{"id":"e9d7293f-3703-4a67-82b9-5a266ec22bb9","arxiv_id":"2412.19439","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A constrained neural network that tunes only the peak wavelength of simulated opsins reproduces qualitative evolutionary color-vision trends and proposes minimal camera filter designs.","lead":"This paper trains a neural network layer that mimics the wavelength sensitivity of eye opsins, using image recognition accuracy as a stand-in for evolutionary fitness. It claims to reproduce evolutionary shifts in color vision (mammal dichromacy, primate trichromacy, fish blue-shift) and to design task-specific camera filters for Mars and cancer imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The trainable C-to-3 layer after the opsin layer confounds the mIoU fitness signal, so the reported lambda_max evolution and camera filter designs may reflect joint network adaptation rather than intrinsic opsin advantages.","rationale":"The reader correctly flags the HSI-reconstruction assumption as a serious data-level risk, and I agree that inaccurate reconstructed spectra would invalidate the color-blindness and camera-design results. However, the most load-bearing concern is internal and affects every experiment, including those using real LIB-HSI data. The C-to-3 convolution, described only in Sec. 7.1 of the supplement, is a trained linear layer between the opsin layer and the encoder. Because it is optimized simultaneously with lambda_max, the segmentation loss can be reduced by adapting this projection rather than by changing opsin sensitivities. Consequently, the reported lambda_max values are not isolated causes of fitness; they are one component of a heavily over-parameterized end-to-end optimization. This undermines the claim that the framework 'quantifies evolutionary pressures' through opsin-filtered recognition accuracy, since the recognition system can compensate for any opsin setting. The gene-duplication experiment is particularly vulnerable: starting from identical 560 nm kernels, the split to ~540 and ~580 nm may be driven by random initialization of the C-to-3 layer and encoder, not by natural selection. This is a concrete, testable confound, and it warrants a major revision before the central claims are accepted. The reader's verdict of CONDITIONAL is appropriate, but the conditions should include ablating the C-to-3 layer or otherwise isolating the opsin contribution.","tokens_in":17541,"tokens_out":9929,"duration_ms":89301,"concrete_test":"Retrain the Sec. 4.2 mammal and Sec. 4.3 gene-duplication experiments with the C-to-3 layer removed (e.g., replicate or zero-pad opsin channels to 3 channels) and with at least 5 random seeds. If the final lambda_max distributions overlap with the reported values and the gene-duplication split consistently yields roughly 540 and 580 nm, the C-to-3 confound is minor; if the trajectories shift or the split disappears, the reported lambda_max evolution is an artifact of the learned projection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that mIoU on opsin-filtered images quantifies evolutionary fitness of the opsin spectral sensitivities (Sec. 1). However, the opsin layer output is projected through a trainable 1x1 convolution (C-to-3 layer, Sec. 7.1, Fig. 5) before the MiT encoder. This layer is trained jointly with the opsin layer and the encoder/decoder, so the network can arbitrarily recombine opsin channels. For C=2 (mammal, Sec. 4.2) and C=3 (primate, Sec. 4.3; camera, Sec. 5), the C-to-3 layer can mix channels; for C=1 (blue-shift, Sec. 4.5.1) it can rescale the single channel. Thus mIoU reflects the whole learned front-end, not the opsin sensitivity functions in isolation. The gene-duplication split in Sec. 4.3 could result from symmetry breaking caused by random initialization of the C-to-3 layer rather than from a selective advantage of two distinct lambda_max values. Without ablating or freezing this layer, or reporting multiple seeds, the biological interpretations are not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a computational framework, called evolutionary conservation optimisation, to model colour vision evolution. It introduces an opsin layer whose convolutional kernels are Gaussian approximations of spectral sensitivity functions, with only the centre wavelength λ_max updated under a per-epoch step cap, and uses semantic segmentation mIoU on hyperspectral images as a proxy for evolutionary fitness. The framework is applied to reconstruct several proposed evolutionary transitions: the loss of two cone opsins in mammals, primate trichromacy via gene duplication, retention of colour blindness, blue-shift of squirrelfish rod opsins, and multiple rod opsins under bioluminescence. It is also used to speculate about Martian colour vision and to design task-specific camera spectral filters for Mars and cancer detection. The paper claims that the model quantitatively verifies long-standing biological hypotheses and provides a minimalist camera-design paradigm.","tokens_in":17780,"tokens_out":4525,"duration_ms":41028,"significance":"If the framework were validated, it would provide a fast, in-silico platform for generating quantitative hypotheses about opsin evolution and a practical recipe for task-specific spectral filter design. The opsin-layer parameterisation is simple and the evolutionary conservation constraint is an interesting way to inject biological plausibility into a differentiable pipeline. The paper also connects with a currently active area of minimalist, task-driven camera design. However, the reported experiments do not yet support the central quantitative claims: the fitness signal is confounded by a trainable layer at the front end; several target phenomena are effectively inserted into the experimental setup; there are no error bars or repeated seeds; and key datasets are generated by RGB-to-HSI reconstruction. The manuscript's strengths are the conceptual framework and the breadth of application scenarios, but the evidence as presented is not sufficient to establish the biological verifications or the practical camera-design recommendations.","major_comments":[{"comment":"The C-to-3 1×1 convolution layer, placed between the opsin layer and the MiT encoder, is trainable and is trained jointly with the opsin layer, encoder, and decoder. For C=2 and C=3 it can arbitrarily recombine the opsin channels, and for C=1 it can rescale the single channel. Therefore the mIoU values in Secs. 4.2–4.6 and Sec. 5 reflect the joint adaptation of the whole front end, not the opsin spectral sensitivities in isolation. The claim in Sec. 1 that recognition performance on images filtered through specific opsins quantifies the opsin advantage is not supported unless this layer is ablated (e.g., frozen to an identity-like matrix), or controlled by random fixed projections, or by multiple seeds that show the outcome is insensitive to the C-to-3 initialization. Without such an ablation, the biological interpretations of the evolved λ_max values are confounded.","section":"Supplementary Sec. 7.1, Fig. 5; main-text Sec. 3.4"},{"comment":"The blue-shift result follows directly from the input model. In Eq. (4), E(d,λ) = E(0,λ)e^{−Kd(λ)d}, and the diffuse downwelling attenuation Kd(λ) attenuates long wavelengths more strongly with depth, so the ambient spectrum is constructed to shift to shorter wavelengths as d increases. Optimizing a single Gaussian λ_max on such inputs is expected to produce a decreasing λ_max with depth; this is an extraction of the known spectral trend, not an independent verification of the blue-shift hypothesis. Similarly, the multi-rod-opsin experiment (Supp. Sec. 8.1) creates bioluminescence by exempting a specific label region from the diffusion attenuation, thereby injecting a local spectral feature that the multi-kernel optimization can lock onto. The paper should present these as illustrative reconstructions or add null controls (e.g., input spectra without depth-dependent attenuation) to show that the optimization does not simply recover the designed-in trend.","section":"Sec. 4.5.1, Eq. (4); Sec. 4.5.2, Supp. Sec. 8.1"},{"comment":"Every quantitative result is reported as a single mIoU or SR value, with no standard deviation, confidence interval, significance test, or multiple random seeds. Given that some claimed improvements are very small (e.g., Tab. 6: 39.15 vs. 39.20; Tab. 4: 41.01 vs. 40.75), the differences may be within run-to-run variation. The conclusions that one visual system 'outperforms' another, or that a designed filter 'enhances performance,' are not statistically supported. The paper should provide repeated runs with distinct initializations, report mean ± std, and ideally perform paired significance tests for the comparisons that drive the biological and camera-design claims.","section":"All result tables (e.g., Tab. 2, Tab. 3, Tab. 4, Tab. 6, Tab. 7)"},{"comment":"The hyperspectral data for MinneApple, VOC2012, and Mars-Seg are reconstructed from RGB using the method of [6]. The optimized λ_max values are therefore attributes of the reconstruction network's estimated spectra, not of the actual scene radiance. This is especially load-bearing for the Mars camera design (Sec. 5.2) and the Martian vision prediction (Sec. 4.6), where the designed filters are trained on synthetic HSI derived from RGB images. The paper should either validate the reconstruction's spectral accuracy on a dataset with measured HSI, or reframe these results as demonstrations on synthetic spectra rather than as recommendations for real Mars-exploration or medical cameras.","section":"Sec. 4.1.2; Sec. 5.2"}],"minor_comments":[{"comment":"The camouflage score SR is introduced in the main text only by name; it should be defined in the main text or the main-text table should explicitly state that it is the reconstruction fidelity score from the supplementary material. This is currently unclear because Tab. 2 uses SR without defining it.","section":"Sec. 4.3, Supp. Sec. 7.2"},{"comment":"There appears to be an inconsistency: Tab. 2 lists di-vision SR 0.2195 and tri-vision SR 0.3161, while Tab. 3 reports the 'normal' column as 0.2195 in bright conditions. If 'normal' is meant to be normal trichromatic vision, the value should match the tri-vision value of 0.3161; if it is meant to be dichromatic, the column label is wrong. Please clarify the correspondence between the two tables.","section":"Tab. 3 vs. Tab. 2"},{"comment":"The main text says 'As shown in Tab.17' but refers to the 3-filter camera result; the table in the main text is Tab. 7, while Tab. 17 appears in the supplementary material. This cross-reference should be corrected.","section":"Sec. 5.3, Tab. 7"},{"comment":"There are some typographical errors, including the duplicated 'to' in the abstract ('adaptations to to more effectively spot fruits') and 'clour-blind' in Sec. 4.4. The manuscript should be proofread for such issues.","section":"Abstract; Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is interesting conceptually but the experimental validation needs substantial strengthening. The C-to-3 confound is a serious issue because it undermines the attribution of fitness differences to the opsin layer itself; the circularity of the blue-shift and bioluminescence setups is also concerning. These are not merely presentation problems, and the authors should be asked to add ablations, controls, repeated seeds, and real or validated HSI data before the claims of quantitative verification are made. That said, the framework may be salvageable as a proof-of-concept platform if the claims are appropriately scaled back and the experiments are made statistically sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously: a Gaussian-parameterized opsin layer trained with lambda_max-only updates, capped at 0.5 nm per epoch, is a clean formalization of spectral tuning biology. That combination is new, and it could become a useful simulation platform for color-vision evolution and task-specific camera filters. The paper is also honest about its simplifications, especially in the multi-rod-opsin section where it admits the 5 nm separation is much smaller than observed data.\n\nThe problems are mostly in the evidence supporting the quantitative claims. Every table reports a single mIoU or SR value with no error bars, and no multiple seeds. For claims like \"trichromatic vision outperforms dichromatic vision\" or \"blue-shift is caused by light diffusion,\" that is not enough. Worse, the HSI for MinneApple, VOC2012, and Mars-Seg is reconstructed from RGB by a neural network, and the opsin layer then operates on those estimated spectra. If the reconstruction is biased, every optimized lambda_max inherits that bias. This is load-bearing, not a minor caveat.\n\nThe stress-test concern about the trainable C-to-3 convolution layer is correct on reading. That layer sits right after the opsin layer and is trained jointly with it, so the network can recombine opsin channels arbitrarily. mIoU reflects the whole learned front-end, not the opsin sensitivities in isolation. I would want to see the C-to-3 layer frozen or ablated (e.g., identity mapping) before trusting the biological interpretations. The gene-duplication split could easily be symmetry breaking in that layer rather than a selective advantage.\n\nThere is also a circularity problem in the evolutionary experiments. The underwater model in Eq. 4 already attenuates longer wavelengths more with depth, so the blue-shift in Table 5 is largely constructed by the data-generation process. The bioluminescence simulation likewise exempts one region from attenuation, which is almost a direct encoding of the hypothesis. The paper should frame these as consistency checks, not independent verifications. The causal language in Sec. 4.5.1 overreaches.\n\nTwo smaller issues: the camera-design gains in the main text are marginal (0.05 and 0.38 mIoU), and Table 3 appears to have swapped labels — the \"normal\" column matches the di-vision value in Table 2, which undermines the colour-blindness advantage claim until corrected.\n\nWho should read this? Biologists working on computational evolution and vision researchers interested in learned spectral filters. It deserves a serious referee, but not acceptance in current form. The authors need multi-seed runs, a frozen or ablated C-to-3 layer, real HSI or a sensitivity analysis on the reconstruction, and much more careful causal language. With those revisions, this could become a valuable tool. Send it to review with a request for major revision.","headline":"Genuinely novel opsin-layer framework with a clean conservation constraint, but the quantitative support is not there yet: single-run tables, RGB-reconstructed HSI, and a trainable C-to-3 layer that confounds the fitness signal.","tokens_in":18340,"tokens_out":3910,"would_cite":false,"duration_ms":34986,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A one-layer opsin model, evolved by segmentation accuracy, reproduces colour-vision transitions and designs task-specific camera filters.","keywords":["colour vision evolution","opsin layer","evolutionary conservation optimisation","spectral sensitivity","hyperspectral image reconstruction","semantic segmentation","task-specific camera design","primate trichromacy"],"falsifier":"Train the same opsin-layer pipeline on a dataset with true measured hyperspectral radiance (for example, real HSI of apples on leaves) and compare the converged $\\lambda_{\\max}$ trajectories with those obtained from RGB-reconstructed spectra; a shift of more than about 10 nm, or failure of a duplicated long-wavelength opsin to split into two distinct peaks, would show that the evolutionary conclusions depend on the reconstruction network rather than on biology.","tokens_in":17301,"feed_emoji":"👁️","tokens_out":16507,"duration_ms":127170,"temperature":0.7,"pith_summary":"Evolutionary hypotheses about colour vision are usually tested through genetic or behavioural experiments. This paper argues that a drastically simplified computational surrogate—a single 'opsin layer' of Gaussian-shaped spectral filters whose only learnable parameter is the peak wavelength $\\lambda_{\\max}$, driven by segmentation accuracy as a stand-in for fitness—can replay those evolutionary transitions in seconds on a GPU. Using hyperspectral images, including spectra reconstructed from ordinary RGB, the authors reproduce the mammalian loss of two cone opsins, primate trichromacy from gene duplication, the advantage of colour blindness in dim light, the blue-shift of fish rod opsins with depth, and a modest opsin split under simulated bioluminescence. The same optimisation, applied to camera filters, yields task-specific spectral response functions for Martian terrain and cancer-tissue segmentation with small but consistent gains over generic RGB filters. If the spectral reconstructions are faithful, the framework offers a fast, quantitative platform for both evolutionary 'what-if' questions and minimalist camera design.","feed_headline":"Opsin model replays colour-vision evolution on a GPU","feed_subtitle":"Segmentation accuracy as fitness reproduces primate trichromacy and fish blue-shift, then tunes camera filters for Mars and cancer.","key_machinery":"The load-bearing machinery is the opsin layer combined with the evolutionary conservation optimisation. The opsin layer is a $1\\times1$ convolutional layer with $C$ Gaussian kernels, each kernel acting as a spectral sensitivity function $\\psi_c$; because integrating a hyperspectral pixel's radiance against an opsin sensitivity function is the same as a dot product, one convolution maps an $H\\times W\\times N$ hyperspectral image to an $H\\times W\\times C$ feature map. The conservative regularisation keeps the Gaussian shape and width $\\sigma$ fixed, allows only the peak wavelengths $\\lambda_{\\max,c}$ to move, and limits each move to $0.5$ nm per epoch, encoding the biology of spectral tuning sites. A MiT-B0 encoder plus a lightweight all-MLP decoder converts those feature maps into segmentation maps, and the segmentation loss supplies the selection pressure. This setup is what converts millions of years of evolution into a few hundred GPU iterations, and it is also the mechanism that lets the authors treat camera filter design as the same optimisation problem.","core_discovery":"The paper's central claim, stated on its own terms, is that the evolutionary history of vertebrate colour vision is, to a first approximation, a smooth drift of a few Gaussian peak wavelengths under selection, and that this drift can be captured by a single differentiable layer. The opsin layer is a $1\\times1$ convolution whose kernels are fixed-width Gaussians $\\psi_c(\\lambda_i)=\\frac{1}{\\sqrt{2\\pi}\\sigma}e^{-(\\lambda_i-\\lambda_{\\max,c})^2/2\\sigma^2}$, and the only parameter allowed to change during optimisation is $\\lambda_{\\max}$, capped at $0.5$ nm per epoch to mimic real spectral tuning sites. Selection is supplied by cross-entropy loss of a MiT-B0 encoder with an all-MLP segmentation decoder, so mIoU plays the role of evolutionary fitness. From ancestral starting points the optimiser moves $\\lambda_{\\max}$ along trajectories that match known transitions: two mammalian opsins shift under dim light, a duplicated long-wavelength opsin separates into the trichromatic pair, squirrelfish rod opsin shifts from roughly 496 nm at the surface to 481 nm at 70 m depth, and a five-opsin layer develops a maximum 5 nm separation under simulated bioluminescence but not without it. The same loop, applied to camera spectral response functions, shifts filter peaks for Mars segmentation (611.88/522.93/425.90 nm versus a 580/540/425 nm baseline) and for cancer detection, supporting the proposal that task-specific camera filters can be evolved rather than hand-designed.","pith_inferences":["If the central claim holds, the same loop could be applied to other visually guided tasks with known spectral signatures—agricultural weed/crop discrimination, underwater monitoring, or industrial sorting—where a hand-designed filter set is currently the default; the paper does not test these.","A direct validation of the weakest link would be to compare converged $\\lambda_{\\max}$ values on true measured hyperspectral radiance versus RGB-reconstructed spectra, since this would show how much of the reconstructed biology is real.","Varying the 0.5 nm-per-epoch cap, or removing the Gaussian shape constraint, would reveal whether the reported evolutionary trajectories are robust biological attractors or artefacts of the regularisation schedule, which the paper keeps fixed.","The framework's fitness signal is segmentation accuracy of a specific deep network, not survival or reproduction; using different encoders or tasks as the fitness signal could change which $\\lambda_{\\max}$ values win, and that sensitivity is not explored."],"forward_implications":["The framework gives evolutionary biologists a cheap in silico testbed: hypotheses about opsin tuning under different light environments can be probed in seconds, without stochastic genetic screens.","It quantitatively supports the selective story behind primate trichromacy—better fruit detection in leaves—while also predicting conditions under which dichromacy outperforms trichromacy (dim light, khaki-coloured terrain), bearing on why colour blindness persists.","It reproduces the depth-dependent blue-shift of squirrelfish rod opsins from roughly 496 nm at the surface to 481 nm at 70 m, and shows a small opsin separation arising under simulated bioluminescence, linking bioluminescence to multiple rod opsin evolution.","The same optimisation pipeline transfers directly to camera sensor design: for Mars terrain and cholangiocarcinoma detection, the evolved filter peaks improve segmentation mIoU/IoU over standard RGB-inspired filters, suggesting that application-specific spectral response functions can be manufactured with small iterative filter changes.","Under the model's assumptions, hypothetical Martian organisms would be better served by dichromatic than trichromatic or tetrachromatic vision, giving a concrete prediction about the visual systems of alien life under Martian illumination."],"supporting_citations":[{"why":"Supplies the Mask-guided Spectral-wise Transformer used to reconstruct hyperspectral images from RGB for MinneApple, VOC2012, and Mars-Seg; every optimized peak wavelength depends on this reconstruction.","marker":"[6]"},{"why":"Establishes the equivalence between spectral integration and a 1x1 convolution, the modelling move that turns opsin sensitivity into a learnable layer.","marker":"[51]"},{"why":"Provides the MiT-B0 encoder and lightweight all-MLP decoder that convert opsin-layer outputs into segmentation maps, defining the fitness signal.","marker":"[69]"},{"why":"Justifies the Gaussian approximation of opsin spectral sensitivity used to parameterise each kernel by its peak wavelength and standard deviation.","marker":"[22]"},{"why":"Gives the ancestral vertebrate cone opsin wavelengths (620, 530, 450, 375 nm) used as initial conditions for the mammalian transition experiment.","marker":"[5]"},{"why":"Supplies the evolutionary history of vertebrate visual pigments used to set up and interpret the mammal and primate transitions.","marker":"[13]"},{"why":"The gene-duplication hypothesis for primate trichromacy that the paper reconstructs by initialising two opsins at the same wavelength and letting them split.","marker":"[30]"},{"why":"Provides the camouflage and reconstruction fidelity score used to quantify fruit-recognition ability in the trichromacy and colour-blindness comparisons.","marker":"[36]"},{"why":"The squirrelfish rhodopsin data (peak wavelengths 481 to 502 nm versus depth) that the blue-shift experiment reproduces and uses as its quantitative target.","marker":"[72]"},{"why":"Documents multiple rod opsins in deep-sea teleosts, the observation that the bioluminescence experiment is designed to explain.","marker":"[49]"}],"fun_headline_variants":["GPU evolution recapitulates colour vision history","Simulated opsin evolution designs better camera filters","One opsin layer replays vision evolution in seconds","Evolution on GPU: from fish rods to Mars cameras"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that segmentation accuracy on hyperspectral images reconstructed from RGB faithfully represents real spectral radiance and can serve as a proxy for evolutionary fitness; if the reconstruction is inaccurate, every optimized $\\lambda_{\\max}$ is an artifact of the reconstruction network.","fun_headline_variants_meta":{"raw":{"variants":["GPU evolution recapitulates colour vision history","Simulated opsin evolution designs better camera filters","One opsin layer replays vision evolution in seconds","Evolution on GPU: from fish rods to Mars cameras"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1690,"prompt_tokens":1069,"completion_tokens":621,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":561}},"tokens_in":685,"tokens_out":621,"duration_ms":6003,"temperature":1.0,"reasoning_tokens":561,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:35:48.880915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same opsin-layer pipeline on a dataset with true measured hyperspectral radiance (for example, real HSI of apples on leaves) and compare the converged $\\lambda_{\\max}$ trajectories with those obtained from RGB-reconstructed spectra; a shift of more than about 10 nm, or failure of a duplicated long-wavelength opsin to split into two distinct peaks, would show that the evolutionary conclusions depend on the reconstruction network rather than on biology.","supporting_citations":[{"cited_title":"Mask-guided spectral-wise transformer for efficient hyper- spectral image reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the Mask-guided Spectral-wise Transformer used to reconstruct hyperspectral images from RGB for MinneApple, VOC2012, and Mars-Seg; every optimized peak wavelength depends on this reconstruction."},{"cited_title":"Deeply learned filter response functions for hyperspectral reconstruction","cited_arxiv_id":null,"evidence_quote":"Establishes the equivalence between spectral integration and a 1x1 convolution, the modelling move that turns opsin sensitivity into a learnable layer."},{"cited_title":"GOV ARDOVSKII, NANNA FYHRQUIST, TOM REUTER, DMITRY G","cited_arxiv_id":null,"evidence_quote":"Justifies the Gaussian approximation of opsin spectral sensitivity used to parameterise each kernel by its peak wavelength and standard deviation."},{"cited_title":"Bowmaker","cited_arxiv_id":null,"evidence_quote":"Gives the ancestral vertebrate cone opsin wavelengths (620, 530, 450, 375 nm) used as initial conditions for the mammalian transition experiment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the evolutionary history of vertebrate visual pigments used to set up and interpret the mammal and primate transitions."},{"cited_title":"Mollon Kanwaljit S","cited_arxiv_id":null,"evidence_quote":"The gene-duplication hypothesis for primate trichromacy that the paper reconstructs by initialising two opsins at the same wavelength and letting them split."},{"cited_title":"The making and breaking of camouflage, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the camouflage and reconstruction fidelity score used to quantify fruit-recognition ability in the trichromacy and colour-blindness comparisons."},{"cited_title":"The Molecular Ba- sis of Adaptive Evolution of Squirrelfish Rhodopsins.Molec- ular Biology and Evolution, 21(11):2071–2078, 2004","cited_arxiv_id":null,"evidence_quote":"The squirrelfish rhodopsin data (peak wavelengths 481 to 502 nm versus depth) that the blue-shift experiment reproduces and uses as its quantitative target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents multiple rod opsins in deep-sea teleosts, the observation that the bioluminescence experiment is designed to explain."}],"review_version":1}