach\u0026thinsp;\u0026ge;\u0026thinsp;5/9 interface residues; maximum observed was 1/9 per seed. Gate 3 (ΔG direction): mean\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;12.357 kcal/mol; PASS (\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;10 kcal/mol). Gate 4 (CV): CV\u0026thinsp;=\u0026thinsp;55.85%; FAIL. Failing gates: 1, 2, 4. outcome_key\u0026thinsp;=\u0026thinsp;POSE_UNSTABLE.\u003c/p\u003e \u003cp\u003eThe pre-specified ceiling was applied without modification:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003e\u0026ldquo;Quercetin does not show reproducible binding at the NF-κB p65 DNA-binding interface under the current gate framework. POSE UNSTABLE \u0026mdash; pose drift observed across replicates, insufficient for credibility claim. p65 branch CLOSED. No Node 5 routing permitted.\u0026rdquo;\u003c/em\u003e \u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe platform claim was updated accordingly: \u0026ldquo;Two-node model (COX-2\u0026thinsp;+\u0026thinsp;IKKβ). p65 branch closed. No cascade evidence.\u0026rdquo; This is a bounded update \u0026mdash; the two-node model is not weakened by the Node 4 closure; it is correctly scoped by it. Gate 3 passing does not alter the verdict: a favorable ΔG direction mean without pose stability or pocket contact is insufficient evidence for a credibility claim under the pre-specified gate structure.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section3\"\u003e \u003ch2\u003e3.4.3 What honest closure demonstrates\u003c/h2\u003e \u003cp\u003eThe value of Node 4 is not its scientific content \u0026mdash; pose instability at p65 does not advance the mechanistic understanding of quercetin. Its value is architectural: it demonstrates that the platform closes a branch under the same gate framework that produces positive results in other branches.\u003c/p\u003e \u003cp\u003eA platform that closes branches honestly is one whose positive results carry evidential weight. If Node 2 and Node 3 had both passed while Node 4 failed but was never documented, a reader of the final publication would have no way to assess the false discovery rate of the pipeline. By documenting Node 4 \u0026mdash; with the pre-specified envelope, the gate failures, and the explicit update to the platform claim \u0026mdash; the pipeline provides the information needed to evaluate the positive results in context.\u003c/p\u003e \u003cp\u003eThe Node 4 receipt trail and verdict are archived as frozen artifacts in the same artifact set as the positive nodes. Node 4 is not a footnote; it is a first-class result of the wrong kind. A cross-node summary of all four experimental branches is presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCross-node summary: all four experimental branches. All values from SHA256-verified FINAL_RESULTS_DECOMP.dat or gate_verdict.json. Claim ceilings locked in verdict_templates.json before verdicts were written.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode / Run\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTarget\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCompound\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMean ΔG (kcal/mol)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCV (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eGate outcome\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eBranch decision\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCOX-2 (5KIR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOleocanthal (lead)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e\u0026minus;61.765\u0026thinsp;\u0026plusmn;\u0026thinsp;1.477\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003ePROMOTE (published)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 2 initial\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIKKβ (4KIK)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLuteolin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e\u0026minus;28.18\u0026thinsp;\u0026plusmn;\u0026thinsp;4.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCV_FAIL_ONLY\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eARCHIVE \u0026rarr; reseed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 2 reseed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIKKβ (4KIK)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLuteolin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e\u0026minus;31.457\u0026thinsp;\u0026plusmn;\u0026thinsp;2.154\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADEQUATE \u0026mdash; exploratory DDT\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIKKβ (4KIK)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQuercetin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e\u0026minus;24.603\u0026thinsp;\u0026plusmn;\u0026thinsp;0.749\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eALL_PASS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADEQUATE \u0026mdash; exploratory DDT\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ep65 (1VKX)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQuercetin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e\u0026minus;12.357\u0026thinsp;\u0026plusmn;\u0026thinsp;6.906\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e55.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePOSE_UNSTABLE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKILL \u0026mdash; branch closed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Governed multi-node discovery versus compound ranking\u003c/h2\u003e \u003cp\u003eA standard computational polypharmacology screen [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] produces a ranked list of compounds by binding affinity. This list has interpretive value, but it lacks governance: the ranking does not specify what claims the ΔG values support, does not document which branches failed, and does not constrain the language available to describe the results. The ranked list is a computation; it is not by itself an evidence-governed claim.\u003c/p\u003e \u003cp\u003eThe platform described here produces a ranked list and enforces governance on every entry. For each compound\u0026ndash;target pair, the operative unit is the four-column ceiling table described in Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e2.3\u003c/span\u003e: what was observed, what may be claimed, what may not be claimed, and what condition would permit an upgrade. This table is not commentary on the results; it is the output of the pipeline, generated deterministically from raw data and pre-specified templates.\u003c/p\u003e \u003cp\u003eThe ceiling table constrains what can be claimed from any given ΔG value: the claim boundary is set before results arrive and cannot be revised based on what the numbers turned out to be. It does not constrain the ΔG values themselves; it constrains the language used to describe them.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Failure as a first-class result\u003c/h2\u003e \u003cp\u003eNode 2's initial CV failure (CV\u0026thinsp;=\u0026thinsp;16.1%, gate threshold\u0026thinsp;\u0026lt;\u0026thinsp;15%; gate_verdict.json, NODE2_DECISION_SHEET.md) produced three subsequent analysis steps: a structural diagnosis of the bimodal ΔG distribution, a KMeans cluster analysis (result: CV worsened to 21.5% and 21.3% by cluster \u0026mdash; variance is not mode-separable), and a pre-specified upgrade gate locked before the reseed was executed.\u003c/p\u003e \u003cp\u003eIf CV\u0026thinsp;=\u0026thinsp;16.1% had been described as \u0026ldquo;modest variability\u0026rdquo; \u0026mdash; as it could have been, and as similar values routinely are in the computational polypharmacology literature \u0026mdash; none of that analysis would have been triggered. The diagnosis, the cluster analysis, the upgrade gate, and the Mode B characterization all follow from treating the CV failure as a failure rather than a success with a softened label.\u003c/p\u003e \u003cp\u003eNode 4 is the same principle applied to branch closure. The p65 result is negative. Its documentation prevents the platform from claiming a mechanistic cascade it has not established, and it provides the denominator needed to assess the positive result rate of the pipeline. A system that reports only successes has no interpretable success rate.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e4.3 The rescue pattern as a reusable protocol contribution\u003c/h2\u003e \u003cp\u003eThe Node 2 rescue sequence follows a three-step structure that can be applied to any node producing outcome_key\u0026thinsp;=\u0026thinsp;CV_FAIL_ONLY with a structural diagnostic candidate: (1) diagnose the source of CV variance using structural clustering or contact timeline analysis; (2) formulate a specific, falsifiable structural hypothesis about what change to the starting conditions would eliminate the variance; (3) lock an upgrade gate \u0026mdash; with the same four criteria as the original gate, in a named file, before executing the reseed.\u003c/p\u003e \u003cp\u003eStep 3 is the one that prevents the reseed from becoming a do-over with softened expectations. Step 2 is the one that prevents step 3 from being justified post-hoc. And step 1 requires the original failure to be documented \u0026mdash; which is why a protocol that encourages quiet retry cannot support this pattern. The Node 2 cluster analysis that made the CV worse is not a failed experiment; it is the evidentiary foundation that makes the upgrade gate legitimate.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Boltz-2 as bounded orthogonal evidence\u003c/h2\u003e \u003cp\u003eThe Boltz-2 predictions for quercetin represent a methodological choice about how to handle orthogonal evidence. The governance rule applied is explicit: structure prediction cannot rescue a failed MD gate or upgrade an MD ceiling, and it cannot downgrade a passed gate. This rule is necessary because the two methods measure different things under different assumptions; allowing one to override the other removes the interpretive boundary between complementary observations and post-hoc rationalization.\u003c/p\u003e \u003cp\u003eWithin this constraint, Boltz-2 functions as a consistency check. For quercetin, the centroid distance (10.1 \u0026Aring; from the pre-specified anchor) and the contact residue overlap (8/9 within 5 \u0026Aring;) together describe a prediction shifted from the MD-derived pocket center but substantially overlapping with the MD contact set (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Two independent methods \u0026mdash; one dynamics-based, one structure-prediction-based \u0026mdash; both locate quercetin in the vicinity of the IKKβ ATP-pocket contact residues. Neither method is upgraded or downgraded by that observation; both are reported at their respective evidence tiers.\u003c/p\u003e \u003cp\u003eThe luteolin Boltz-2 prediction stalled due to a CUDA/PyTorch incompatibility and produced no scientific output. This outcome is documented in the artifact set alongside the positive Boltz result. The absence of a Boltz prediction for luteolin does not affect the MD ceiling.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Self-improving formatter: architecture versus deployment\u003c/h2\u003e \u003cp\u003eThe self-improving verdict formatter is included here as an architectural contribution, not as a deployed system. The first complete training run \u0026mdash; Qwen/Qwen2.5-7B [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], LoRA [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] rank\u0026thinsp;=\u0026thinsp;16, three verified cycles \u0026mdash; produced an adapter that outputs correct per-seed ΔG values, includes SHA256 provenance, and generates no fabricated numbers. It does not yet produce the required ceiling phrase or markdown template format: three training examples are insufficient to anchor a specific phrase pattern in a 7B model.\u003c/p\u003e \u003cp\u003eThe adapter produces evidence of learning without meeting the promotion gate (eval_report.json overall_pass\u0026thinsp;=\u0026thinsp;True). It is not promoted. The production formatter remains the rule-based system (format_verdict.py) until the corpus reaches sufficient size for ceiling template anchoring \u0026mdash; estimated 5\u0026ndash;10 cycles, including at least one CV_FAIL_ONLY example. The training loop accumulates each new closed experimental cycle; the next fine-tune trigger will be data-driven, not calendar-driven.\u003c/p\u003e \u003cp\u003eThe architecture is described here so that the pattern is available to other groups building governance-compliant computational pipelines. The current training state is described as it is, so that no reader mistakes the architectural description for a claim about a production-ready system.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec35\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Rational polypharmacology context\u003c/h2\u003e \u003cp\u003eThe experimental design \u0026mdash; two independent molecular targets, multiple compounds tested at each \u0026mdash; reflects the 2026 rational polypharmacology paradigm [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]: multi-target engagement as a feature to be characterized and governed, not incidentally discovered. The COX-2 and IKKβ nodes represent independent anti-inflammatory mechanisms (prostaglandin synthesis [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] and NF-κB activation [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], respectively), and the compound panel includes structures \u0026mdash; luteolin, quercetin \u0026mdash; with prior literature support for activity at both target classes [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. The IKKβ ATP-binding pocket targeted here corresponds to the conserved kinase hinge region and catalytic cleft, which are established binding sites for small-molecule inhibitors [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] and have prior literature support for flavonoid engagement [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe platform's contribution to this context is not the specific compound results but the methodology for generating bounded, auditable, reproducible results at multiple nodes. The highest supported claim of this work is that a computationally reproducible two-node profile exists for select fig and olive phytochemicals at COX-2 and IKKβ, documented under governance-constrained conditions. The appropriate next step is wet-lab validation at each node independently before any cross-node inference is drawn.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Limitations","content":"\u003cdiv id=\"Sec37\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Computational evidence only\u003c/h2\u003e \u003cp\u003eAll binding free energy values in this paper are MM-GBSA estimates from classical force-field molecular dynamics simulations. MM-GBSA is an end-point free energy method with known limitations [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]: it does not account for full conformational entropy of the protein, it uses implicit solvent, and its absolute values are sensitive to the GB model and salt concentration parameters. The values reported here are internally consistent within the stated protocol (igb\u0026thinsp;=\u0026thinsp;2, saltcon\u0026thinsp;=\u0026thinsp;0.150 M) and reproducible across independent replicates, but they are not experimental binding affinities. MM-GBSA ΔG values reported here are not directly comparable to experimental binding free energies (ΔG_exp, \u0026minus;RT\u0026middot;ln(K_i), or IC₅₀ equivalents); rank-ordering within this study under a fixed protocol is valid, but cross-study numerical comparison of absolute values is not. No IC₅₀, K_i, K_d, or surface plasmon resonance data were collected for any compound in this paper.\u003c/p\u003e \u003cp\u003eAll results are classified at the EXPLORATORY DDT (Drug Discovery Tool) tier: appropriate for target selection and hypothesis prioritization, not for therapeutic claims. Advancement beyond this tier requires experimental validation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec38\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Force-field coverage gaps\u003c/h2\u003e \u003cp\u003eThe GAFF2 force field [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] failed for glycosidic linkages in oleuropein and rutin, producing unphysical positive ΔG values (+\u0026thinsp;310 and +\u0026thinsp;576 kcal/mol, respectively). These compounds were excluded with documentation rather than silently dropped. The exclusions do not affect the ranked results for the nine quantified compounds, but they represent a genuine coverage gap: glycoside-containing phytochemicals cannot be reliably evaluated under the current protocol without a force-field extension or alternative parameterization. No instability or unphysical energy terms were observed for oleocanthal or other non-glycosidic compounds, suggesting the GAFF2 limitation is specific to glycosidic linkages rather than a global parameterization issue.\u003c/p\u003e \u003cp\u003eBergapten and psoralen are included in the ranked results with a standing caveat: both are furanocoumarins with documented CYP450 inhibitory and phototoxic liabilities [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The computational binding result at COX-2 is accurate under the stated protocol, but the safety signal is not addressed by any analysis in this paper. Advancement of either compound requires explicit safety characterization prior to any in vitro testing.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec39\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Starting-geometry dependence\u003c/h2\u003e \u003cp\u003eThe Node 2 rescue demonstrated directly that MD results at IKKβ are starting-geometry dependent: the initial run from a shallow-pocket geometry produced a bimodal ΔG distribution (CV\u0026thinsp;=\u0026thinsp;16.1%), while the deep-pocket reseed from a constrained geometry produced a convergent single-mode result (CV\u0026thinsp;=\u0026thinsp;6.85%). The quercetin Node 3 result \u0026mdash; run from the same deep-pocket starting geometry \u0026mdash; sampled both Mode A and Mode B, confirming that mode selection is not fully determined by the starting pose.\u003c/p\u003e \u003cp\u003eResults at IKKβ are therefore conditional on the starting geometry. The two-flavonoid result described in Section \u003cspan refid=\"Sec24\" class=\"InternalRef\"\u003e3.3.4\u003c/span\u003e is valid within the stated starting conditions; it does not characterize the full conformational landscape of either compound at the IKKβ pocket, and alternative starting geometries may produce different mode distributions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec40\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Fixed crystal structure and single-conformation constraint\u003c/h2\u003e \u003cp\u003eEach node in this platform tests a single molecular target with a fixed crystal structure (5KIR for COX-2 (rofecoxib-bound complex) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], 4KIK for IKKβ [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], 1VKX for p65 [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]; all entries retrieved from the PDB [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]). Crystal structures represent a single conformational snapshot; induced fit, allosteric effects, and protein flexibility not captured by the starting structure are not modeled. The gate criteria and ceiling templates are calibrated to the specific structure used; results may not transfer directly to alternative structures of the same target. For COX-2, alternative crystal structures (e.g., celecoxib-bound 3LN1 or arachidonic-acid-bound 1DIY) were not explored; all COX-2 results are conditional on the 5KIR conformational state.\u003c/p\u003e \u003cp\u003eFor IKKβ in particular, the 4KIK structure (655-residue kinase domain, chain A) was used with a residue offset of \u0026minus;\u0026thinsp;8 between PDB numbering and the GRO system used in analysis. This offset is documented in the decision sheet and applied consistently; it is a potential source of numbering confusion when cross-referencing literature that uses PDB-canonical residue numbers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec41\" class=\"Section2\"\u003e \u003ch2\u003e5.5 Training pool size\u003c/h2\u003e \u003cp\u003eThe self-improving verdict formatter has been trained on three experimental cycles. Three cycles are insufficient to anchor the required ceiling phrase pattern in a 7B model; the gap is characterized in Section \u003cspan refid=\"Sec34\" class=\"InternalRef\"\u003e4.5\u003c/span\u003e and represents a genuine limitation on the current state of the formatter component. The formatter in production use is the rule-based system (format_verdict.py), not the adapter; the gap does not affect any result reported in this paper, but the self-improving loop has not yet reached the state where it can substitute for the rule-based formatter without human review.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec42\" class=\"Section2\"\u003e \u003ch2\u003e5.6 Absence of wet-lab validation\u003c/h2\u003e \u003cp\u003eNo compounds in this paper have been tested in vitro or in vivo. There is no biochemical binding data, no cell-based NF-κB reporter assay, no kinase selectivity panel, and no safety characterization against the LTKB or EDKB toxicity databases. All claims are bounded to computationally reproducible under the stated protocol. This is the correct claim for the evidence tier. It is not a temporary limitation that will be removed by additional computation; it is a ceiling that can only be raised by experimental work.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Conclusions","content":"\u003cp\u003eThis paper describes a governed multi-node computational discovery platform and reports its gated outcomes across four experimental branches: a nine-compound COX-2 panel (Node 1, ALL_PASS; Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), luteolin at IKKβ following a pre-registered rescue from an initial CV failure (Node 2, CV_FAIL_ONLY \u0026rarr; ALL_PASS; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e), quercetin at IKKβ as a generalization test (Node 3, ALL_PASS; Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e), and quercetin at p65 as a pre-specified branch closure (Node 4, POSE_UNSTABLE; Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The compound results are bounded to the evidence tier at which they were obtained. They are illustrations of the platform producing correct gated outcomes \u0026mdash; including failure.\u003c/p\u003e \u003cp\u003eThe platform contribution comprises seven interlocking components: pre-specified gates frozen before any simulation runs; outcome-keyed ceiling templates that remove interpretive latitude from the result-to-claim step; a 46-check regression oracle that verifies every published number against the raw output file it came from; a SHA256 provenance chain from raw data to published claim; a retrieval-first knowledge graph that prevents memory-only responses to questions that can be grounded in the evidence record; a NemoClaw constitutional review layer that scans every formatted verdict for forbidden language before archiving; and a self-improving formatter that accumulates each closed experimental cycle as a training example under hash-verified admission criteria. The integrated pipeline is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and the full gate outcomes are in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eEach component targets one of the five failure modes named in Section \u003cspan refid=\"Sec1\" class=\"InternalRef\"\u003e1\u003c/span\u003e: pre-specified gates address post-hoc threshold setting; ceiling templates address claim ceiling collapse; receipt trails and branch closure records address silent omission; SHA256 oracle addresses provenance breaks; N\u0026thinsp;\u0026ge;\u0026thinsp;3 velocity-seed replicates with pre-specified CV gate address single-trajectory reporting. None of these failure modes is addressed by better statistics alone; all require governance built into the pipeline structure.\u003c/p\u003e \u003cp\u003eThe scientific output is a bounded two-node model: oleocanthal at COX-2 (ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;61.8 kcal/mol, CV\u0026thinsp;=\u0026thinsp;2.4%, N\u0026thinsp;=\u0026thinsp;4) and both luteolin (ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;31.5 kcal/mol, CV\u0026thinsp;=\u0026thinsp;6.85%, N\u0026thinsp;=\u0026thinsp;3) and quercetin (ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;24.6 kcal/mol, CV\u0026thinsp;=\u0026thinsp;3.04%, N\u0026thinsp;=\u0026thinsp;3) at the IKKβ deep-cavity ATP pocket. These results are computationally reproducible under the stated protocol and are EXPLORATORY DDT tier. The appropriate next steps are enzymatic binding assays at COX-2 and IKKβ, a kinase selectivity panel for the IKKβ hits, and safety characterization (CYP450 profiling for furanocoumarins [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], LTKB/EDKB screening for all confirmed nodes) before any advancement claim is made.\u003c/p\u003e \u003cp\u003eThe full frozen artifact set \u0026mdash; 28 SHA256-verified files including decision sheets, receipt trails, verdict templates, the regression oracle, and the KG export \u0026mdash; is archived at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.19076569\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.19076569\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and is sufficient to audit and verify all reported results (a representative subset is enumerated in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). The governance layer is not optional; it is structurally embedded in every stage of the experimental cycle. The pipeline implementation is proprietary and not publicly released; the methodology, protocols, and gate framework are described in full in this paper so that the approach can be independently implemented by other groups.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRepresentative frozen artifacts (FREEZE_MANIFEST.json, 2026-03-24). Full set: 28 files. All chmod 444 at freeze. SHA256 hashes in FREEZE_MANIFEST.json (Zenodo: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.19076569\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.19076569\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArtifact\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePath\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRole\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaster outcomes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMASTER_OUTCOMES_FROZEN.md\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAuthoritative ΔG, gate verdicts, ceilings (all nodes)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 2 decision sheet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enode2_IKKb/NODE2_DECISION_SHEET.md\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePre-specified gates; upgrade record; diagnosis trail\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 3 decision sheet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enode2_IKKb/NODE3_DECISION_SHEET.md\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQuercetin IKKβ gates and verdict\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 4 decision sheet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enode4_p65/NODE4_DECISION_SHEET.md\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ep65 branch gates and closure record\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVerdict templates\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epipeline/templates/verdict_templates.json\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCeiling phrases; NOT SAFE lists\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClaim ceiling rules\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epipeline/manifests/claim_ceiling_rules.json\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGovernance rules for ceiling assignment\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegression oracle\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epipeline/oracle/verify_node1.py\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46-check Node 1 oracle (PASS 2026-03-24)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 1 oracle manifest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epipeline/oracle/node1_manifest_v1.json\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExpected values for all 46 oracle checks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRaw results\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003emd_runs/PAPER_RESULTS_FINAL.txt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAll Node 1 ΔG values, individual replicates\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNode 2 receipt trail\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enode2_IKKb/RECEIPT_TRAIL.md\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFull audit trail Node 2 (590 lines)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCycle 1 gate verdict\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epipeline/training_pool/cycle_1/inputs/gate_verdict.json\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSHA256-verified verdict luteolin dp reseed\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCycle 2 gate verdict\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epipeline/training_pool/cycle_2/inputs/gate_verdict.json\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSHA256-verified verdict quercetin IKKβ\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBoltz quercetin sheet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enode2_IKKb/BOLTZ_EVIDENCE_SHEET_QUERCETIN.md\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCentroid 10.1 \u0026Aring;; 8/9 contacts; pLDDT\u0026thinsp;=\u0026thinsp;84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFreeze manifest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFREEZE_MANIFEST.json\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSHA256 of all 28 critical artifacts\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eWhat this paper does not claim: that any compound inhibits its target in vitro or in vivo; that the two-node profile constitutes multi-target synergy; that the training adapter is production-ready; that the computationally reproducible result is transferable to alternative structures of the same target; or that any result in this paper is sufficient to advance a compound beyond the EXPLORATORY DDT tier without experimental validation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eFunding Declaration\u003c/p\u003e\n\u003cp\u003eThis work received no specific grant, contract, or sponsorship from any funding agency in the public, commercial, or not-for-profit sectors. The research was conducted independently by the author as an Independent Researcher, using self-funded computational resources. No external financial support, institutional backing, or industry sponsorship contributed to the design, execution, analysis, or writing of this study.\u003c/p\u003e\n\u003cp\u003eConflicts of Interest\u003c/p\u003e\n\u003cp\u003eThe author declares no conflicts of interest. The pipeline implementation is proprietary (developed solely by the author) but is not commercialized; no financial or professional relationships exist that could be perceived to influence the work reported.\u003c/p\u003e\n\u003cp\u003eData and Code Availability\u003c/p\u003e\n\u003cp\u003eAll frozen artifacts supporting the reported results \u0026mdash; decision sheets, receipt trails, gate verdicts, verdict templates, the regression oracle, and the knowledge-graph export (28 SHA256-verified files) \u0026mdash; are archived at https://doi.org/10.5281/zenodo.19076569. A representative subset is enumerated in Table 5. The pipeline implementation source code is proprietary and not publicly released; the methodology, protocols, and gate framework are described in sufficient detail in Sections 2 and 3 to permit independent re-implementation.\u003c/p\u003e\n\u003cp\u003eAI Use Statement\u003c/p\u003e\n\u003cp\u003eDuring manuscript preparation, generative AI tools were used to assist in drafting, revising, and language editing of human-directed text. All underlying scientific analyses, numerical results, interpretations, and conclusions were independently reviewed and verified by the author, who takes full responsibility for the final content. AI tools were not used to generate results, perform trajectory analysis, evaluate gate criteria, or make scientific decisions; all such work was performed by the author using the computational tools and protocols described in the Methods section. No AI tool is listed as an author.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHuang J, MacKerell AD Jr. (2013) CHARMM36 all-atom additive protein force field: validation based on comparison to NMR data. J Comput Chem. ;34(25):2135\u0026ndash;2145. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/jcc.23354\u003c/span\u003e\u003cspan address=\"10.1002/jcc.23354\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 23832629\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbraham MJ, Murtola T, Schulz R, P\u0026aacute;ll S, Smith JC, Hess B, Lindahl E (2015) GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 1\u0026ndash;2:19\u0026ndash;25 (Tool paper; not indexed in PubMed.)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE (2000) The Protein Data Bank. Nucleic Acids Res. ;28(1):235\u0026ndash;242. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/nar/28.1.235\u003c/span\u003e\u003cspan address=\"10.1093/nar/28.1.235\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 10592235\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O et al (2021) Highly accurate protein structure prediction with AlphaFold. Nature. ;596(7873):583\u0026ndash;589. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41586-021-03819-2\u003c/span\u003e\u003cspan address=\"10.1038/s41586-021-03819-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 34265844\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHopkins AL (2008) Network pharmacology: the next paradigm in drug discovery. Nat Chem Biol. ;4(11):682\u0026ndash;690. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nchembio.118\u003c/span\u003e\u003cspan address=\"10.1038/nchembio.118\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 18936753\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnighoro A, Bajorath J, Rastelli G (2014) Polypharmacology: challenges and opportunities in drug discovery. J Med Chem. ;57(19):7874\u0026ndash;7887. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1021/jm5006463\u003c/span\u003e\u003cspan address=\"10.1021/jm5006463\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 24946140\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGenheden S, Ryde U (2015) The MM/PBSA and MM/GBSA methods to estimate ligand-binding affinities. Expert Opin Drug Discov. ;10(5):449\u0026ndash;461. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1517/17460441.2015.1032936\u003c/span\u003e\u003cspan address=\"10.1517/17460441.2015.1032936\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 25835573\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVald\u0026eacute;s-Tresanco MS, Vald\u0026eacute;s-Tresanco ME, Valiente PA, Moreno E, gmx_MMPBSA: (2021) A New Tool to Perform End-State Free Energy Calculations with GROMACS. J Chem Theory Comput. ;17(10):6281\u0026ndash;6291. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1021/acs.jctc.1c00645\u003c/span\u003e\u003cspan address=\"10.1021/acs.jctc.1c00645\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 34586825\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOrlando BJ, Malkowski MG (2016) Crystal structure of rofecoxib bound to human cyclooxygenase-2. Acta Crystallogr F Struct Biol Commun 72(Pt 10):772\u0026ndash;776. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1107/S2053230X16014230\u003c/span\u003e\u003cspan address=\"10.1107/S2053230X16014230\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003ePMID: 27710942. [PDB 5KIR]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKurumbail RG, Stevens AM, Gierse JK, McDonald JJ, Stegeman RA, Pak JY et al (1996) Structural basis for selective inhibition of cyclooxygenase-2 by anti-inflammatory agents. Nature. ;384(6610):644\u0026ndash;648. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/384644a0\u003c/span\u003e\u003cspan address=\"10.1038/384644a0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 8967954\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen YQ, Ghosh S, Ghosh G (1998) A novel DNA recognition mode by the NF-κB p65 homodimer. Nat Struct Biol 5(1):67\u0026ndash;73. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nsb0198-67\u003c/span\u003e\u003cspan address=\"10.1038/nsb0198-67\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003ePMID: 9437432. [PDB 1VKX]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeauchamp GK, Keast RSJ, Morel D, Lin J, Pika J, Han Q, Lee CH, Smith AB III, Breslin PAS (2005) Phytochemistry: ibuprofen-like activity in extra-virgin olive oil. Nature. ;437(7055):45\u0026ndash;46. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/437045a\u003c/span\u003e\u003cspan address=\"10.1038/437045a\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 16136122\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMontoya T, Castej\u0026oacute;n ML, S\u0026aacute;nchez-Hidalgo M, Gonz\u0026aacute;lez-Benjumea A, Fern\u0026aacute;ndez-Bola\u0026ntilde;os JG, Alarc\u0026oacute;n-de-la-Lastra C (2019) Oleocanthal modulates LPS-induced murine peritoneal macrophages activation via regulation of inflammasome, Nrf-2/HO-1, and MAPKs signaling pathways. J Agric Food Chem 67(19):5552\u0026ndash;5559. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1021/acs.jafc.9b00771.\u003c/span\u003e\u003cspan address=\"10.1021/acs.jafc.9b00771.\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e PMID: 31042377\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang J, Bonvino NP, Hung A, Karagiannis TC (2020) In silico characterisation of olive phenolic compounds as potential cyclooxygenase modulators. Part 1. J Mol Graph Model. ;101:107719. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jmgm.2020.107719\u003c/span\u003e\u003cspan address=\"10.1016/j.jmgm.2020.107719\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 32898836\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaragiannis TC, Ververis K, Liang JJ, Pitsillou E, Kagarakis EA, Yi DTZ, Xu V, Hung A, El-Osta A (2024) Investigation of the anti-inflammatory properties of bioactive compounds from olive mill waste: in silico evaluation of cyclooxygenase enzyme inhibition and pharmacokinetic profiling. Molecules. ;29(15):3502. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/molecules29153502\u003c/span\u003e\u003cspan address=\"10.3390/molecules29153502\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 39124908\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JS, Jobin C (2005) The flavonoid luteolin prevents lipopolysaccharide-induced NF-κB signalling and gene expression by blocking IκB kinase activity in intestinal epithelial cells and bone-marrow derived dendritic cells. Immunology. ;115(3):375\u0026ndash;387. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1365-2567.2005.02156.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1365-2567.2005.02156.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 15946255\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen CY, Peng WH, Tsai KD, Hsu SL (2007) Luteolin suppresses inflammation-associated gene expression by blocking NF-κB and AP-1 activation pathway in mouse alveolar macrophages. Life Sci. ;81(23\u0026ndash;24):1602\u0026ndash;1614. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.lfs.2007.09.028\u003c/span\u003e\u003cspan address=\"10.1016/j.lfs.2007.09.028\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 17977562\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun ZJ, Chen G, Hu X, Zhang W, Liu Y, Zhu LX, Zhou Q, Zhao YF (2010) Activation of PI3K/Akt/IKK-α/NF-κB signaling pathway is required for apoptosis-evasion in human salivary adenoid cystic carcinoma: its inhibition by quercetin. Apoptosis. ;15(7):850\u0026ndash;863. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s10495-010-0497-5\u003c/span\u003e\u003cspan address=\"10.1007/s10495-010-0497-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 20386985\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEndale M, Park SC, Kim S, Kim SH, Yang Y, Cho JY, Rhee MH (2013) Quercetin disrupts tyrosine-phosphorylated phosphatidylinositol 3-kinase and myeloid differentiation factor-88 association, and inhibits MAPK/AP-1 and IKK/NF-κB-induced inflammatory mediators production in RAW 264.7 cells. Immunobiology. ;218(12):1452\u0026ndash;1467. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.imbio.2013.04.019\u003c/span\u003e\u003cspan address=\"10.1016/j.imbio.2013.04.019\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 23735482\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaldwin AS Jr (1996) The NF-κB and IκB proteins: new discoveries and insights. Annu Rev Immunol. ;14:649\u0026ndash;683. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1146/annurev.immunol.14.1.649\u003c/span\u003e\u003cspan address=\"10.1146/annurev.immunol.14.1.649\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 8717528\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu F, Xia Y, Parker AS, Verma IM (2012) IKK biology. Immunol Rev. ;246(1):239\u0026ndash;253. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1600-065X.2012.01107.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1600-065X.2012.01107.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 22435559\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBastys T, Gapsys V, Doncheva NT, Kaiser R, de Groot BL, Kalinina OV (2018) Consistent prediction of mutation effect on drug binding in HIV-1 protease using alchemical calculations. J Chem Theory Comput. ;14(7):3397\u0026ndash;3408. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1021/acs.jctc.7b01109\u003c/span\u003e\u003cspan address=\"10.1021/acs.jctc.7b01109\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 29847122\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMesser A, Raquet N, Lohr C, Schrenk D (2012) Major furocoumarins in grapefruit juice II: phototoxicity, photogenotoxicity, and inhibitory potency vs. cytochrome P450 3A4 activity. Food Chem Toxicol. ;50(3\u0026ndash;4):756\u0026ndash;760. doi:10.1016/j.fct.2011.11.023. PMID: 22155270\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, Wang L, Chen W, LoRA (2022) Low-rank adaptation of large language models. In: Proceedings of the International Conference on Learning Representations (ICLR). arXiv:2106.09685. (Preprint; not indexed in PubMed.)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQwen Team Qwen2.5 Technical Report. arXiv:2412.15115 [cs.CL]. 2024. (Preprint; not indexed in PubMed.)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9587405/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9587405/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eComputational multi-target profiling is technically accessible but methodologically ungoverned: thresholds set post-hoc, claims exceeding the data, failed branches omitted, and ΔG values untraceable to source files. We describe a platform that addresses these failure modes by construction.\u003c/p\u003e \u003cp\u003eThe platform enforces pre-specified gates (frozen in named decision sheets before any simulation runs), outcome-keyed ceiling templates (format_verdict.py, NemoClaw constitutional review), a 46-check regression oracle, and a SHA256 provenance chain linking every published number to the ΔTOTAL line of the source FINAL_RESULTS_DECOMP.dat file. All closed cycles feed a retrieval-first knowledge graph.\u003c/p\u003e \u003cp\u003eApplied to nine fig and olive phytochemicals across two targets: (Node 1) nine compounds versus COX-2 (5KIR), N\u0026thinsp;=\u0026thinsp;4 replicates each, ALL_PASS; oleocanthal ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;61.8 kcal/mol, CV\u0026thinsp;=\u0026thinsp;2.4%, ΔΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;22.9 kcal/mol versus celecoxib. (Node 2) luteolin versus IKKβ (4KIK): initial CV\u0026thinsp;=\u0026thinsp;16.1% (FAIL); structural diagnosis and pre-registered deep-pocket reseed produced ALL_PASS (mean ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;31.457 kcal/mol, CV\u0026thinsp;=\u0026thinsp;6.85%, N\u0026thinsp;=\u0026thinsp;3). (Node 3) quercetin versus IKKβ: generalization, ALL_PASS (mean ΔG\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;24.603 kcal/mol, CV\u0026thinsp;=\u0026thinsp;3.04%, N\u0026thinsp;=\u0026thinsp;3). (Node 4) quercetin versus p65 (1VKX): POSE_UNSTABLE, branch closed; platform claim bounded to two-node model.\u003c/p\u003e \u003cp\u003eAll results are EXPLORATORY DDT tier: computationally reproducible under the stated protocol, not biologically validated. Appropriate next steps are enzymatic binding assays, kinase selectivity profiling, and safety characterization. Frozen artifact set (28 SHA256-verified files): \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.19076569\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.19076569\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e","manuscriptTitle":"A Governed Multi-Node Computational Discovery Pipeline: Pre-Specified Gates, Bounded Claims, and Honest Branch Closure in Phytochemical Polypharmacology","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-18 10:04:54","doi":"10.21203/rs.3.rs-9587405/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e4352fab-d059-44fa-935c-2791e387c73c","owner":[],"postedDate":"May 18th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-23T13:04:50+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-17T13:26:20+00:00","index":32,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-13T06:44:43+00:00","index":31,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-11T10:28:53+00:00","index":30,"fulltext":""},{"type":"reviewerAgreed","content":"102010346335572076107027293266810775244","date":"2026-05-11T00:30:46+00:00","index":28,"fulltext":""},{"type":"reviewerAgreed","content":"334433666274539763252693066711754522200","date":"2026-05-10T15:54:17+00:00","index":27,"fulltext":""},{"type":"reviewerAgreed","content":"340188918200967734416323802914164202629","date":"2026-05-10T15:49:20+00:00","index":26,"fulltext":""},{"type":"reviewerAgreed","content":"142643565920744764546622655690359949086","date":"2026-05-10T14:16:32+00:00","index":25,"fulltext":""},{"type":"reviewerAgreed","content":"15046804172852408800133309444540190938","date":"2026-05-08T19:02:51+00:00","index":22,"fulltext":""},{"type":"reviewersInvited","content":"14","date":"2026-05-08T12:15:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T12:17:00+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-04T12:16:07+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-23T13:10:28+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-18 10:04:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9587405","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9587405","identity":"rs-9587405","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}