{"paper_id":"28ac4dbb-e58f-43f0-8b2d-9af22298cc69","body_text":"1 \nSignal Versus Noise: Evaluating iNaturalist Photos as a Source of \nQuantitative Phenotypic Data in Plethodon  Salamanders using \nAutoresearch and Agentic AI \nKyle A. O'Connell \nBiomedical Data Science Team, Deloitte LLP \n1919 N Lynn St. \nArlington, VA 22209 \nORCID: https://orcid.org/0000-0002-0464-9259 \nKeywords: citizen science, iNaturalist, phenotyping, color variation, Plethodon, image analysis, \necogeographic rules, observer bias \n \nAbstract \nCommunity-science platforms such as iNaturalist now contain tens of millions of georeferenced, \nphotographically vouchered biodiversity records, yet extracting reliable quantitative measurements from \nopportunistic photographs remains methodologically challenging. Here, I evaluate the signal-to-noise \nratio of iNaturalist photos for phenotyping Plethodon salamanders across two trait classes: continuous \ndorsal brightness (a proxy for ecogeographic clines predicted by Gloger's rule and the thermal melanism \nhypothesis) and discrete color morph frequency in P. cinereus. I optimized a color-extraction pipeline \nusing an agent-guided parameter search adapted from the autoresearch framework (Karpathy 2026; \nSchmidgall et al. 2025), exploring crop fraction, color space, normalization, and quality-control \nthresholds across 50 bounded micro-experiments. Applying the production HSV pipeline to 103,653 \nobservations of 34 species, I found negligible geographic structure in dorsal brightness (R² = 0.001), even \nwithin P. cinereus alone (n = 71,627). Variance decomposition showed that photographer identity \nexplains 23.3% of brightness variance, geography 5.1%, species 1.6%, and time of day 0.3%, with 69.7% \nresidual. In contrast, a hue-threshold morph classifier recovered a significant geographic signal in red-\nback frequency (R² = 0.008, p < 0.001), 7× stronger than the brightness result, though still weaker than \nthe supervised CNN of Hantak et al. (2022; pseudo-R² \n≈  0.04). These results indicate that citizen-science \nphotographs are poorly suited to continuous quantitative phenotyping under current collection conditions, \nwhereas discrete categorical traits remain recoverable with appropriate classifiers. The autoresearch loop \nclarified the failure mode: no tested parameter configuration recovered a meaningful brightness signal \nfrom a dataset dominated by observer effects. \nIntroduction \nColor is among the most tractable phenotypic traits in ecology and evolutionary biology. It is directly \nobservable, ecologically meaningful, and predictably distributed across geography in ways that link \nphysiology, behavior, climate, and natural selection (Endler 1980; McLean and Stuart-Fox 2014). Two \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n2 \nclassic ecogeographic rules make explicit, opposing predictions about how coloration should vary with \nlatitude and climate in terrestrial ectotherms. Gloger's rule predicts darker pigmentation in warmer, more \nhumid environments, consistent with adaptive responses to UV radiation, desiccation, and parasitism \n(Delhey 2019). The thermal melanism hypothesis predicts darker coloration at higher latitudes and \nelevations, where darker bodies absorb solar radiation more efficiently and support faster warming in cool \nenvironments (Clusella-Trullas et al. 2007). In a genus of animals spanning ~25° of latitude across eastern \nNorth America, the direction of any observed latitudinal brightness gradient would immediately inform \nwhich selective pressure dominates. \nPlethodon (family Plethodontidae) is the most species-rich vertebrate genus in eastern North America, \ncomprising ~55 species of lungless salamanders that occupy forest floor microhabitats from Florida to \nNewfoundland and from sea level to the high Appalachians. The genus is exceptionally well studied from \ncommunity ecology and biogeographic perspectives (Hairston 1951; Jaeger 1971; Highton 1995), but \nquantitative range-wide analysis of dorsal coloration has been limited by the difficulty of obtaining \nstandardized photographs across the full geographic range. iNaturalist (Di Cecco et al. 2021; iNaturalist \ncontributors 2026) now hosts >154,000 Research Grade Plethodon observations, each linked to a \ngeoreferenced photograph, representing orders of magnitude more geographic coverage than any museum \ncollection or directed field survey. \nThe central challenge in using this resource for quantitative phenotyping is that iNaturalist photographs \nare taken by thousands of observers under heterogeneous conditions, with uncontrolled lighting, camera \nsettings, and subject positioning. Extracting measurements that reflect biological rather than photographic \nvariation therefore requires a pipeline that is robust to image heterogeneity and explicitly evaluated for its \nability to recover known signals. This evaluation step, namely determining what the photos can and \ncannot reveal, is not yet routine in citizen-science phenotyping studies. \nRecent work has begun to address this gap for discrete phenotypic classes. Hantak et al. (2022) \ndemonstrated that an ensemble convolutional neural network (CNN) trained on 4,000 human-labeled \niNaturalist images could classify the striped/unstriped color polymorphism in P. cinereus with ~98% \naccuracy across 20,318 images, and used the resulting annotations to document climatic niche \nassociations between morphs at range-wide scale (pseudo-R² \n≈  0.04). The success of this approach \ndepended on supervised training with expert-labeled data and a whole-image deep learning architecture \ncapable of implicitly localizing the relevant dorsal region. Whether simpler, unsupervised photometric \nmethods — mean pixel brightness, hue, or saturation extracted from a central crop — can recover \ncomparable signals remains unknown. \nFor continuous traits, the question is more fundamental: are the photos informative at all? Observer-level \nheterogeneity in exposure, flash use, and photo composition creates additive noise that may overwhelm \nsubtle quantitative gradients. McCormick and Riley (2025) documented systematic observer bias in the \nsame system: in a concurrent field survey in New Brunswick, Canada, 95.9% of encountered striped-\nmorph individuals were striped, whereas only 83.2% of iNaturalist observations were striped — a \nsignificant overrepresentation of rarer morphs in the community science dataset (\nχ ² = 11.92, p < 0.01). At \nthe level of specific rare morphs, the unstriped form was observed at 6.6% on iNaturalist versus 3.3% in \nfield surveys (χ ² = 5.83, p = 0.02). This suggests that observer selection preferences distort even \ncategorical measurements. Whether analogous biases distort continuous brightness measurements has not \nbeen formally evaluated. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n3 \nHere, I address these questions using three complementary analyses. First, I optimize and benchmark a \ncolor extraction pipeline for iNaturalist Plethodon photographs using a bounded, agent-guided parameter \nsearch adapted from Karpathy's (2026) autoresearch framework and related LLM-agent research-\nassistant workflows (Schmidgall et al. 2025), which treat parameter optimization as a logged sequence of \nmicro-experiments evaluated against a composite geographic signal score. Second, I apply a stable \nproduction extraction pipeline to 103,653 observations and use the pilot-optimized alternative as a \nsensitivity check on the validation subset. Third, I decompose brightness variance into biological and \nphotographic sources using hierarchical ICC analysis and a linear mixed model, and compare the \nresulting signal-to-noise profile against a simple hue-threshold morph classifier applied to the same \nphotographs. My goal is to provide an explicit empirical characterization of what iNaturalist photographs \ncan and cannot reveal about phenotypic variation in a well-studied salamander clade — and to \ndemonstrate a reproducible, computationally formalized approach to optimizing extraction pipelines for \ncitizen science image data. \nMethods \nStudy System and Data Source \nAll observations were obtained from iNaturalist through the iNaturalist API (Di Cecco et al. 2021; \niNaturalist contributors 2026). I downloaded metadata for 103,669 Research Grade observations spanning \n34 species and covering the eastern and Pacific Plethodon radiations across latitudes 25–50°N. Research \nGrade status requires that an observation be accompanied by a usable photograph and that species identity \nbe confirmed by a minimum of two independent community reviewers. Preprocessing included removal \nof observations with missing or obscured geographic coordinates, deduplication of near-identical records \nwithin 100 m, standardization of observation fields, and flagging of potential out-of-range occurrences \nbased on known species distributions. After preprocessing and image-level quality-control filtering \n(brightness and entropy thresholds; see Color Feature Extraction below), the final analysis dataset \ncomprised 103,653 observations — a net loss of only 16 records, reflecting the permissive nature of the \nautomated QC thresholds documented in the crop quality audit. Cleaned observations were assigned to \nH3 hexagonal grid cells at resolution 5 (~252 km² per cell) using the h3-py library (Brodsky 2018). This \nspatial binning unit provides a geographically consistent footprint for estimating within-location variance \nand supports the geographic signal score used in pipeline optimization. \nFrom the cleaned observation table, I generated a photo manifest containing the first available image URL \nfor each record. Images were downloaded at a rate-limited throughput of ~10 requests/second using a \nshared token-bucket limiter across six concurrent threads, with per-request retry logic and checkpoint \nrecovery. Downloaded images were stored by observation identifier to maintain a one-to-one mapping \nbetween extracted color measurements and source records. \nAgent-Guided Optimization of the Color Extraction Pipeline \nThe image analysis workflow was framed as a constrained optimization problem adapted from the \nautoresearch framework (Karpathy 2026), in which parameter search is structured as a logged sequence \nof bounded micro-experiments rather than ad hoc manual tuning. An LLM agent (claude-opus-4-6) \niteratively proposed one parameter change at a time to a restricted configuration dictionary, executed the \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n4 \nrevised pipeline on a fixed validation subset of ~859 photographs, evaluated the result against a \npredefined composite score, and retained or reverted the change before proposing the next experiment. \nThe agent was constrained to modify only the following parameters: \n• central_crop_fraction [0.15–0.7]: Fraction of image area retained in the centered dorsal crop \n• color_space [HSV, LAB, RGB]: Color representation used for brightness extraction \n• normalize_brightness [bool]: Whether to apply per-channel histogram equalization before \nextraction \n• background_mask [none, green_threshold, saturation_threshold]: Optional pixel exclusion \nstrategy \n• min_brightness, max_brightness, min_entropy: Quality-control rejection thresholds \n• percentile_trim [0–10%]: Fraction of extreme pixel values excluded from mean calculation \nThe composite optimization score rewarded geographic signal while penalizing local noise: \n\\mathrm{Score} = R^2(\\mathrm{brightness} \\sim \\mathrm{latitude}) - \\lambda \\times \n\\mathrm{NormalizedWithinCellVariance} \nwhere $\\lambda = 0.5$. Each experiment returned extracted brightness values linked to observation \ncoordinates and H3 cells; from these, I computed the latitude regression R² and mean within-cell \nbrightness variance across cells with at least three observations. All experiments were logged with their \nfull parameter configuration, score, regression statistics, and QC summary. \nThe pilot optimization loop ran for 50 iterations on 859 photographs from an initial test dataset. I then \ndeployed the same workflow on a Google Cloud Platform VM (n2-highmem-32, 32 vCPUs, 256 GB \nRAM, 200 GB SSD boot disk, Debian 12, us-east1-b) to handle large-scale photo download, \nadditional optimization, and production extraction. The VM operated autonomously through a startup \nscript with no interactive access: it cloned the repository, installed dependencies, downloaded the full \nphoto corpus, ran the optimization loop, performed final extraction, uploaded results, and self-terminated. \nAn initial spot-instance deployment was preempted after 2.5 hours during photo download; the final \nproduction run therefore used standard on-demand provisioning (~$0.33/hr spot vs. ~$2.10/hr on-demand \nfor n2-highmem-32), with an estimated total cost of ~$21 for ~10 compute-hours. All local preprocessing, \ndownstream statistical analysis, and figure generation were performed on a MacBook Air (M3, 16 GB \nRAM, macOS 26.3). \nColor Feature Extraction \nThe pilot autoresearch loop identified CIE L\\a\\b\\* color space and histogram normalization as the largest \nimprovements to the composite score. For the primary manuscript analyses, however, I report the stable \nHSV production extraction (40% central crop, no normalization) that had already completed across the \nfull dataset when results were frozen for analysis. The pilot-optimized LAB configuration is therefore \ntreated as a validation-subset sensitivity check rather than as the production pipeline. Although the \noptimized configuration substantially reduced within-cell variance on the validation subset, both HSV and \nLAB pipelines led to the same qualitative conclusion: geographic brightness signal remained negligible \nrelative to photographic noise. All reported brightness values therefore use HSV V-channel measurements \non a 0–255 scale. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n5 \nFor each downloaded photograph, I extracted dorsal color features from a centered crop using the \nfollowing procedure: (1) load image as RGB using Pillow; (2) extract the central 40% × 40% fraction; (3) \nconvert to HSV using scikit-image (H in degrees 0–360, S and V on 0–255 scales); (4) compute Shannon \nentropy of the full crop as an image-quality indicator; (5) apply quality-control filters (brightness \n/i3  [15, \n245], entropy ≥  4.0) and exclude images that fail. Retained images contributed five metrics per \nobservation: mean_brightness (V channel), mean_hue (H channel), mean_saturation (S \nchannel), entropy, and image dimensions. After QC filtering and merging to observation metadata (lat, \nlon, species, H3 cell, observer, date), the final analysis dataset comprised 103,653 observations. \nGeographic Brightness Analysis \nI tested for geographic clines in dorsal brightness using ordinary least squares (OLS) regression of \nmean_brightness against latitude and longitude (brightness ~ lat + lon), implemented in statsmodels. \nI fit this model (1) across all 34 species simultaneously (n = 103,653) and (2) for P. cinereus alone (n = \n71,627), the most-photographed species and the one for which observer bias and morph composition are \nbest characterized. At the cell level, I aggregated mean brightness per H3 res-5 cell and re-fit the model \non cell means (n = 6,474 cells). A positive latitude coefficient would indicate brighter dorsal coloration at \nhigher latitudes (consistent with Gloger's rule); a negative coefficient would indicate darker coloration \n(consistent with thermal melanism). I report R², adjusted R², F-statistic p-value, and regression \ncoefficients with standard errors. Cross-species heterogeneity in brightness was tested with a Kruskal–\nWallis test across all species with \n≥ 10 observations. \nVariance Decomposition \nTo partition brightness variation into biological versus photographic sources, I computed intraclass \ncorrelation coefficients (ICC) for four grouping variables: species identity (34 groups), individual \nobserver (user_id, 33,496 groups), hour of day (24 groups, derived from the timezone-aware \nobserved_on timestamp), and H3 res-5 geographic cell (6,474 groups). The ICC(1) estimator for each \nvariable was: \n\\mathrm{ICC} = \\frac{MS_{between} - MS_{within}}{MS_{between} + (n_0 - 1) \\times \nMS_{within}} \nwhere $n_0$ is the mean group size and $MS_{between}$/$MS_{within}$ are the between- and within-\ngroup mean squares from a one-way ANOVA. I also fit a linear mixed model (statsmodels MixedLM, \nREML) with user_id as a random intercept and latitude, longitude, species code, hour of day, image \nentropy, and aspect ratio as fixed effects, to estimate the proportion of total variance attributable to the \nphotographer random effect. Feature engineering added hour_of_day (extracted from \nobserved_on), aspect_ratio (width/height), and log_obs_per_user (log of observations per \nobserver, as an experience proxy) to the analysis dataset. ICC values were floored at zero; residual \nvariance was estimated as 100% minus the sum of all ICC terms. \nColor Morph Classification and Geographic Validation \nAs a positive control for geographic signal detection, I applied a hue-threshold classifier to the 71,627 P. \ncinereus observations to distinguish red-back (striped) from lead-back (unstriped) morphs. Following the \ndocumented pigment biology of this polymorphism (Lotter and Scott 1977), I classified observations as \nred-back if mean_hue fell within the warm range (0–60° or 300–360°) and mean_saturation ≥  30 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n6 \n(on a 0–255 scale). Observations with mean_saturation < 30 were classified as lead-back regardless \nof hue; observations with hue in the clearly cool range (90–270°) and saturation ≥  30 were classified as \nlead-back; all remaining observations were classified as ambiguous and excluded from geographic \nanalysis. These conservative thresholds prioritize classification confidence over coverage, retaining only \nobservations with unambiguous photometric evidence of morph identity. \nClassified observations (excluding ambiguous) were aggregated per H3 res-5 cell to compute red-back \nmorph frequency (n_redback / n_total). I fit an OLS model of red-back frequency ~ mean cell latitude + \nmean cell longitude using cells with \n≥ 5 classified observations (n = 1,776 cells). The resulting R² was \ncompared directly to the brightness R² as the primary signal-versus-noise contrast. All analyses were \nimplemented in Python using pandas, numpy, scipy, and statsmodels; figures were generated with \nmatplotlib and seaborn. Code and derived analysis outputs are available at \nhttps://doi.org/10.5281/zenodo.19050224. \nCrop Quality Validation \nTo assess whether the center-crop protocol reliably isolates dorsal salamander body pixels, I conducted a \nmanual audit of 200 randomly sampled locally cached photographs (seed 20260323). Each crop was \nscored by a single rater using a four-category rubric: (1) good — crop contains mostly salamander \ntrunk/dorsal pixels, acceptable for brightness extraction; (2) partial — salamander present but substantial \nnon-body content remains; (3) fail — crop is not usable dorsal body pixels; (4) in_hand — hand or glove \nis a major crop component. I additionally recorded brightness_usable (yes/no/unclear) and \nmorph_usable (yes/no/unclear) per image, derived conservatively from crop-quality category. This \naudit was designed as a single-rater protocol failure-rate estimate, not an inter-rater validation study. \nAudit results were used to compute: (a) concordance between the automated passed_qc flag and \nmanual crop-quality labels; (b) a test of whether poor-quality crops are geographically structured (Mann–\nWhitney U comparing latitude distributions of good versus not-good observations); and (c) two \nsensitivity regression analyses. Path A restricted to brightness_usable = yes observations from \nthe audit-matched parquet subset (n = 75). Path C trained a logistic regression classifier (L2 \nregularization, C = 1.0) on five features available in the merged parquet (mean_brightness, \nmean_hue, mean_saturation, entropy, aspect_ratio = width/height) using the 195 \naudit-labeled observations, evaluated by 5-fold cross-validation, and applied to the full 103,653-\nobservation dataset. Sensitivity regressions on the Path C–filtered subset were compared to full-dataset \nresults, and the species composition of predicted-good versus predicted-bad subsets was inspected to \nassess whether the classifier introduced morph-sampling bias. \nResults \nAgent-Guided Pipeline Optimization \nThe pilot autoresearch loop completed 50 experiments on 859 photographs, of which 12 (24%) were \naccepted as improvements to the composite score. The baseline configuration (40% central crop, HSV \ncolor space, no normalization, no masking) achieved a composite score of −0.257, driven primarily by \nhigh within-cell brightness variance (mean = 1,101). Over the course of accepted experiments, the \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n7 \ncomposite score improved to +0.010 — a +0.267 absolute improvement — while the latitude R² remained \nstable at 0.018 and within-cell variance fell 97% to 33 (Table 1). \nTable 1. Pilot autoresearch loop summary (composite score uses λ  = 0.5). \nParameter Baseline Optimized \nColor space HSV CIE L\\*a\\*b\\* \nCrop fraction 0.40 0.64 \nHistogram normalization No Yes \nBackground masking None None \nPercentile trim 0% 4.7% \nComposite score −0.257 +0.010 \nR² (brightness ~ latitude) 0.018 0.018 \nWithin-cell variance 1,101 33 \n \nThe most consequential single change was histogram normalization (+0.163 to composite score), which \nstandardized variable exposure across photos. Switching from HSV to CIE L\\a\\b\\* color space (+0.058) \nimproved the perceptual uniformity of brightness estimates, and a larger crop fraction (+0.021) captured \nmore dorsal surface area. Background-masking strategies (green-channel threshold, saturation threshold) \nwere explored but never improved the score and were ultimately excluded from the final configuration. \nThe 24% acceptance rate is consistent with a well-constrained search: most proposed changes either had \nnegligible effect or degraded the within-cell noise metric. \nGeographic Brightness Analysis \nApplying the HSV production extraction pipeline to 103,653 observations across 34 species, I found \nnegligible geographic structure in dorsal brightness. The OLS model brightness ~ latitude + longitude \nexplained R² = 0.001 across all species (adj. R² = 0.001, F-test p < 0.001; Table 2). The result held within \nP. cinereus alone (n = 71,627; R² = 0.001, p < 0.001), where species composition is held constant. At the \nH3 cell level (n = 6,474 cells), R² = 0.003 across all species and R² = 0.001 for P. cinereus cells. \nEstimated regression coefficients were near zero (latitude: β  ≈  0.04 V/degree, longitude: β  ≈  0.01 \nV/degree) with confidence intervals spanning both positive and negative values. A Kruskal–Wallis test \nconfirmed that species differ significantly in mean brightness (H = 830.1, p = 1.56 × 10\n/i3 ¹/i3/i3 ), indicating \nthat substantial cross-species signal exists in the dataset — it is the within-species, across-geography \ngradient that is absent. The regression panels, range-wide brightness map, and species-comparison plot \n(Figures 2-4) all reinforce the same pattern: substantial between-species spread but no clear within-\nspecies geographic cline. \nTable 2. OLS brightness ~ latitude + longitude results. \nDataset n R² adj. R² F p-value \nAll species \n(observation level) \n103,653 0.0011 0.0011 < 0.001 \n*P. cinereus* only \n(observation level) \n71,627 0.0011 0.0011 < 0.001 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n8 \nAll species (cell \nlevel, n≥ 3) \n6,474 0.003 0.003 < 0.001 \n \n \nFigure 2. Dorsal brightness versus geography across analytic levels, including observation-level \nregressions, within-species slopes, and species means. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n9 \n \nFigure 3. Geographic variation in Plethodon dorsal brightness across the study extent. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n10 \n \nFigure 4. Distribution of dorsal brightness values for the 15 most-observed Plethodon species. \nVariance Decomposition \nICC analysis partitioned brightness variance into four identifiable sources (Figure 1A, Table 3). Observer \nidentity (user_id) showed the largest ICC (0.233), indicating that individual photographers account for \n23.3% of total brightness variance — more than all other modeled sources combined. Geographic cell \n(ICC = 0.051) explained 5.1%, reflecting a mixture of real spatial structure and spatially correlated \nphotography conditions. Species identity explained 1.6% (ICC = 0.016) and hour of day only 0.3% (ICC \n= 0.003). Estimated residual variance (unaccounted by any of these groupings) was 69.7%, attributable to \ncamera settings, flash, angle, zoom level, and substrate variation within individual photos. \nTable 3. ICC variance decomposition. \nSource ICC % Variance N Groups N Obs \nObserver (user_id) 0.233 23.3% 33,496 103,653 \nGeography (H3 res-\n5) \n0.051 5.1% 6,474 103,653 \nSpecies 0.016 1.6% 34 103,653 \nHour of day 0.003 0.3% 24 103,653 \nResidual — 69.7% — — \n \nNote: ICCs were estimated from separate one-way ANOVAs; grouping variables are not orthogonal (e.g., \nobserver and geography are correlated). The residual is defined as 100% minus the sum of individual \ner \nto \ng., \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n11 \nICCs. A single multilevel model with crossed random effects would yield a different decomposition. The \nmixed model estimate of 25.8% for observer (below) provides a partial correction. \nThe linear mixed model with user_id as a random intercept confirmed the observer variance estimate: \nthe photographer random effect accounted for 25.8% of total brightness variance, whereas the fixed \neffects collectively added little explanatory power beyond the grand mean. Hour-of-day analysis showed \na modest brightening trend from dawn through midday followed by decline through \nevening, but the ICC \nof 0.003 indicates that time-of-day effects are minor relative to observer identity. The distribution of per-\nobserver mean brightness (among observers with \n≥ 5 observations) had a standard deviation of ~16.7 V \nunits on a 0–255 scale, confirming that photographers introduce systematic additive offsets amounting to \nroughly 6.5% of the full brightness range. \n \nFigure 1. Composite summary of variance decomposition and morph-frequency validation. (A) ICC\nvariance decomposition across observer, geography, species, hour-of-day, and residual sources. (B) \nH3-cell map of classified red-back morph frequency in Plethodon cinereus. (C) Red-back morph \nfrequency as a function of mean cell latitude for cells with at least five classified observations.  \nCrop Quality Validation and Sensitivity Analysis \nManual scoring of 200 audited crops revealed that only 76 (38.0%) were clearly acceptable for dorsal-\nbody brightness extraction (crop_quality = good; Figure 5A-B; Table 4). The remaining 62.0% \nwere partial (n = 58; 29.0%), fail (n = 24; 12.0%), or in_hand (n = 42; 21.0%). The automated \nQC filter (passed_qc: brightness /i2  [15, 245], entropy ≥  4.0) failed to reject a single image in the 200-\nimage audit sample: 100% of images across all four quality categories, including all 42 in_hand \nC \nC \n) \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n12 \nimages, passed the automated screen (Figure 5C). This confirms that passed_qc functions as a data-\nquality filter (rejecting overexposed or uniformly dark images) rather than a localization check. \nTable 4. Manual crop quality audit (n = 200 images). \nLabel n % brightness_usable morph_usable \n`good` 76 38.0% yes yes \n`partial` 58 29.0% unclear yes \n`fail` 24 12.0% no no \n`in_hand` 42 21.0% no unclear \n \nThe latitude distributions of good and not-good crops were not significantly different (Mann–Whitney \nU = 4,096, p = 0.29; good mean lat = 39.3°, bad mean lat = 40.0°; Figure 5D), indicating that crop \nfailures are geographically random rather than systematically concentrated at particular latitudes. This \nsupports treating bad crops as additive measurement noise rather than a source of directional geographic \nbias. \nPath A (audit clean subset). Restricting to the 75 observations scored brightness_usable = \nyes, OLS regression of brightness ~ lat + lon returned R² = 0.027 (n = 75, p = 0.37). The estimate is not \nstatistically significant and is consistent with the null result from the full dataset (R² = 0.001), though the \nsmall sample precludes strong inference. On the 67 P. cinereus observations scored morph_usable = \nyes, the ambiguous classification rate dropped to 16.4% (vs. 20.3% in the full dataset), consistent with \nthe interpretation that most ambiguous calls originate from in-hand photographs. A cell-level morph \nregression on the classified subset returned R² = 0.360 across only 6 cells (p = 0.51), which is too sparse \nto support strong inference despite the larger point estimate. \nPath C (classifier-filtered full dataset). A logistic regression classifier trained on 195 audit-labeled \nobservations achieved 5-fold cross-validation accuracy of 0.65 (majority-class baseline = 0.39), indicating \nmodest above-chance discriminability. Applied to the full 103,653-observation dataset, the classifier \nretained 10,823 observations (10.4%). Brightness R² on the filtered subset was 0.0008 (n = 10,823), \nessentially identical to the full-dataset result. The classifier introduced a notable morph-sampling bias: the \npredicted-good P. cinereus subset was 67.5% red-back versus 49.2% in the predicted-bad subset, \nreflecting the spectral overlap between human skin-tone hues and red-back dorsal coloration in HSV \nspace. Morph frequency R² on the classifier-filtered subset was 0.022 (n = 255 cells, p = 0.06) — \nnumerically higher than the full-dataset result (0.008) but not statistically significant, and likely inflated \nby the morph-composition bias rather than reflecting genuine improvement in localization quality. \nTable 5. Sensitivity analysis comparison. \nDataset n (obs) Brightness R² Morph R² (cells) \nFull dataset 103,653 0.0011 0.0080 \nPath A: audit-clean subset 75 0.027 (p = 0.37) 0.360 (p = 0.51; 6 cells) \nPath C: classifier-filtered 10,823 0.0008 0.022 (p = 0.06) \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n13 \n \nFigure 5. Crop-audit validation of the center-crop protocol. (A) Representative audited crops \nscored `good`. (B) Representative problematic crops, including partial-body and in-hand cases. (C) \nConcordance between automated `passed_qc` and manual crop-quality labels. (D) Geographic \ndistribution of audited crop-quality labels, showing no strong latitudinal bias in crop failures.  \nColor Morph Classification and Geographic Signal \nThe hue-threshold classifier applied to 71,627 P. cinereus observations classified 29,004 (40.5%) as red-\nback, 28,062 (39.2%) as lead-back, and 14,561 (20.3%) as ambiguous. The 20.3% ambiguous rate is \nclosely consistent with the 21.0% in_hand rate observed in the manual crop audit, indicating that most \nunclassifiable observations reflect photographic rather than phenotypic ambiguity: crops dominated by \nhand or glove pixels lack the hue/saturation signal needed to resolve morph identity. Red-back frequency \nacross 3,746 H3 cells with \n≥ 1 classified observation was 0.51 (mean across cells). Aggregating to cells \nwith ≥ 5 classified observations (n = 1,776), OLS regression of red-back frequency ~ mean cell latitude + \nmean cell longitude returned R² = 0.008 (lat-only R² = 0.003; F-test p < 0.001; Figure 1B, 1C). This is 7× \nlarger than the brightness R² from the same photographs (0.008 vs. 0.001), and statistically significant, \nconfirming that a detectable geographic signal in morph composition exists in these photos — one that \nthe continuous brightness analysis completely fails to recover. \nThe morph map (Figure 1B) shows spatially coherent variation in red-back frequency across the P. \ncinereus range, with a modest northward and westward increase in red-back frequency consistent with \n) \n× \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n14 \npublished regional and range-wide surveys (Gibbs and Karraker 2006; Hantak et al. 2019; Hantak et al. \n2022; Moore and Ouellet 2015). However, classified red-back frequency (51%) falls substantially below \npublished field-survey estimates for this species, which range from ~74–80% range-wide (Gibbs and \nKarraker 2006) to >95% in northern populations (McCormick and Riley 2025), indicating that the \nclassifier systematically underestimates red-back prevalence. I attribute this to observer novelty bias: \nlead-back individuals are photographed disproportionately relative to their population frequency because \nthey appear unusual to citizen scientists, inflating the lead-back fraction of the iNaturalist sample \n(McCormick and Riley 2025). \nDiscussion \nThe Autoresearch Loop as a Phenotyping Optimization Tool \nThe autoresearch framework (Karpathy 2026) was designed for machine-learning hyperparameter search, \nin which a clear, automatable loss function evaluates candidate configurations within a constrained search \nspace. Here, I show that the same architecture transfers naturally to ecological image-pipeline \noptimization, where the objective function combines geographic signal strength with within-cell \nmeasurement noise. The 50-experiment pilot loop identified four actionable improvements, namely CIE \nL\\a\\b\\* color space, histogram normalization, a larger crop fraction, and moderate percentile trimming, \nthat reduced within-cell variance by 97% while leaving the latitude R² unchanged. This outcome \nillustrates a useful property of the approach: a well-specified metric can distinguish changes that improve \nprecision from changes that merely amplify noise or compress signal. \nThe 24% acceptance rate is also informative about the optimization landscape. Most parameter \nperturbations either failed to improve the composite score or traded signal for noise reduction. \nBackground masking, although initially appealing as a way to exclude leaf litter and substrate pixels from \nthe crop, never improved the metric, likely because the centered crop already excluded most background \nand masking introduced edge artifacts that inflated within-cell variance. That conclusion would have been \ndifficult to reach through informal manual iteration; the loop established it quickly and transparently. \nMore broadly, the autoresearch pattern should be useful for phenotyping pipelines in which (1) the \nparameter space is sufficiently bounded for sequential search, (2) a composite evaluation metric encoding \nboth sensitivity and reproducibility can be formalized, and (3) per-experiment cost is low enough to \npermit many iterations. The ~$21 total cloud cost for the full production run, including a preemption-\nrecovery restart, suggests that this approach is feasible without unusually large compute budgets. \nWhy Continuous Brightness Fails: The Noise Floor \nThe primary geographic analysis returned a well-powered null result under the present measurement \napproach: brightness ~ latitude + longitude explains R² = 0.001 across 103,653 observations and 34 \nspecies, and the same non-result holds within P. cinereus alone across 71,627 individuals spanning the \nspecies' full geographic range. This is not a statistical power problem — at n = 71,627, this provides \n>99% power to detect R² \n≥  0.01. The result indicates that no detectable geographic brightness cline \nemerges from these photos once measurement noise is taken into account. \nThe variance decomposition explains this result. Observer identity accounts for 23.3% of brightness \nvariance, a value corroborated by the mixed model estimate of 25.8%. The 33,496 photographers \nrepresented in this dataset introduce characteristic brightness offsets on the order of ±16.7 V units among \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n15 \nobservers with at least five observations, plausibly reflecting device-specific exposure behavior, flash use, \nand preferred shooting distance. Any real latitudinal gradient in dorsal brightness would therefore need to \nexceed this observer noise floor to be recoverable from raw pixel values. At the 0–255 scale used here, \nthe observed inter-observer offsets imply that only comparatively large geographic shifts in mean \nbrightness would be detectable. \nThe remaining 69.7% residual variance captures additional uncontrolled heterogeneity in citizen-science \nphotography, including viewing angle, zoom level, substrate leakage into the crop, and background \nillumination. The geographic cell ICC (5.1%) is larger than the species ICC (1.6%), suggesting that \nlocations differ not only biologically but also in the characteristic conditions under which observations are \nrecorded. Hour of day (0.3% variance) proved comparatively unimportant, likely because most Plethodon \nobservations are made during evening and nighttime coverboard searches, which restrict the range of \nambient lighting conditions. \nA manual crop quality audit of 200 images provides direct evidence that a substantial fraction of this \nresidual variance originates from localization failures. Only 38% of audited crops were clearly suitable \nfor dorsal brightness extraction; 21% showed the animal in hand, and an additional 41% were partial or \nunusable frames. Critically, the automated QC filter (brightness + entropy thresholds) did not reject a \nsingle image in the audit sample — confirming that passed_qc is not a localization check. Two \nsensitivity analyses support the robustness of the null brightness result. Restricting to the 75 audit-verified \nclean observations returned a non-significant R² = 0.027 (p = 0.37), consistent with the null. A logistic \nregression classifier applied to the full 103K dataset retained 10,823 predicted-good observations and \nreturned R² = 0.0008. A geographic bias test showed no significant difference in latitude between good \nand poor-quality crops (Mann–Whitney p = 0.29), indicating that crop failures are randomly distributed \nacross the range. Random localization failures inflate measurement noise but cannot create a geographic \nbrightness cline where none exists; the null result is therefore not an artifact of the crop failure rate. \nThe optimized extraction parameters — histogram normalization especially — reduced within-cell \nvariance by 97% in the pilot, but this compression of photographic noise did not reveal a previously \nhidden geographic signal. This is the critical diagnostic: if photographic noise were masking a real signal, \nnormalization would have uncovered it by improving the signal-to-noise ratio. The fact that R² remained \nstable at ~0.018 (pilot) and 0.001 (full dataset) despite dramatically reduced noise indicates that the \ngeographic brightness signal in iNaturalist Plethodon photos is either absent or smaller than any \nphotometric normalization can recover. \nWhat Discrete Morphs Can Detect \nThe hue-threshold morph classifier recovered a geographic signal in P. cinereus red-back frequency (R² = \n0.008) that is 7× larger than the brightness R² from the same photos. This serves as a narrow positive \ncontrol: the same photographs that fail to reveal a continuous brightness gradient still contain sufficient \ncolor information to classify morphs and detect geographic variation in morph frequency. The signal is \nmodest (R² = 0.008 at the cell level), but statistically significant and directionally consistent with \npublished range-wide surveys (Gibbs and Karraker 2006; Moore and Ouellet 2015). \nHowever, direct comparison with Hantak et al. (2022) reveals the limits of the unsupervised threshold \napproach. Their ensemble CNN, trained on 4,000 human-labeled images, achieved ~98% classification \naccuracy and recovered climatic associations with pseudo-R² \n≈  0.04 — roughly 5× stronger than the \npresent threshold-based result from the same image source (though note that OLS R² and pseudo-R² are \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n16 \nnot directly comparable metrics, so this ratio is approximate). The performance gap is attributable to \nseveral factors. First, the CNN implicitly localizes the dorsal stripe by learning to attend to the relevant \nimage region, whereas the central crop approach averages over the entire frame including irrelevant areas. \nSecond, the CNN's whole-image representation is robust to common failure modes (blur, partial \nocclusion, oblique angles) that degrade mean hue and saturation measurements. Third, the CNN was \ntrained specifically on this classification task, encoding expert knowledge about what constitutes \"red-\nback\" versus \"lead-back\" appearance under real photographic conditions. The threshold classifier embeds \na simplified model of morph appearance (warm hue + adequate saturation) that holds for well-lit, dorsum-\nvisible photos but degrades under adverse conditions, producing the conservative 40.5%/39.2% red-\nback/lead-back split (versus 75.9%/24.1% in Hantak et al.) and an ambiguous fraction of 20.3%. The crop \naudit provides a direct explanation for this ambiguous rate: 21.0% of audited images were scored \nin_hand, and crops dominated by hand or glove pixels lack the color signal needed to resolve morph \nidentity. The CNN approach of Hantak et al. (2022) does not produce an analogous ambiguous class \nbecause the whole-image deep learning architecture implicitly localizes the dorsal region and is trained to \nclassify every input; the threshold approach propagates crop failures directly into ambiguous calls. \nThe progression across methods, from continuous brightness (R² = 0.001) to a simple hue threshold (R² = \n0.008) to a supervised CNN (pseudo-R² \n≈  0.04), quantifies the signal-to-noise problem from a \nmethodological perspective: more sophisticated classifiers extract more signal from the same \nphotographs, but require correspondingly greater investment in training data and model development. \nObserver Novelty Bias and Sampling Distortion \nMcCormick and Riley (2025) documented that iNaturalist observations of P. cinereus in New Brunswick \noverrepresent rarer morphs relative to concurrent field surveys (unstriped: 6.6% on iNaturalist vs. 3.3% in \nthe field; χ ² = 5.83, p = 0.02), attributing this to citizen scientists' preference for photographing visually \nunusual individuals. The present dataset shows a related pattern: even after conservative morph \nclassification, 39.2% of classified observations are lead-back. Published estimates of unstriped morph \nfrequency vary substantially across the species' range — from <5% in many northern populations (Lotter \nand Scott 1977; McCormick and Riley 2025) to 20–26% range-wide (Gibbs and Karraker 2006, \ncompiling 50,960 individuals from 558 sites; see also Hantak et al. 2021) — but the 39.2% observed here \nsubstantially exceeds even the highest published field estimates. Moore and Ouellet (2015), analyzing \n236,109 observations from 1,148 localities, found no significant climatic or geographic influence on \nmorph proportions, suggesting that the geographic pattern itself remains contested. Hantak et al.'s (2022) \nmore accurate CNN classifier obtained 24.1% lead-back in their iNaturalist dataset — still above field \nsurvey estimates, confirming that the bias operates at the sampling (who gets photographed) rather than \nthe classification (who gets correctly identified) level. \nThis sampling bias has direct consequences for geographic inference. If the probability that an observer \nphotographs an unusual morph varies by region — for example, if birding communities in the \nnortheastern United States are more attentive to unusual herps than communities in the mid-Atlantic — \nthen morph frequency estimates from iNaturalist will reflect photographer demographics as well as true \nbiology. The geographic cell ICC from the variance decomposition (5.1%) is consistent with this: part of \nwhat looks like \"geographic\" brightness variation is likely geographically structured photographer \nbehavior. The morph frequency R² = 0.008 should therefore be interpreted cautiously; the modest \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n17 \ngeographic gradient may partly reflect spatial variation in observer novelty-seeking behavior rather than \ntrue morph frequency clines. \nImplications for Citizen Science Phenotyping \nThese results support a simple, practical framework for evaluating citizen science photos as phenotypic \ndata sources. Continuous quantitative traits (brightness, size measurements, meristic counts) are \nvulnerable to observer variance because the measurement is a direct function of the raw pixel values, \nwhich are dominated by photographic conditions. Without standardized photography protocols — \nconsistent backgrounds, calibrated color cards, known focal distances — observer identity will typically \nexplain more variance in continuous trait estimates than any biological variable of interest. Photometric \nnormalization can reduce within-photographer variance substantially, but cannot recover geographic \nsignal that is smaller than the inter-photographer offset distribution. \nDiscrete categorical traits are more robust because classification thresholds can absorb photographic \nvariation. Red-back and lead-back morphs differ dramatically in dorsal hue — warm orange-red versus \ndesaturated gray-brown — a contrast that is perceptually and photometrically large enough to survive \nsubstantial photographic noise. However, as McCormick and Riley (2025) and the present data both \ndemonstrate, the sampling process itself introduces a second layer of bias — observer preferences — that \nclassification accuracy alone cannot correct. The implication is that citizen-science photo data may be \nmost valuable for documenting the existence and distribution of discrete phenotypic variants rather than \ntheir frequencies or quantitative values. \nThe autoresearch loop provides a general way to determine where a given trait falls along this spectrum. \nBy formalizing pipeline evaluation as a composite metric and running a bounded search, researchers can \nassess whether any parameter configuration extracts geographic signal above the noise floor before \ncommitting to large-scale analysis. In that sense, the negative result reported here, namely that no tested \nparameter configuration recovered a brightness cline in Plethodon, is itself informative. \nFuture Directions \nThree extensions would substantially strengthen the present analysis. First, adding a background-\nsubtraction step using a pretrained segmentation model (e.g., SAM; Kirillov et al. 2023) would allow \nbrightness extraction from confirmed dorsal pixels only, reducing the substrate-leakage component of \nresidual variance. Given the magnitude of observer effects, this would likely improve precision more than \nreveal new geographic signal, but it would sharpen the inference. Second, a stratified subsampling design \nthat selects one photo per observer per H3 cell would decorrelate observer identity from geography, \nreducing the observer ICC at the cost of sample size. At the estimated 23.3% observer ICC and 33,496 \nunique observers, such a dataset would still exceed 10,000 observations. Third, extending the Hantak et \nal. (2022) CNN framework to the full Plethodon genus, through species-specific or transfer-learned \nmodels for additional polymorphic species, would test whether the geographic morph signal observed in \nP. cinereus generalizes across the clade. \nThe methodological contribution of this study, namely the formalization of citizen-science image-pipeline \noptimization as a bounded, logged, multi-experiment search, is independent of the biological result. I \nrecommend this approach as a standard component of research programs that seek to extract quantitative \nphenotypic data from opportunistically collected photographs. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n18 \nData and Code Availability \nCode for data acquisition, image processing, optimization, and analysis is archived at \nhttps://doi.org/10.5281/zenodo.19050224. Derived experiment logs and manuscript figures referenced \nhere were generated from that repository. The observation metadata analyzed in this study were retrieved \nfrom iNaturalist through the iNaturalist API; the platform-level occurrence dataset is cited in the \nReferences as a general source description for iNaturalist records. \nAcknowledgments \nI thank iNaturalist and its community of contributors for making this dataset possible, Alex Pyron for his \ncritical review, and Jessica Nadler for her improvements on content.  \nConflict of Interest \nThe author declares no competing interests. This research was conducted independently and does not \nrepresent the views of Deloitte LLP. \nFunding \nThis research received no external funding. Cloud computing costs (~$21) were funded by the Deloitte \nFederal Health AI initiative.  \nReferences \nBrodsky, I., Friend, A. J., and the h3-py contributors. 2018. h3-py: Python bindings for H3, a hierarchical \nhexagonal geospatial indexing system. Uber Technologies. https://github.com/uber/h3-py \nClusella-Trullas, S., van Wyk, J. H., and Spotila, J. R. 2007. Thermal melanism in ectotherms. Journal of \nThermal Biology 32:235–245. https://doi.org/10.1016/j.jtherbio.2007.01.013 \nDelhey, K. 2019. A review of Gloger's rule, an ecogeographical rule of colour: definitions, interpretations \nand evidence. Biological Reviews 94:1294–1316. https://doi.org/10.1111/brv.12503 \nDi Cecco, G. J., Barve, V., Belitz, M. W., Stucky, B. J., Guralnick, R. P., and Hurlbert, A. H. 2021. \nObserving the observers: how participants contribute data to iNaturalist and implications for \nbiodiversity science. BioScience 71:1179–1188. https://doi.org/10.1093/biosci/biab093 \nEndler, J. A. 1980. Natural selection on color patterns in Poecilia reticulata. Evolution 34:76–91. \nhttps://doi.org/10.1111/j.1558-5646.1980.tb04790.x \nGibbs, J. P., and Karraker, N. E. 2006. Effects of warming conditions in eastern North American forests \non red-backed salamander morphology. Conservation Biology 20:913–917. \nhttps://doi.org/10.1111/j.1523-1739.2006.00375.x \nHairston, N. G. 1951. Interspecies competition and its probable influence upon the vertical distribution of \nAppalachian salamanders of the genus Plethodon. Ecology 32:266–274. \nhttps://doi.org/10.2307/1930418 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint \n\n19 \nHantak, M. M., Page, R. B., Converse, P. E., Anthony, C. D., Hickerson, C. M., and Kuchta, S. R. 2019. \nDo genetic structure and landscape heterogeneity impact color morph frequency in a polymorphic \nsalamander? Ecography 42:1383–1394. https://doi.org/10.1111/ecog.04534 \nHantak, M. M., Federico, N. A., Blackburn, D. C., and Guralnick, R. P. 2021. Rapid phenotypic change in \na polymorphic salamander over 43 years. Scientific Reports 11:22681. \nhttps://doi.org/10.1038/s41598-021-02124-2 \nHantak, M. M., Guralnick, R. P., Zare, A., and Stucky, B. J. 2022. Computer vision for assessing species \ncolor pattern variation from web-based community science images. iScience 25:104784. \nhttps://doi.org/10.1016/j.isci.2022.104784 \nHighton, R., Maha, G. C., and Maxson, L. R. 1989. Biochemical evolution in the slimy salamanders of \nthe Plethodon glutinosus complex in the eastern United States. Illinois Biological Monographs 57:1–\n153. \nHighton, R. 1995. Speciation in eastern North American salamanders of the genus Plethodon. Annual \nReview of Ecology and Systematics 26:579–600. \nhttps://doi.org/10.1146/annurev.es.26.110195.003051 \niNaturalist contributors, iNaturalist. 2026. iNaturalist Research-grade Observations. Occurrence dataset. \nAccessed 2026-03-12. https://doi.org/10.15468/ab3s5x \nJaeger, R. G. 1971. Competitive exclusion as a factor influencing the distributions of two species of \nterrestrial salamanders. Ecology 52:632–637. https://doi.org/10.2307/1934151 \nKarpathy, A. 2026. autoresearch. GitHub repository. Accessed 2026-03-12. \nhttps://github.com/karpathy/autoresearch \nKirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. \nC., Lo, W.-Y., Dollar, P., and Girshick, R. 2023. Segment anything. 2023 IEEE/CVF International \nConference on Computer Vision (ICCV), 3992–4003. \nhttps://doi.org/10.1109/ICCV51070.2023.00371 \nLotter, F., and Scott, N. J., Jr. 1977. Correlation between climate and distribution of the color morphs of \nthe salamander Plethodon cinereus. Copeia 1977:681–690. https://doi.org/10.2307/1443166 \nMcCormick, A., and Riley, J. L. 2025. Integrating ecological and community science data to understand \npatterns of colour polymorphism and social behaviour at the northern range limit of a plethodontid \nsalamander. PLOS ONE 20(9):e0332501. https://doi.org/10.1371/journal.pone.0332501 \nMcLean, C. A., and Stuart-Fox, D. 2014. Geographic variation in animal colour polymorphisms and its \nrole in speciation. Biological Reviews 89:860–873. https://doi.org/10.1111/brv.12083 \nMoore, J.-D., and Ouellet, M. 2015. Questioning the use of an amphibian colour morph as an indicator of \nclimate change. Global Change Biology 21:566–571. https://doi.org/10.1111/gcb.12744 \nSchmidgall, S., Su, Y., Wang, Z., Sun, X., Wu, J., Yu, X., Liu, J., Moor, M., Liu, Z., and Barsoum, E. \n2025. Agent Laboratory: Using LLM agents as research assistants. Findings of the Association for \nComputational Linguistics: EMNLP 2025, 5977–6043. https://doi.org/10.18653/v1/2025.findings-\nemnlp.320 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 27, 2026. ; https://doi.org/10.64898/2026.03.24.713936doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}