The promise book. Frontier AI labs publish commitments with dates on them; this book seals those commitments as forecast rows before their outcomes, under the same gate, the same controls and the same adjudication as every other row on the desk. Stated versus operational, applied to the people building the measurement stack. AI-claims book (KK17 #4), reopened by operator ruling 2026-08-27 (desk-local 2026-08-26).
Each row quotes the promise as the lab stated it and names the page; the operator prices it and a climatological control seals in the same run. Rows resolve on the register's frozen sources, not on press.
| id | statement | p | deadline | status | control |
|---|---|---|---|---|---|
KKR-20260827-113 | Anthropic records the provable inference prototype goal of its Frontier Safety Roadmap, stated as a prototype to be developed by 2026-09-30, as completed with a completion date on or before 2026-09-30, on the Frontier Safety Roadmap updates page at anthropic.com. | 70% | 2026-10-14 | open | KKR-20260827-116 |
KKR-20260827-114 | Anthropic records the Leveling up across the board security goal of its Frontier Safety Roadmap, stated with a target of 2027-07-01, as completed with a completion date on or before 2027-07-01, on the Frontier Safety Roadmap updates page at anthropic.com. | 50% | 2027-07-15 | open | KKR-20260827-117 |
KKR-20260827-115 | OpenAI publishes a revised Preparedness Framework, described by OpenAI as a revision or new version of the framework that dates largely to December 2023, on openai.com between 2026-08-28 and 2026-12-31. | 35% | 2027-01-07 | open | KKR-20260827-118 |
KKR-20260827-121 | OpenAI publishes a system card or preparedness report for the Astra model on openai.com between 2026-08-28 and 2026-12-31 that names at least one external organization, a government agency or an independent AI safety organization, as having evaluated the Astra model. | 25% | 2027-01-07 | open | KKR-20260827-122 |
Promises that resolved before the book could seal them. Logged with outcomes, excluded from every score: a row sealed after its outcome would be retrodiction, and the book keeps the same law as the ledger.
| promise | outcome | why not sealed |
|---|---|---|
| publish a technical report on its Hugging Face investigation 'in the coming days' (Aug 18, 2026) | KEPT - 37-page technical report and blog published 2026-08-26 (TechCrunch, CNBC, Axios, Fortune); 8 days | resolved before the book opened; already-decided at seal |
| METR and Redwood Research 'planning to publish their own reports on the incident' (TechCrunch, Aug 26) | KEPT - 91-page independent analysis published 2026-08-26 (Fortune); same day | resolved same day |
| lightly redacted transcript 'within the next week' (Anthropic, 2026-07-30) | TO OBSERVE - promise verified at the primary (anthropic.com, 30 Jul: 'within the next week, we will release a lightly redacted transcript'); no release located by desk search on 2026-08-27. The earlier 'none by 2026-08-05 (The Hacker News)' line rested on an unlinked snippet and is retained as superseded (5.07). | promise window lapsed before the book opened |
Every promise resolves against the page as it stood when the book opened. Freeze discipline at book opening: web.archive.org snapshot where the host permits; local SHA-256 held outside the repo where it does not (packet pattern: local by design, hash in the register); secondary reporting named where the primary URL was not identified. Gaps printed, not filled.
| key | page | frozen |
|---|---|---|
ANTH-ROADMAP | https://www.anthropic.com/responsible-scaling-policy/roadmap | snapshot 2026-08-27 17:04Z |
ANTH-ROADMAP-UPDATES | https://www.anthropic.com/responsible-scaling-policy/updates | snapshot 2026-08-27 17:04Z |
OAI-JUL21-DISCLOSURE | https://openai.com/index/hugging-face-model-evaluation-security-incident/ | snapshot 2026-08-27 17:07Z |
OAI-AUG18-AS-REPORTED | OpenAI primary URL for the 2026-08-18 safety-practice announcement not identified at seal | as reported 2026-08-18 |
OAI-AUG26-TECHNICAL-REPORT | https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf | 2026-08-27 local hash |
— Rows are forecast-shaped and gated like every other row; domain ai-claims; priced by operator/human and paired with control/baserate. Frontier arms do not price rows about their own developer; cross-lab pricing by frontier arms is deferred to v2 (assistant ruling at delegation, disclosed).
— Every row's statement quotes the promise as the lab stated it and names the page; this register carries the source URL and a web.archive.org snapshot taken at seal, so the promise text cannot drift.
— Two layers per promise where both exist: STATED (the lab's own status page records the goal met by its target) and OPERATIONAL (an external artifact exists). v1 seeds the STATED layer; OPERATIONAL rows follow when a named venue exists.
— Promises that resolve before the desk can seal them are logged here as OBSERVED (kept / late / broken, with days), never sealed as forecasts; they feed the control arm's base rate for 'lab keeps a dated promise'.
— Drafting disclosure: the seed rows were drafted by an Anthropic model (Claude) from the labs' public pages, verbatim; pricing is the operator's alone.
Register ai_claims_register_2026-08-27.json sha256 aec3a19b8b1cfd52; ledger as of 2026-09-02T00:09:26Z; page generated 2026-09-02 00:14Z. Everything above recomputes from the register and the ledger in the repository.