{"claim":{"id":"6837adaf-3802-4da9-9d4c-248c1b639e71","submission_id":"89f3e616-47b5-4fec-83c3-6e97bb825e39","submission_revision_id":"b8748ad6-fef6-4620-8813-993797eb35b2","builder":"builder-01","status":"active","claimed_at":"2026-10-15T15:18:30.177116Z","delivered_at":null,"ended_at":null,"delivery_note":null,"end_reason":null},"task":{"submission_id":"89f3e616-47b5-4fec-83c3-6e97bb825e39","submission_revision_id":"b8748ad6-fef6-4620-8813-993797eb35b2","revision_version":4,"submission_status":"under_review","submitted_at":"2026-09-05T04:02:25.143498Z","submitter_name":"Gün Kaynar","submitter_email":"gkaynar@andrew.cmu.edu","task":{"status":"submitted","version":4,"extra_fields":{"difficulty_self_test":{"model":"Claude Opus 5","harness":"Claude Code","estimated_score":"<1%"}},"submitted_at":"2026-09-05T04:02:25.143498+00:00","industry_domain":"Computational Biology / Genomics (translation regulation, ribosome profiling)","task_short_name":"riboseq_translational_buffering_instance_1","operating_system":"Linux","task_description":"**The authoritative instruction is `input/task_description.md`, shipped inside the input\nmaterials.** It is a 1,261-word brief and it is the single text a solver sees. Its SHA-256 is\n7e311ad4a98347b7d9eb3b32d3877bd2cffb09e1b1a6514cad1a9f66e4b04b8b, and that is byte-for-byte the\ntext used in the difficulty self-test. What follows is a summary for the reviewer; where the two\ndiffer, the brief governs.\n\nNothing about the analysis method or the hidden conditions of the experiment appears in the\nbrief. Those are set out under \"What the task actually tests\" in the verification section below,\nwhich is evaluator-only material and must not be shown to a solver.\n\n**Objective.** Recover the 100 translationally buffered genes from a matched ribosome-profiling\nand RNA-sequencing experiment spanning twenty donor-derived human cell lines.\n\nFor a gene, let x be its log2 mRNA level and y its log2 ribosome-footprint level across the twenty\nlines. The quantity of interest is beta, the slope of the true, error-free y on the true,\nerror-free x. beta = 1 means ribosome load tracks mRNA proportionally. beta < 1 means the gene is\nbuffered: translation partly cancels the mRNA difference. Rank genes by the evidence that\nbeta < 1 and report the strongest 100.\n\n**Starting state.** Eighty raw, untrimmed FASTQ libraries in `input/fastq/`: twenty cell lines, two\nassays, two replicates each. Footprint libraries were sequenced at 36 or 50 cycles, RNA libraries\nat 2 x 50. A sample sheet gives per-library assay, replicate, cycle count, three-prime adapter,\nunique-molecular-identifier layout and strandedness. The cell-line column is withheld for every\nfootprint library. A GENCODE v47 reference is provided: annotation, protein-coding transcript\nFASTA with coding-sequence spans in the headers, full transcript FASTA, a contaminant FASTA of\nribosomal, transfer and small RNA, and a variant-site panel for donor matching. A gene panel file\nlists the genes the libraries were restricted to and discloses how reads were subset.\n\n**Deliverables.**\n\n    output/library_assignment.tsv   40 rows: footprint library -> cell line, with the score\n                                    it was matched on\n    output/qc_metrics.tsv           80 rows, one per library: contaminant share, duplicate rate,\n                                    retained read lengths, P-site offset per read length,\n                                    in-frame share, and the foreign fraction and its sources\n    output/te_table.tsv             translation efficiency per analysis-set gene per cell line\n    output/gene_slopes.tsv          one row per analysis-set gene: beta, its standard error, and\n                                    the evidence statistic. This is the full result\n    output/buffered_genes.tsv       exactly 100 ranked rows, the top of gene_slopes.tsv\n    output/method_summary.json      estimated measurement-error variances, the median of the beta\n                                    column of gene_slopes.tsv, and the analysis-set size\n\nThe brief additionally fixes the reporting conventions that are units rather than insight: the\ncounting window, the normalisation and pseudocounts, the definition of the analysis set, and the\nexact output schemas. It leaves every methodological choice to the analyst.\n\n### Why the reads are simulated\n\nThe reads are simulated and the submission does not disguise that. Two properties of this\nparticular question make it necessary, and a third makes it defensible.\n\n**The estimand has no ground truth in real data.** The target is beta, the slope of the true,\nerror-free log2 footprint level on the true, error-free log2 mRNA level across cell lines. On a\nreal experiment nobody knows that number. An answer key would have to be one pipeline's estimate,\nand at twenty cell lines a per-gene slope carries a standard error large enough that such a key\nwould encode that pipeline's noise as though it were truth. Grading a second analyst against it\nwould reward agreement with one set of arbitrary choices rather than correctness. The simulator's\nparameter table is the only exact key available for this quantity.\n\n**A real-data version is partly answerable by retrieval.** Every public matched\nribosome-profiling and RNA-sequencing dataset publishes its per-gene count matrix, and several\npublish translation-efficiency tables outright. An agent that recognises the dataset can skip the\nanalysis.\n\n**Nothing in the data is an invented obstacle.** Every property the analysis has to handle is a\ndocumented feature of real ribosome profiling, reproduced at a rate the simulator records: the\nribosomal RNA that dominates footprint libraries, the non-periodic mRNA background that is the\nreason the field checks three-nucleotide periodicity at all, the read-length censoring that\nfollows from arithmetic, the two nucleases with different offset profiles, the identifiers that\ndo not fit in a 36-cycle read, the paralogous families that share coding sequence, the isoform\ncomposition that shifts between lines, the technical dispersion of library preparation, and the\nsample cross-contamination and lost labels that are the reason variant panels are used for\nidentity checking in the first place. `README.md` in the reference bundle gives the observed rate\nfor each so it can be compared against practice.\n\nOne parameter sits above what is typical and is flagged rather than hidden: the foreign fraction\nin the affected footprint libraries runs 11 to 25 percent, higher than index hopping alone would\nproduce. It reflects a handling failure, which is what the brief tells the analyst, and it was\nset in that range by measurement: below it the effect is not correctable at twenty lines, above it\nunmixing destroys the signal it is meant to recover.\n\n**What is real.** The workflow, the tool chain, the reference sequence, the coding-sequence spans,\nthe isoform structure, the paralog families and the contaminant references all come from GENCODE\nv47 and the human ribosomal and transfer RNA references. The adapter behaviour and insert-length\ncensoring were fitted against a real ribosome-profiling run, GEO GSE65912, run SRR1803147. What is\nsimulated is the reads.","software_licensing":"Free/Open Source","evaluation_criteria":null,"verification_method":"Gate and score, fully deterministic, no model judge. A gate failure scores zero.\n\n### What the analysis has to handle — EVALUATOR ONLY, not visible to the solver\n\nThese are properties of the experiment, not puzzles added to it; `README.md` in the reference\nbundle gives the rate at which each occurs here and why it occurs in real data. They are listed\nhere rather than in the task description because telling the analyst which ones are present would\nhand over most of the analysis.\n\nTwelve of them, each shown by the variant evidence below to change the result. The brief states\nthe facts, the reporting conventions and the output contract, and leaves the analysis methodology\nto the analyst. It does not say that the axes carry measurement error, which transcripts belong\nin a gene's mRNA, how to treat multimapped reads, or which read lengths are genuine footprints.\nThe counting window and the normalisation ARE fixed in the brief, because they are units rather\nthan insight, and because the length-bias gate cannot otherwise separate a defensible convention\nfrom a mistake.\n\nA second brief that spells out the method is included in the reference bundle. It is not the task;\nit is the control that measures how much of the difficulty comes from withholding methodology, and\na frontier model passes it.\n\n1. Establish library purity. Sixteen of the twenty cell lines have footprint libraries carrying\n   material from one other line, at 11 to 25 percent. The fraction is the same in both replicates\n   of a line, because both came from the same pellet, so replicate differences cannot reveal it,\n   and the RNA libraries are unaffected, so it does not cancel between the two axes of the\n   regression. Ordinary quality control does not see it: alignment rates, periodicity, duplicate\n   rates and the library-to-line assignment all look normal. The variant panel is the only place\n   it shows, which is why such panels are used for identity checking. Left uncorrected it drags\n   the reported slope to 0.83 against a true 0.99.\n2. Decide *whether* each library is affected, not merely by how much. Four lines are clean.\n   Assuming contamination everywhere is its own error: it invents a source on the clean libraries\n   and on all forty RNA libraries.\n3. Assign each footprint library to a cell line. The labels were lost, so the pairing must be\n   recovered from the data itself, either by genotyping the variant panel from the reads or by\n   matching expression deviations against the labelled RNA libraries.\n4. Trim adapters and collapse duplicates according to each library's own chemistry. Footprint and\n   RNA libraries do not share a chemistry, and unique molecular identifiers are present on some\n   libraries only.\n5. In the 36-cycle libraries the longest footprints run off the end of the read and carry no\n   adapter at all. Those reads have to go, because their true insert length is unknown; keeping\n   them injects a spurious pile of full-length inserts.\n6. Deplete contaminants against the provided contaminant FASTA. A size cut alone will not do it,\n   because the contaminant fragments overlap the footprint size range.\n7. Separate real footprints from the non-footprint background. Fifteen percent of each footprint\n   library is ordinary mRNA fragments that survived size selection: positioned anywhere on the\n   transcript, untranslated regions included, with no reading-frame preference. They align to the\n   transcriptome exactly as footprints do, so no contaminant filter removes them, and they sit\n   just outside each nuclease's own footprint range, so the usable set of read lengths differs\n   between the two nucleases and a single global size window picks them up. Only the\n   three-nucleotide periodicity of genuine footprints distinguishes them.\n8. Calibrate one P-site offset per read length from the start-codon metagene. Two footprint\n   libraries, RPF_L017 and RPF_L025, were digested with micrococcal nuclease rather than RNase I\n   and have a longer footprint with a shifted offset profile. They belong to two different cell\n   lines, so neither is paired with a same-line library sharing its chemistry.\n9. Align footprints to the protein-coding transcriptome and quantify mRNA at isoform level.\n   Several genes have two isoforms with different untranslated regions and different coding\n   sequences, and the isoform mix changes across lines.\n10. Paralogous families share near-identical coding sequence, a property of the human genome, so\n   their footprint reads have no unique best alignment. Discarding them removes reads from an\n   identifiable set of genes rather than shifting all genes together.\n11. Count P-sites inside the coding sequence with the first 15 and the last 5 codons excluded.\n12. Estimate beta for every gene in the analysis set, correcting for measurement error on both\n   axes. Library preparation does not measure every gene equally well, so the mRNA axis carries\n   technical error with a median of 0.22 log2 units and a tenth-to-ninetieth range of 0.09 to\n   0.55. An uncorrected regression is attenuated and reports a median slope of 0.63 against a\n   true 0.99. The error-variance ratio is a **per-gene** quantity, because both variances scale\n   like one over the gene's read count and counts span orders of magnitude; using one global\n   ratio instead costs 0.20 of F1. Replicate libraries are present in the data, which is what\n   makes the two error variances estimable.\n\n\n    Gate 1  All six output files present and conforming to the contract the brief states.\n            Headers must match exactly, in order, with no extra columns.\n            library_assignment.tsv must hold exactly the 40 footprint libraries, once each;\n            qc_metrics.tsv exactly the 80 libraries, once each. te_table.tsv and\n            gene_slopes.tsv must cover the same gene set, with no repeats, and every gene in\n            all three gene-level files must come from the panel, so the reported set cannot be\n            padded with invented rows. Standard errors must be positive. buffered_genes.tsv\n            must hold exactly 100 unique genes with rank running 1 to 100, and its beta, se and\n            z must all agree with gene_slopes.tsv, so the ranked list is a faithful view of the\n            full result rather than a separate claim. method_summary.json must carry all four\n            required keys as finite numbers, its n_analysis_genes must equal the row count of\n            gene_slopes.tsv, and its median_beta must equal that file's median. Every numeric\n            field must be finite, and on a footprint library every quality-control field must\n            be filled and in range; an RNA row may leave everything but the identifiers empty.\n            Genes outside the reference's own analysis set are allowed, since analysis sets can\n            differ at the margin; they simply cannot count as hits.\n    Gate 2  At least 38 of the 40 footprint libraries assigned to the correct cell line.\n    Gate 3  The read lengths named in retained_lengths must be exactly those named in\n            psite_offsets, and a blank or non-finite in-frame share is rejected outright rather\n            than compared, because a comparison against a missing value succeeds\n            silently. For every read length that the reference and the submission\n            both\n            retain, the reported P-site offset must equal the offset the simulator actually\n            used, and the reported in-frame share must be within 0.05 of the reference.\n            Checking against the simulator rather than against the reference pipeline matters:\n            an early trial was failed for being right, because the reference had itself\n            miscalled one library and read length.\n    Gate 4  Spearman correlation between the agent's translation-efficiency table and the\n            reference is at least 0.85, computed on genes in the analysis set.\n    Gate 5  Length-bias check. Regress the agent's translation efficiency minus the reference\n            against log coding-sequence length. The slope must lie within 0.05 of zero. This\n            catches a wrong counting window or a wrong P-site offset, which move short and long\n            genes in opposite directions.\n    Gate 6  Attenuation check. The median slope, RECOMPUTED BY THE GRADER from the submitted\n            per-gene slopes and taken over the reference's own analysis set, must lie between\n            0.88 and 1.15. Two things are deliberate here. It is recomputed rather than read\n            from method_summary.json, because a self-reported summary can disagree with the\n            result it claims to summarise. And it is taken over the evaluator's gene set rather\n            than the submission's, because a median over a set the submission chooses can be\n            moved simply by reporting more genes: padding an attenuated result with a few\n            hundred panel genes at a slope of 1.35 shifts a failing 0.80 to a passing value.\n            The submission must cover at least 80 percent of the reference analysis set for the\n            gate to be judged at all. The band is one-sided with a sanity cap because the bias\n            it polices pushes the slope downward. Measured on the released data: the true value\n            is 0.99, a full analysis reports 0.99, ignoring measurement error on the mRNA axis\n            reports 0.63, and ignoring cross-contamination reports 0.83.\n    Gate 7  Purity check. For each footprint library, the set of contaminating cell lines above a\n            0.02 floor must match the truth, and the total foreign share must be finite and\n            within 0.05; a blank is rejected rather than compared. At\n            most 3 of the 40 may be wrong on either count. This grades the contamination\n            directly rather than through its effect on the slope, which is what\n            frees gate 6 to police\n            attenuation alone.\n\n    Score   Full pass requires every gate plus F1 at least 0.40 against the true buffered genes\n            in the analysis set. For partial credit, report 0.7 * F1 + 0.3 *\n            Spearman correlation\n            between the reported ranking and the true slopes, clipped at zero.\n\n    The score is a floor against nonsense; the gates are the discriminator. The reference reaches\n    0.470 and chance is about 0.11. The floor is deliberately not placed high enough to catch the\n    one variant below that fails no gate, because doing so would leave a competent but slightly\n    different analysis two hundredths from failing for no stated reason.\n\n    Evidence that the rules are load-bearing. Variants of the reference pipeline, each violating\n    exactly one requirement, graded by the released grader against the released answer key.\n\n            variant                              median slope  gates failed   F1     pass\n            reference                                0.989        none        0.470  yes\n            contamination ignored                    0.825        6, 7        0.470  no\n            contaminant forced on every library      0.949         7          0.459  no\n            one global error-variance ratio          1.006        none        0.273  no\n            rank by slope, no correction             0.633         6          0.022  no\n            rank by evidence, no correction          0.633         6          0.131  no\n            whole coding sequence counted            0.988         5          0.459  no\n            one P-site offset for all                0.994         3          0.470  no\n            no periodicity filter                    0.981         3          0.393  no\n            mRNA summed over all isoforms            0.953         5          0.328  no\n            multimapped reads discarded              0.978         5          0.470  no\n            eight libraries mislabelled              0.812        2, 6, 7     0.317  no\n            multimapper sent to first hit            0.958        none        0.437  YES\n\n    Twelve of thirteen are rejected. Four of them keep a score at or above the reference's and are\n    caught only by a gate: ignoring the contamination, using one P-site offset for every read\n    length, discarding multimapped reads, and counting the whole coding sequence. The gates do\n    that work, not the score, which is why a submission cannot stumble into a pass. Conversely one\n    variant, the single global error-variance ratio, fails no gate at all and is caught only by\n    the score, at 0.273 against the reference's 0.470.\n\n    One limitation, stated rather than hidden. Sending each multimapped read to its first\n    alignment fails no gate and scores 0.437, so it passes. It is a lazy convention rather than an\n    outright error, and its damage is confined to the 76 analysis-set genes that sit in\n    near-identical paralog families, which is too small a share to move any gate. Making it fail\n    would mean either enlarging the paralog panel or setting the score bar at 0.45, and the latter\n    would put a fair analysis at risk for the sake of catching one unfair one.\n\n    Where the gates stop being able to judge. Two counting windows that are conventions rather\n    than mistakes were also run, and both fail the length-bias gate:\n\n            counting window of 20 and 10 codons   slope +0.102   fails\n            counting window of 10 and 3 codons    slope -0.100   fails\n\n    The reference sits at 0.000 and the smallest genuine error, summing mRNA over all isoforms,\n    at 0.084, so the gate cannot separate a defensible convention from a mistake. That is\n    precisely why the counting window is pinned in the task brief and is not left to the analyst,\n    and it marks the boundary of what this evaluation can fairly judge.\n\n    The gates were tested adversarially, not merely written. `grader/test_gates.py` takes the\n    reference output, breaks exactly one rule at a time, and asserts rejection. All 22 cases are\n    caught. They cover the schema (an extra column, a duplicated library row, a missing library,\n    a rank column that is not 1 to 100), the numbers (a non-finite slope, a non-positive standard\n    error, a blank in-frame share, a blank foreign fraction, a foreign fraction outside zero to\n    one), the internal consistency of the deliverables (gene sets that disagree between the two\n    gene-level files, beta or se or z disagreeing between the ranked list and the full result,\n    a method summary missing a key or carrying a non-finite value or contradicting the row count\n    or the median it claims to report), and the two ways of evading the attenuation gate: slopes\n    attenuated by 30 percent and honestly reported, which fails gate 6 on the science, and the\n    same slopes padded with several hundred extra panel genes at 1.35 to drag the median back\n    into the band, which fails gate 6 because the median is taken over the reference's gene set.\n\n    Control submissions, run against the released grader. An empty directory fails on schema and\n    scores zero. A submission with correct upstream work and honest per-gene statistics, but a\n    random hundred genes chosen from them, scores 0.214, F1 0.109. One with wholly invented\n    per-gene statistics, internally consistent so that it passes every structural check and\n    squeaks through the attenuation gate at 0.890, scores 0.097, F1 0.098. Both sit far below the\n    0.40 bar and are rejected on substance rather than on formatting.\n\nThe answer key is the simulator's parameter table, not the output of any pipeline. The reference\noutputs are a genuine pipeline run on the released data.\n\nThese gates are not theoretical. A measured trial against the released brief passed six of seven\nand failed gate 5 at +0.143 per log10 coding-sequence length, because the agent filtered\nalignments at MAPQ >= 2 and so discarded the paralog families. Its deviation from the reference\ncorrelates 0.989 with the discard-multimappers variant above. The full account, the agent's own\nreport, and the grader's per-gate verdict are in the difficulty self-test evidence upload.","software_and_version":"Linux command line, no GUI. Every version below is the one the reference analysis was actually produced with, and every one is constrained in the provided environment manifest (`input/runtime_env/environment.yml`).  Python 3.11 with pysam 0.24.0, numpy 2.4.6, pandas 3.0.5, scipy 1.17.1, statsmodels 0.15.0, biopython 1.88, scikit-learn 1.9.0. cutadapt 4.9. bowtie2 2.5.4. salmon 1.10.3. samtools 1.21. bcftools 1.21. bedtools 2.31. subread 2.0.6 (featureCounts). pigz 2.8.  No GPU, no network. CPU only.","specific_requirements":null,"expected_completion_time":null},"files":[{"file_id":"2e7dfdb4-f351-4efe-ab21-3ff9b5c1402d","filename":"input_part_03_rpf_3.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":457668507},{"file_id":"d207757b-b1c8-4947-87e8-f46e910018be","filename":"input_part_01_rpf_1.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":803024708},{"file_id":"c29a7b40-2766-4c3e-aca1-802db94f05e6","filename":"input_part_02_rpf_2.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":815239320},{"file_id":"a96ed65c-46ae-43c6-ab32-23e4e635687b","filename":"input_part_04_rna_1.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":847724116},{"file_id":"3408dbfc-83b4-4bef-a972-e595450ed806","filename":"input_part_05_rna_2.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":846022904},{"file_id":"60f47726-d035-4308-91ce-c2ba3d171579","filename":"input_part_06_rna_3.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":849114215},{"file_id":"5ddc4e1f-e7f2-46b6-9512-84df9452b82d","filename":"MANIFEST.txt","category":"input_materials","content_type":"text/plain","size_bytes":2535},{"file_id":"21fa1f5c-88d9-4bdf-b5fa-4ab8e517d75e","filename":"input_part_00_metadata_and_reference.tar.gz","category":"input_materials","content_type":"application/x-gzip","size_bytes":12422264},{"file_id":"1dc9b91b-8a6c-46b6-98c0-e3338f5665dd","filename":"SHA256SUMS.txt","category":"input_materials","content_type":"text/plain","size_bytes":668},{"file_id":"85a5319f-19d7-498f-8d07-b5ab434974a7","filename":"reference_and_evaluation.tar.gz","category":"reference_output","content_type":"application/x-gzip","size_bytes":6062660},{"file_id":"a88334cf-6abe-4630-bdd1-1a0574a6666e","filename":"difficulty_self_test_evidence.tar.gz","category":"difficulty_self_test_evidence","content_type":"application/x-gzip","size_bytes":715410},{"file_id":"ff0d483d-a46a-42b5-a50e-ca23d183ff61","filename":"difficulty_self_test_report.md","category":"difficulty_self_test_evidence","content_type":"text/markdown","size_bytes":10549}],"deep_review":{"overall_verdict":"Strong Accept","review_text":"**Overall Verdict:** Strong Accept — The submission is complete, professionally constructed, fully reachable, and supported by coherent inputs, reference artifacts, deterministic grading, and substantive difficulty evidence.\n\n**Completeness Check:**\n- Instruction: Complete — The authoritative brief is clear, internally consistent, and matches the declared SHA-256; it defines the estimand, analysis set, normalization, and exact six-file output contract.\n- Inputs: Complete — All seven input archives match their declared sizes and authoritative checksums; they contain 129 expected files representing 80 libraries, with valid references, metadata, and 120 readable FASTQ files.\n- Toolchain: Complete — A pinned, open-source Linux/CPU environment covers preprocessing, alignment, quantification, statistics, and grading without requiring a GPU or runtime network access.\n- Reference output: Complete — All six required deliverables are present with matching schemas and coherent dimensions, including 1,004 analysis genes, 100 ranked genes, 40 assignments, and 80 QC rows; separate truth tables support objective evaluation.\n- Rubric: Complete — Seven deterministic gates cover structure and key scientific failure modes, followed by an F1 threshold and defined partial-credit score; adversarial variants, baselines, and two documented self-tests substantiate discrimination.\n\n**File and Task Reachability:**\nAttachment classification: `reachable`. Task classification: `reachable`. I inspected `MANIFEST.txt`, `SHA256SUMS.txt`, all seven `input_part_*.tar.gz` archives, `reference_and_evaluation.tar.gz`, `difficulty_self_test_evidence.tar.gz`, and `difficulty_self_test_report.md`. Declared byte counts matched, the seven authoritative input checksums matched, archive paths and entry types were safe, all 120 inner FASTQ gzip streams validated, and representative records conformed to FASTQ structure and declared read layouts. The reference and self-test bundles were inventoried and their key instructions, schemas, truth tables, grader logic, trial reports, and representative records were inspected without executing submitted programs. No attachment was omitted; individual sequencing records were sampled rather than biologically interpreted exhaustively. Solver-visible inputs are separated from evaluator-only truth and grading assets. The completion path is practical on one CPU machine: unpack inputs, construct the pinned environment, preprocess and quantify the reads, infer assignments and purity, estimate gene slopes, write the six outputs, and apply the deterministic grader. License keys, accounts, tokens, and other credentials are not required from the contributor. Confidential, proprietary, or internal material may be evaluated as submitted evidence without requiring additional disclosure.\n\n**Special Screening:**\nNone. Public-reference provenance was not independently checked through search, but the inspected evidence provides no affirmative basis for a screening label; the simulated experiment is explicitly disclosed and is built around a professional genomics workflow rather than an AI-fabricated or defect-only fixture.\n\n**Recommended Changes:**\nThe submission already provides unusually strong input validation, evaluator separation, adversarial gate testing, and measured difficulty evidence. No changes are required.","reviewed_at":"2026-09-05T04:54:26.812909Z"},"claim":{"id":"6837adaf-3802-4da9-9d4c-248c1b639e71","submission_id":"89f3e616-47b5-4fec-83c3-6e97bb825e39","submission_revision_id":"b8748ad6-fef6-4620-8813-993797eb35b2","builder":"builder-01","status":"active","claimed_at":"2026-10-15T15:18:30.177116Z","delivered_at":null,"ended_at":null,"delivery_note":null,"end_reason":null}},"revision_is_current":true}