Observable event and deadline
Exact binary question, scope, population or AI system, deadline in UTC and clear yes/no outcomes. Language must be set before locking; open-ended wording stays a draft.
CORTIST / PUBLIC STANDARD / V1.0Forecasts need
Forecasts need
rules before results.
The Cortist Forecast Protocol defines how we publish, preserve, lock, verify and score AI predictions. It is a public operating standard, not a claim of forecasting accuracy.
FORECAST PROTOCOL / PROSPECTIVE RULES · The protocol defines locking, resolution and scoring rules. Two separate milestone forecasts (CM-003 and CM-004) now have editorial publication locks; none of the eight numeric topic drafts is locked or scored. See results or editorial standards. Editorial responsibility ↗
THE STANDARD / 22 OCT 2026
Forecast Protocol v1.0.
Stable questions. Visible changes. Independent evidence. Transparent errors.
Numerical drafts8
F-001–F-008; not scored
Publisher-locked2
CM-003 and CM-004; not externally reviewed
Resolved and scored0
No score or accuracy claim is possible
01 / THE FORECAST CONTRACT
What every record must include.
A reproducible commitment
Probability from 0 to 1, forecast owner, publication time, rationale, method and supporting sources. Year-slider interpolation is an illustration, not a separately locked probability for every year.
Preregistered tests
Named domains, thresholds, comparison groups, permitted sources, replication requirements and rules for conflicting or missing evidence. A company's own AGI/ASI announcement alone is not sufficient.
Independence and disputes
Identify reviewers, disclose conflicts, publish a written decision and supporting evidence. When evidence cannot determine the outcome, leave the question unresolved rather than calling it false.
02 / VERSIONING
No quiet rewriting.
- Draft. A record can display a provisional estimate and open methodological questions. Each published snapshot gets its own path (for example
/data/forecasts/F-001/v0.1.json). It is not scored. - Review and lock. Resolve all material open questions, preregister the evaluation, identify the accountable editorial approver (an outside reviewer is optional), publish an exact UTC lock time and a snapshot hash. A versioned original with a transparent prospective publication timestamp and content digest is required to call a record genuinely locked.
- Update. Create another version at a new URL, preserving every old snapshot. Explain what changed and why. Subsequent probability updates must be distinguished from the original locked commitment and not silently substituted in its score.
- Resolve. Publish a yes/no evidence decision or mark unresolved/disputed/void with justification. A post-hoc rule change cannot convert an old draft into a scored forecast.
- Report. Publish both successes and failures, sample sizes, unresolved cases and baselines. Independent adjudication evidence links must remain inspectable.
03 / SCORING
Errors become public.
The Brier score applies to resolved, prospectively locked binary forecasts only.
Brier score = (probability - outcome)2
Outcome is 1 if an event happens under the locked rules, otherwise 0. Lower scores are better.
| Forecast | Observed | Brier score |
|---|---|---|
| 80% yes | Yes | 0.04 |
| 50% yes | Yes | 0.25 |
| 10% yes | Yes | 0.81 |
These are hypothetical arithmetic examples, not Cortist forecast results. We will report score count, representative baselines and data exclusions. Calibration comparisons require enough independently resolved questions.
04 / TRANSPARENCY
Machine-readable by default.
Public forecast manifest (JSON) ↗Record schema (JSON) ↗Adjudication register (JSON) ↗Reference scoring code ↗
Record hashes are SHA-256 digests of the exact bytes of each versioned JSON snapshot. They detect discrepancies against the published manifest. A manifest controlled by the same site is not an independent timestamp, digital signature, tamper-proof ledger or legally immutable archive. R2 retention is now configured and readback-verified. Independent external anchoring and expert signoff remain useful optional assurance measures, not requirements for an editorial publication lock under the updated 27 October 2026 policy. A site-controlled digest alone does not prove independent time of publication.