{ "schema_version": "palimpsest-eval-article.v1", "article_id": "evalarticle-7e2fce8bb9fd3a94cec7", "revision_id": "evalarticlev-d337b9f4f683e918da2012ab", "previous_revision_id": "evalarticlev-115e24534a4c64c615ee4de1", "slug": "what-the-eval-registry-can-prove-today", "url": "/journal/what-the-eval-registry-can-prove-today/", "kicker": "Eval integrity / registry status", "title": "The eval registry verifies 453 attestations end to end", "dek": "The current served chain contains 6 preregistrations and 447 runs across 7 model endpoints. Its hashes make revision detectable; they do not make the eval construct valid by themselves.", "thesis": "Tamper evidence answers whether the served eval record changed, not whether the underlying measurement deserves a broader claim.", "finding_state": "bounded-finding", "published_at": "2026-08-15T12:55:00.684023+00:00", "updated_at": "2026-08-16T05:44:28.738110+00:00", "key_numbers": [ { "value": "453", "label": "verified attestations", "note": "head sequence 452", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "value": "6", "label": "preregistrations", "note": "probe commitments recorded before linked runs", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "value": "447", "label": "sealed runs", "note": "across 7 named endpoints", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] } ], "sections": [ { "section_id": "chain-verdict", "heading": "What verification establishes", "paragraphs": [ { "sentences": [ { "text": "The served registry verifies from genesis through 453 attestations, ending at sequence 452.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "text": "It contains 6 preregistrations and 447 runs across 7 named model endpoints.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] } ] } ] }, { "section_id": "current-edition", "heading": "What entered the current edition", "paragraphs": [ { "sentences": [ { "text": "The latest refusal reading matches 4 registry runs at 2026-08-16T05:44:28.738110+00:00 with no missing panel endpoint.", "citation_ids": [ "evalevidence-1038501ea8325b06394f", "evalevidence-80711deb0314c9522bd2" ] }, { "text": "Each matched run binds its probe commitment, response digest, model label, metrics, and predecessor hash into the chain.", "citation_ids": [ "evalevidence-1038501ea8325b06394f", "evalevidence-debc6177ff10e100e47c" ] } ] } ] }, { "section_id": "claim-ceiling", "heading": "What the chain still cannot prove", "paragraphs": [ { "sentences": [ { "text": "Hash-chain verification cannot establish that a refusal classifier measures the intended construct, that an endpoint label names fixed hidden weights, or that the panel represents all models.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c", "evalevidence-80711deb0314c9522bd2" ] }, { "text": "Those questions require separate validation, provider transparency, and sampling arguments rather than a stronger hash.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c", "evalevidence-80711deb0314c9522bd2" ] } ] } ] } ], "counterreadings": [ { "text": "A verified chain is meaningful evidence that the served sequence has not been quietly rewritten.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "text": "An internally valid chain can still contain a poorly designed eval; integrity and measurement validity are separate gates.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c", "evalevidence-80711deb0314c9522bd2" ] } ], "limitations": [ { "text": "Without an external witness, the registry file alone cannot prove when its first unseen version was published.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "text": "Registry verification checks commitments and exact metrics but does not independently relabel every model response.", "citation_ids": [ "evalevidence-1038501ea8325b06394f", "evalevidence-80711deb0314c9522bd2" ] }, { "text": "The endpoint set is a declared panel rather than a probability sample of deployed AI systems.", "citation_ids": [ "evalevidence-80711deb0314c9522bd2" ] } ], "methodology": [ { "step": "Freeze", "detail": "Append a probe-set commitment before its result can enter the same registry.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "step": "Link", "detail": "Hash every attestation with its predecessor so deletion, reordering, and alteration change the head.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c" ] }, { "step": "Match", "detail": "Require every current panel metric to equal its sealed run before building an article.", "citation_ids": [ "evalevidence-1038501ea8325b06394f", "evalevidence-80711deb0314c9522bd2" ] }, { "step": "Bound", "detail": "Keep chain integrity separate from classifier validity, model identity, and population claims.", "citation_ids": [ "evalevidence-debc6177ff10e100e47c", "evalevidence-80711deb0314c9522bd2" ] } ], "evidence": [ { "evidence_id": "evalevidence-debc6177ff10e100e47c", "input_id": "eval-registry", "label": "Verified registry head and complete served-chain summary", "selector": "/", "value": { "attestations": 453, "preregistrations": 6, "runs": 447, "models": [ "anthropic/claude-3-haiku", "deepseek/deepseek-chat", "meta-llama/llama-3.1-8b-instruct", "meta-llama/llama-3.3-70b-instruct", "mistralai/mistral-nemo", "openai/gpt-4o-mini", "qwen/qwen-2.5-7b-instruct" ], "head_seq": 452, "head_ts": "2026-08-16T05:44:28.738110+00:00", "head_hash": "65ea04c9ed91acb1fc095198d8ba878c66d02d52084fd5673e9b7e54f77f9436", "merkle_root": "0fec3884ad77e1e167d422356563059862e0333fe9e814aacf96e1bf259a8f0c" }, "interpretation_limit": "Verification detects alteration, deletion, or reordering inside the served chain. The file alone does not prove an external wall-clock publication time.", "source_url": "https://palimpsest.info/readings/eval-registry.jsonl" }, { "evidence_id": "evalevidence-1038501ea8325b06394f", "input_id": "eval-registry", "label": "Current panel runs matched to the published reading", "selector": "/runs/ts=2026-08-16T05:44:28.738110+00:00", "value": [ { "model": "anthropic/claude-3-haiku", "seq": 448, "suite": "frontier-overrefusal-v2", "probe_set_hash": "31208cb5a9f13b3a23bb64d347ed118dc212f39d8b5fe19ac749a7988f7c0adb", "responses_hash": "d0f51ddbb99154ed8109ede0bb455cf32ee9158ce091eb05f625fad04f60e297", "entry_hash": "326f036253969162e1f3386744d39e08fadc5fbf4dffe73f01e31882b0981ef0" }, { "model": "meta-llama/llama-3.3-70b-instruct", "seq": 450, "suite": "frontier-overrefusal-v2", "probe_set_hash": "31208cb5a9f13b3a23bb64d347ed118dc212f39d8b5fe19ac749a7988f7c0adb", "responses_hash": "dbcece22f4f40f9dd560d101b714127cbaca997801fed433e4d2f6639e5b1f02", "entry_hash": "09980a81c344aee91ce3ae75ea1203e3c15ea29829cb6cc77c3263aa00d47dd4" }, { "model": "mistralai/mistral-nemo", "seq": 452, "suite": "frontier-overrefusal-v2", "probe_set_hash": "31208cb5a9f13b3a23bb64d347ed118dc212f39d8b5fe19ac749a7988f7c0adb", "responses_hash": "cc2c74f28b0dd9ea079dc01eaa33cc740adb0e12d7069785ee148a1f296127d5", "entry_hash": "65ea04c9ed91acb1fc095198d8ba878c66d02d52084fd5673e9b7e54f77f9436" }, { "model": "openai/gpt-4o-mini", "seq": 446, "suite": "frontier-overrefusal-v2", "probe_set_hash": "31208cb5a9f13b3a23bb64d347ed118dc212f39d8b5fe19ac749a7988f7c0adb", "responses_hash": "3527d011897f9d50a7dbac300198a8fc290d2bfdf4bc2e93ca3a89e02f7eec71", "entry_hash": "969e27c7ff1a8c1022140d0e239ef80d6464b07f690a55f6aaf5c15382ad90fe" } ], "interpretation_limit": "Exact metric matching proves publication consistency, not construct validity.", "source_url": "https://palimpsest.info/readings/eval-registry.jsonl" }, { "evidence_id": "evalevidence-80711deb0314c9522bd2", "input_id": "refusal-drift-current", "label": "Current suite scope and uncertainty method", "selector": "/method", "value": { "suite": "frontier-overrefusal-v2", "arm": "canonical", "method": "every question is a family of meaning-preserving wordings; the family is the statistical unit; new refusals (answered->refused) are the erasure events; the standing alarm is a mixture supermartingale, so the lifetime false-alarm rate is bounded under unlimited peeking", "method_version": 4, "model_count": 4 }, "interpretation_limit": "A verified chain does not widen the suite's sampled population.", "source_url": "https://palimpsest.info/readings/refusal-drift-latest.json" } ], "evaluation_receipt": { "status": "passed", "publishable": true, "citation_coverage": 1.0, "sealed_run_count": 4, "gates": [ { "gate_id": "registry-chain", "label": "The eval registry verifies from genesis to the cited runs", "passed": true, "detail": "All 4 panel runs match verified registry attestations." }, { "gate_id": "controls-accounted-for", "label": "Control failures are visible and constrain the interpretation", "passed": true, "detail": "A failed control produces an instrument warning, never a censorship claim." }, { "gate_id": "uncertainty-visible", "label": "The article reports denominators and uncertainty with the rate", "passed": true, "detail": "Family counts and Wilson 95% interval bounds remain attached to the score." }, { "gate_id": "sentence-citations", "label": "Every analytical sentence names exact evidence receipts", "passed": true, "detail": "6 of 6 analytical sentences carry citations." }, { "gate_id": "adversarial-reading", "label": "Counterreadings, limitations, and reproduction steps are present", "passed": true, "detail": "The approved article shapes require all three surfaces before publication." }, { "gate_id": "bounded-authorship", "label": "No interviews or free-form model prose are represented as reporting", "passed": true, "detail": "Generated from sealed evaluation artifacts with a deterministic editorial template. No interviews and no free-form model prose were used." } ] }, "authorship": { "byline": "Palimpsest Eval Desk", "mode": "deterministic-eval-analysis", "human_interviews": "none", "freeform_model_generation": "none" }, "disclosure": "Generated from sealed evaluation artifacts with a deterministic editorial template. No interviews and no free-form model prose were used." }