Register / draft forecasts and commitments
Register of forecasts and commitments
The register has two kinds of record. A forecast states a proposition, the date by which it resolves and the source that will decide it. A commitment states something the lab has undertaken to do, the date by which it is due, and the check that shows whether it was done.
The register separates predictions of assessor behavior (what bodies that assess or regulate AI, including developers assessing their own systems, will decide or publish) from capability measurements. Neither is, by itself, a safety finding.
- P1
EU delegated act
The European Commission will not adopt, on or before 2 August 2027, a delegated act under Article 51(3) of Regulation (EU) 2024/1689 that either changes the numerical 10^25-FLOP threshold in Article 51(2) or supplements Article 51 with a benchmark or indicator for determining high-impact capabilities. Resolve by Commission adoption date, allowing 30 days for documentary publication; drafts, consultations, guidance, designations, implementing acts, and ordinary legislation do not count.
Candidate probability against a declared 0.50 ignorance baseline; no empirical reference class was supplied, so 0.50 stands as the fallback comparator until a checker accepts or replaces it.
Assessor behavior · Not registered
40%candidate p - P2
Public lab threshold attainment
By 2027-09-30, Anthropic will expressly conclude that the “Automated R&D in key domains” threshold in RSP v3.4 has been met, or Google DeepMind will expressly conclude that a named model reached ML R&D Automation Level 1 or Acceleration Level 1 under FSF v3.1. Only explicit public attainment counts; alert thresholds, precautionary safeguards, or later framework renaming do not.
Candidate probability against a declared 0.50 ignorance baseline; no empirical reference class was supplied, so 0.50 stands as the fallback comparator until a checker accepts or replaces it.
Assessor behavior · Not registered
22%candidate p - P3
Joint evaluation report
By 2027-09-30, NAAIMES will publish or formally co-issue a joint evaluation report identifying at least three member institutes, naming at least three model versions from at least two developers, reporting a common numeric model-level metric for every named model, and disclosing the task set, scoring rule, model/version, and principal inference or elicitation settings.
NAAIMES, as named by its source: “the International Network for Advanced AI Measurement, Evaluation and Science (NAAIMES) – formerly the International Network of AI Safety Institutes” (UK AI Security Institute, “International evaluation best practice and open questions in AI measurement”, retrieved ). This expansion is not part of the proposition.
The resolver for this forecast is not yet bound. The linked source page is the U.S. CAISI centre, not the network’s archive.
Candidate probability against a declared 0.50 ignorance baseline; no empirical reference class was supplied, so 0.50 stands as the fallback comparator until a checker accepts or replaces it.
Assessor behavior · Not registered
42%candidate p - P4
California statutory report
By 2027-01-31, Cal OES will file or publicly post its first report required by California Business and Professions Code §22757.13(g). A zero-incident report counts; a portal, press release, or legislative description without the statutory report does not.
The proposition’s date, 31 January 2027, is read as 23:59:59 Pacific time, which is 2027-02-01T07:59:59Z. The statute sets a start date (‘Beginning January 1, 2027, and annually thereafter’) and transmission to the Legislature and the Governor, not a public filing deadline. The 31 January window is this forecast’s own choice.
Candidate probability against a declared 0.50 ignorance baseline; no empirical reference class was supplied, so 0.50 stands as the fallback comparator until a checker accepts or replaces it.
Assessor behavior · Not registered
92%candidate p - P5
A 24-hour task horizon
By 2027-09-30, METR will publish a task-completion-time-horizon estimate for a named publicly available model with a P50 time horizon of at least 24 hours under TH1.1 or an explicitly bridged successor methodology. Internal or unreleased models and non-estimate lower bounds do not count.
Candidate probability against a declared 0.50 ignorance baseline; no empirical reference class was supplied, so 0.50 stands as the fallback comparator until a checker accepts or replaces it.
Capability measurement · Not registered
68%candidate p - P6
Final evaluation practices
By 2027-09-30, NIST or CAISI will publish a final, non-draft edition of NIST AI 800-2, Practices for Automated Benchmark Evaluations of Language Models, or an officially renumbered successor with substantially the same scope. Revised drafts and requests for comment do not count.
Candidate probability against a declared 0.50 ignorance baseline; no empirical reference class was supplied, so 0.50 stands as the fallback comparator until a checker accepts or replaces it.
Assessor behavior · Not registered
76%candidate p
Commitments
Each commitment names the part of the lab’s process that owes it and a check that anyone can run from public pages. Its outcome is published later as a separate attestation record that cites it, because published records are never edited. A cancelled commitment is recorded as cancelled and does not count as kept.
- com-2026-0001
Return of the withdrawn 10 May 2026 lab journal entry
On 12 May 2026 the lab marked its 10 May lab journal entry as "temporarily withdrawn from public claim status". Its withdrawal notice states: "A revised public version may be republished after legal/accounting facts, grant-submission facts, money-rail claims, and privacy boundaries are reviewed against dated evidence." No revised version has been published as of 26 September 2026.
Deadline · obliged: E3 Ledger
cancelled - com-2026-0002
Return of the withdrawn Entry 006 (10 to 11 May 2026)
On 12 May 2026 the lab marked Entry 006 as "temporarily withdrawn from public claim status". Its withdrawal notice states: "A repaired journal version can return after the archive hardening pass verifies that its claims are scoped to audit-supported evidence and marked as hypothesis, operator narrative, runtime observation, or public evidence as appropriate." No repaired version has been published as of 26 September 2026.
Deadline · obliged: E3 Ledger
kept - com-2026-0004
Identity of the lab's public properties
Every public property in scope states the identity of Hapax Research Lab and links hapaxresearch.com by 1 October 2026. The properties are hapaxresearch.com, the hapax-systems GitHub organization, the hapax omg.lol profile and weblog, and the Bluesky account hapaxresearch.com.
Deadline · obliged: E1 Identity
open (last check failed) - com-2026-0005
Resolution of the two withdrawn May 2026 entries
The lab resolves both withdrawn May 2026 entries by 9 October 2026, by republishing each or by publishing a cancellation for it. A republished entry counts as kept. A cancelled entry is recorded as cancelled and does not count as kept.
Deadline · obliged: E3 Ledger
open (last check failed) - com-2026-0006
Seven consecutive notebook entries
The lab publishes one notebook entry on each of seven consecutive days by 7 October 2026, and each entry is read back from its public address after publication.
Deadline · obliged: E4 Notebook
open (last check failed) - com-2026-0007
Response dispositions within 48 hours
Every response addressed to a notebook entry during the first seven days receives a recorded disposition within 48 hours. The dispositions are answered, folded into a correction, folded into this register, noted, or declined. The register publishes counts only and keeps no list of who responded, so anyone can recount the responses from the public Bluesky service and the weblog at the time of a check, but not later. Because no response is linked to its disposition and the count of responses received is the lab's own observation, the coverage this commitment states is the lab's attested total matched against a recount at check time, not an external response-by-response audit.
Deadline · obliged: E5 Capture
open (last check undetermined) - com-2026-0008
Drafting and review families on each notebook entry
From its first notebook entry onward, each notebook entry carries a line that names the model family that drafted it and the model family that reviewed it, and the two families differ. The line is the lab's own statement and is not by itself evidence that a review took place. The first review of this commitment covers entries published from 26 September to 7 October 2026, and each later review covers the following three months.
Deadline · obliged: E4 Notebook
open (last check undetermined) - com-2026-0009
Corrections keep the corrected text readable
From 26 September 2026 each correction record in this register quotes the corrected text in full, and the corrected entry stays readable at its address with a notice that links the correction. The first review of this commitment is on 7 October 2026, and each later review covers the following three months.
Deadline · obliged: E3 Ledger
fulfilled at last check
Read the record directly.
The human view and machine exports come from the same Markdown records. Candidate probabilities are marked as drafts in both. A content digest identifies this snapshot; it does not prove a proposition true.