Domains

registry v1 · schema v0.7 · 3 registered

A domain is a subject area the log records. The mechanism is the same in every one: a claim, a citation, a snapshot hash, and the signatures of operators no interested party controls. What a domain owns is the tables: which categories exist, how long each stays fresh, which carry a transcript, who is too close to judge, and how a subject is named. Every entry names its domain in its signed core, so a fact can never be moved from one domain to another, by anyone. The same window applies in every domain too: an entry's proof is public from the first minute, and its content is released 30 days after the seal that covers it (decision D-100).

01 · ai-ecosystemThe AI ecosystem What models cost and do. registered 02 · ai-governanceAI governance What states and intergovernmental bodies require. registered 03 · ai-safetyAI safety What non-state parties committed to, and what their systems do. registered

01The AI ecosystem ai-ecosystem registered

What it records: What models and AI products cost and do. A release, a deprecation, a price, a limit, an outage, or a measured behavior of a model, stated by the provider's own page or observed under a receipt and a frozen transcript.
Use case: An agent that calls other models needs the price, the limit, and the deprecation state that hold today, without trusting one vendor's changelog and without re-reading the web. A continual learner syncs the delta, weights a measured behavior above a quoted one, and unlearns a fact when a dispute overturns it.
Who reads it: Agent frameworks and model gateways that route across providers; evaluation and observability tooling; teams that keep a model's knowledge of the ecosystem current; anyone who needs a citable, dated answer to what a provider said and whether it held.
Who validates it: Operators outside every model provider, attested in this domain. The genesis call names operators for this domain.
Subject: <provider>/<model or product>, as openai/gpt-5.
Categories: releasedeprecationpricinglimitbehavioroutagemisbehaviorcorrection
Transcripts: behavior, misbehavior — every other category carries its measurement in the observation.
Freshness: pricing 90 days · limit 90 days · behavior 30 days. The rest never go stale: release, deprecation, outage, misbehavior, correction.
Excluded parties: 22 parties: openai.com, anthropic.com, google.com, deepmind.google, meta.com, microsoft.com, x.ai, and the rest on the policy page.
Attestation: “No model provider holds control of, or a beneficial stake in, this operator.” nomankind-independence-v1
Sources: Official source required: pricinglimitdeprecationreleaseoutage Every other category may cite any host, and the citation is labeled official, recognized, or other.
In this log: 0 entries · 0 trusted operators

Browse entries Policy tables Join this domain

GET /read?domain=ai-ecosystem&subject=openai/gpt-5&category=release

02AI governance ai-governance registered

What it records: Instruments issued by states and intergovernmental bodies, binding or soft law: the EU AI Act, the Council of Europe Framework Convention on AI, state laws, executive orders, ISO 42001, the NIST AI Risk Management Framework, the OECD Principles, UNESCO's Recommendation. What is in force, what was amended or repealed, what guidance was issued, what was enforced.
Use case: A system that must act under the rules of a jurisdiction needs to know which instrument is in force today and what changed since it last looked, cited to the issuing body's own gazette rather than to commentary about it. The record answers what the law says with the text the state published, dated and sealed.
Who reads it: Compliance and legal tooling; policy researchers and trackers; agents deployed across jurisdictions; the bodies that issue the instruments, reading how their own text is being cited.
Who validates it: Operators attested in this domain. The issuing body of an instrument may not validate an entry about it, and model providers are excluded here as everywhere.
Subject: <jurisdiction or body>/<instrument slug>, as eu/ai-act.
Categories: in_forceamendedrepealedguidance_issuedenforcement_actioncorrection
Transcripts: none — no category carries a transcript; every entry is confirmed against the cited instrument.
Freshness: in_force 365 days · guidance_issued 365 days. The rest never go stale: amended, repealed, enforcement_action, correction.
Excluded parties: 22 parties: openai.com, anthropic.com, google.com, deepmind.google, meta.com, microsoft.com, x.ai, and the rest on the policy page.
Attestation: “No model provider, and no body that issues an instrument this record checks, holds control of, or a beneficial stake in, this operator.” nomankind-independence-v1
Sources: Official source required: in_forceamendedrepealedguidance_issued Every other category may cite any host, and the citation is labeled official, recognized, or other.
In this log: 0 entries · 0 trusted operators

Browse entries Policy tables Join this domain

GET /read?domain=ai-governance&subject=eu/ai-act&category=in_force

03AI safety ai-safety registered

What it records: What non-state parties committed to about harm to people, and what their deployed systems and guardrail products actually do. Lab constitutions, scaling and preparedness policies, usage policies, industry safety commitments, civil-society principles, guardrail products; and the measured conduct beside them: refusals, filters, safety evaluations, incidents.
Use case: Hold a commitment against conduct. A published policy is a stated fact; a refusal or a filter behavior is an observed one with a transcript; both sit on the same subject, so the gap between what a party said and what its system did is a query over the log rather than an argument.
Who reads it: Safety researchers and red teams; civil-society monitors; buyers comparing guardrail products; journalists; the parties themselves, whose commitments are on the record they cannot edit.
Who validates it: Operators attested in this domain. Model providers, guardrail vendors, and any party they fund are excluded from validating here.
Subject: <party>/<document or model>, or <party>/<model>/<version>, as anthropic/usage-policy.
Categories: commitment_publishedcommitment_changedcommitment_withdrawnconduct_observedrefusal_behaviorfilter_behaviorsafety_evalincidentcorrection
Transcripts: conduct_observed, refusal_behavior, filter_behavior — a transcript holding a working jailbreak is archived with its payload redacted and published after the disclosure window, its hash intact.
Freshness: conduct_observed 30 days · refusal_behavior 30 days · filter_behavior 30 days · safety_eval 90 days. The rest never go stale: commitment_published, commitment_changed, commitment_withdrawn, incident, correction.
Excluded parties: 30 parties: openai.com, anthropic.com, google.com, deepmind.google, meta.com, microsoft.com, x.ai, and the rest on the policy page.
Attestation: “No model provider, no guardrail vendor, and no party funded by one, holds control of, or a beneficial stake in, this operator.” nomankind-independence-v1
Sources: Official source required: commitment_publishedcommitment_changedcommitment_withdrawn Every other category may cite any host, and the citation is labeled official, recognized, or other.
Delayed disclosure: conduct_observedrefusal_behaviorfilter_behavior A transcript in one of those may be submitted with its payload replaced by a hash. The payload itself is archived at submission and published 90 days after it, so the evidence becomes public on a stated clock and the transcript's own hash never changes.
Staleness on a version change: conduct_observedrefusal_behaviorfilter_behaviorsafety_eval A subject in one of those names the version it was observed against, and the entry goes stale the moment an entry about another version of the same model verifies: what was measured was measured on a version that is gone, and a reconfirmation cannot bring it back.
In this log: 0 entries · 0 trusted operators

Browse entries Policy tables Join this domain

GET /read?domain=ai-safety&subject=anthropic/usage-policy&category=commitment_published

Adding a domain

A new domain is a published decision, not a code change: its tables are written into the registry first, its slug is added to the schema, and the policy module is extended to match, pinned by a test so the two never drift. Existing entries are untouched; their domain is in their signed core. Operators join a later domain by signing that domain's attestation, a public event in the log.