Corpus
What each document said, on each date it was read
Every document behind a score or an assessment is fetched, hashed and kept — every version, not just the latest. A vendor who rewrites a safety framework cannot rewrite what was already recorded. This is the only thing here that cannot be built retroactively: starting in two years gets you two years from then, never the two years before.
76
Documents watched
93
Versions on file
12
Days of history
since 2026-09-08
Changed since first recorded
The output nothing else in this market can produce, because producing it requires having held the earlier version.
| Document | Versions | First recorded | Latest change | |
|---|---|---|---|---|
| https://epoch.ai/data/ai-models | 14 | 2026-09-08 | 2026-09-20 | Compare → |
| https://www.apolloresearch.ai/science | 3 | 2026-09-08 | 2026-09-20 | Compare → |
| https://www.anthropic.com/transparency | 2 | 2026-09-08 | 2026-09-18 | Compare → |
| https://mlcommons.org/benchmarks/ailuminate/ | 2 | 2026-09-08 | 2026-09-10 | Compare → |
What is watched
51 model cards · 9 regulation · 7 independent evaluations · 4 usage policies · 4 vendor governance pages · 1 safety frameworks. Model cards come from the automated scan; everything else is a document an assessment rests on, and is watched because an assessment cites a document as it read on a date.
| Kind | Document | Readings | Size | Why it is watched |
|---|---|---|---|---|
| Vendor governance pages | Responsibility & Safety — Google DeepMind https://deepmind.google/responsibility-and-safety/ | 17 | 6 KB | A vendor responsibility and safety hub — the page where documentation an assessment could not locate would appear if it were published |
| Vendor governance pages | GPT-5.6 System Card - OpenAI Deployment Safety Hub https://deploymentsafety.openai.com/gpt-5-6/evaluations-with-challenging-prompts | 17 | 128 KB | OpenAI Deployment Safety Hub — per-category evaluations and the named external evaluators |
| Vendor governance pages | Introducing Shieldstral. | Mistral https://mistral.ai/news/shieldstral/ | 17 | 7 KB | Shieldstral — the only substantial safety publication located for Mistral |
| Vendor governance pages | Anthropic’s Transparency Hub \ Anthropic https://www.anthropic.com/transparency | 13 | 117 KB | Anthropic transparency hub — the source for the RSP, system cards and external red-team results in the Claude assessment |
| Vendor governance pages | Anthropic’s Transparency Hub \ Anthropic https://www.anthropic.com/transparency | 4 | 132 KB | Anthropic transparency hub — the source for the RSP, system cards and external red-team results in the Claude assessment |
| Independent evaluations | Stanford CRFM https://crfm.stanford.edu/2024/11/08/helm-safety.html | 17 | 21 KB | Stanford CRFM HELM Safety — the source of the only independently published fairness evidence in the registry |
| Independent evaluations | Foundation Model Transparency Index https://crfm.stanford.edu/fmti/December-2025 | 17 | 11 KB | Stanford CRFM Foundation Model Transparency Index — the single source behind ten of the eighteen independent claims in this registry, which is a concentration worth watching |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 3 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 1 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Data on AI Models | Epoch AI https://epoch.ai/data/ai-models | 2 | 7 KB | Epoch AI model database — independent longitudinal data on model releases and compute, useful for corroborating vendor claims about lineage and scale |
| Independent evaluations | Updates - METR https://metr.org/blog/ | 17 | 6 KB | METR — named in the OpenAI system card for AI self-improvement evaluations; publishes methodology a third party could rerun |
| Independent evaluations | AILuminate - MLCommons https://mlcommons.org/benchmarks/ailuminate/ | 4 | 4 KB | MLCommons AILuminate — a consortium safety benchmark, and the closest thing to an industry-standard independent safety score |
| Independent evaluations | AILuminate - MLCommons https://mlcommons.org/benchmarks/ailuminate/ | 13 | 4 KB | MLCommons AILuminate — a consortium safety benchmark, and the closest thing to an industry-standard independent safety score |
| Independent evaluations | AISI Research & Publications | The AI Security Institute https://www.aisi.gov.uk/research | 17 | 8 KB | UK AI Safety Institute research — named as an external evaluator in the OpenAI system card, and one of the few public bodies running frontier evaluations |
| Independent evaluations | Science – Apollo Research https://www.apolloresearch.ai/science | 6 | 11 KB | Apollo Research — named in the OpenAI system card for sandbagging and scheming evaluations |
| Independent evaluations | Science – Apollo Research https://www.apolloresearch.ai/science | 10 | 10 KB | Apollo Research — named in the OpenAI system card for sandbagging and scheming evaluations |
| Independent evaluations | Science – Apollo Research https://www.apolloresearch.ai/science | 1 | 10 KB | Apollo Research — named in the OpenAI system card for sandbagging and scheming evaluations |
| Model cards | Gemma 3 model card | Google AI for Developers https://ai.google.dev/gemma/docs/core/model_card_3 | 17 | 17 KB | The Gemma 3 card, open, where the Hugging Face copy is gated and returns 401 |
| Model cards | https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama3_1/MODEL_CARD.md | 33 | 26 KB | Llama 3.1 card, open, where the Hugging Face copy is gated |
| Model cards | https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama3_2/MODEL_CARD.md | 33 | 25 KB | Llama 3.2 card, open, where the Hugging Face copy is gated |
| Model cards | https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama3_3/MODEL_CARD.md | 33 | 17 KB | Llama 3.3 card — most of the Llama family assessment rests on it |
| Regulation | AI RMF - AIRC https://airc.nist.gov/airmf-resources/airmf/ | 17 | 2 KB | NIST AI RMF — one of the four frameworks D7 scores mappability against |
| Regulation | Article 26: Obligations of Deployers of High-Risk AI Systems | EU Artificial Intelligence Act https://artificialintelligenceact.eu/article/26/ | 17 | 50 KB | EU AI Act Article 26 — deployer obligations, including the six-month log retention the router exists to satisfy. A change here changes the product |
| Regulation | Article 27: Fundamental Rights Impact Assessment for High-Risk AI Systems | EU Artificial Intelligence Act https://artificialintelligenceact.eu/article/27/ | 17 | 18 KB | EU AI Act Article 27 — the fundamental rights impact assessment, the largest product gap on the roadmap and a deliverable every Annex III deployer must produce |
| Regulation | Article 53: Obligations for Providers of General-Purpose AI Models | EU Artificial Intelligence Act https://artificialintelligenceact.eu/article/53/ | 17 | 6 KB | EU AI Act Article 53 — GPAI provider obligations, the upstream half of what a deployer can obtain |
| Regulation | Article 6: Classification Rules for High-Risk AI Systems | EU Artificial Intelligence Act https://artificialintelligenceact.eu/article/6/ | 17 | 22 KB | EU AI Act Article 6 — high-risk classification. It decides which systems the router's policies have to be strict about |
| Regulation | Article 72: Post-Market Monitoring by Providers and Post-Market Monitoring Plan for High-Risk AI Systems | EU Artificial Intelligence Act https://artificialintelligenceact.eu/article/72/ | 17 | 6 KB | EU AI Act Article 72 — post-market monitoring, the obligation the audit chain produces evidence for |
| Regulation | Signatory Taskforce of the General-Purpose AI Code of Practice | Shaping Europe’s digital future https://digital-strategy.ec.europa.eu/en/policies/signatory-taskforce-gpai-code-practice | 17 | 3 KB | GPAI Code of Practice signatory taskforce — signatory status cannot be confirmed from this page today. If it ever names companies, several claims change |
| Regulation | Regulation - 2023/1230 - EN - EUR-Lex https://eur-lex.europa.eu/eli/reg/2023/1230/oj/eng | 9 | 332 KB | EU Machinery Regulation 2023/1230, applying 20 January 2027. Annex I Part A item 5 pulls machine-learning safety components into mandatory third-party conformity assessment — the open question about whether this is a market |
| Regulation | The Fed - FRB: Supervisory Letter SR 26-2 on Revised Guidance on Model Risk Management -- April 17, 2026 https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm | 17 | 4 KB | SR 26-2 itself. Verified against this page on 2026-09-08; four marketplace artifacts and a framework row rest on it, so a revision or withdrawal changes the product |
| Safety frameworks | Introducing the Frontier Safety Framework — Google DeepMind https://deepmind.google/discover/blog/introducing-the-frontier-safety-framework/ | 17 | 7 KB | Google DeepMind Frontier Safety Framework — Critical Capability Levels and mitigations |
| Usage policies | Gemma Prohibited Use Policy | Google AI for Developers https://ai.google.dev/gemma/prohibited_use_policy | 17 | 4 KB | A published prohibited-use policy that an assessment scores against, so a change to it changes a claim |
| Usage policies | Llama 2 Community License Agreement - Meta AI https://ai.meta.com/llama/license/ | 17 | 7 KB | Llama licence — every Llama assessment binds out-of-scope use to it. A licence change is the most consequential quiet edit a vendor can make |
| Usage policies | Welcome to our Legal Center https://legal.mistral.ai/terms/get-started/ | 17 | 1 KB | Mistral terms — no prohibited-use statement for Mistral Large was located during assessment; this is where one would appear |
| Usage policies | Usage Policy \ Anthropic https://www.anthropic.com/legal/aup | 17 | 17 KB | Anthropic usage policy — the prohibited-use claim rests on it |
47 further documents are model cards recorded by the automated scan and are not listed individually here.
How a reading is recorded
- Content-addressed. A body is stored once under its own SHA-256, so a document unchanged across twenty readings is one file and one row with a counter. A changed document opens a second row and both bodies are kept.
- HTML is normalised before hashing, and the raw fingerprint is kept too. Navigation, scripts and build tokens move on every request, so hashing raw bytes would report a change every run and the signal would be noise. What is stored is the readable text — which is also what the assessments rest on — and raw_sha256 records the bytes as they arrived.
- Site chrome is not the document. Extraction prefers the page’s main content region and drops navigation, headers, footers and promo panels. Before it did, a “Commerce” item added to one vendor’s site navigation reported as a change to two of their governance documents on the corpus’s first day. Each snapshot records which extraction generation produced it, so a future reader can tell a vendor’s edit from a change in how we read.
- A document that cannot be read is recorded nowhere. A page that returns an error, or renders only on the client and extracts to almost nothing, is reported and skipped. An unreadable document is not a document that says nothing.