Evidence Press research metrics

Version: 1.0 Effective: 29 August 2026 Status: prospective measured-release supplement

Evidence Press will publish a scoped forecast and outcome receipt with every new release. The purpose is to make tractability estimates calibratable, expose the cost of failed routes, and distinguish active work from waiting. It is not a productivity leaderboard and does not turn a release into a validated theorem.

The observation unit

The unit is a registered attempt, not a release. Every in-scope attempt is entered in the public work ledger before substantive work begins and remains there if it is stopped, abandoned, superseded, null, negative or never released. Release-only statistics omit failures and therefore exaggerate tractability.

When a release packages research that predates registration, the receipt is scoped to publication or assurance only. The earlier discovery clock remains left-censored and is not reconstructed from commits, chat records or release dates.

The enforced boundary begins at 2026-08-29T11:12:44Z, the first intake with the complete forecast and telemetry contract. Earlier work-ledger attempts are retained as denominator records but are not retrospectively enriched.

What is frozen at intake

Each attempt records:

• the exact target outcome and measurement scope;
• a procedure class and reference class;
• a low, central and high active-minute forecast;
• the Fermi components whose products sum to those three forecasts;
• the expected unattended wait;
• separate probabilities of a positive signal and full target closure within a stated active-time horizon;
• assumptions, reforecast triggers and a stop rule.

The forecast timestamp must equal the attempt registration timestamp. An after-the-fact estimate is an outcome narrative, not a forecast.

What is recorded at a terminal state

Every published, stopped, abandoned or superseded attempt records:

• calendar elapsed, active agent, unattended wait, blocked and rework minutes;
• active human and substantive compute minutes when they were actually captured;
• agent runs, maximum parallelism and model turns;
• fork-aware task-local model tokens when the runtime exposes them;
• positive, negative and inconclusive research cycles;
• falsification gates, materially distinct candidate architectures and rejected architectures;
• substantive review rounds, P0/P1 findings and pre-publication claim corrections;
• the scoped result state, whether a positive signal occurred and whether the exact frozen target was reached;
• forecast error, actual-to-forecast ratio, interval coverage and the reason for a material variance.

Release representations also derive Brier scores for the frozen positive-signal and target-closure probabilities when the corresponding terminal outcome is adjudicated. The score is computed from the source receipt rather than entered again. A single score is not calibration; a reference class of comparable prospective attempts is.

Optional telemetry may be unknown only with a field-specific reason. Core clocks, cycle counts and calibration fields are mandatory.

How to read the metrics

Discovery, assurance, publication and translation are separate clocks. Compute may overlap active agent work, human work may overlap waiting, and elapsed time is not their sum. A fast result at an early assurance endpoint is not a matched comparison with conventional peer-reviewed research.

The records are suitable for prospective reference classes and probability calibration. A comparative acceleration or productivity claim still requires a registered comparator, matched scope and assurance, attempted-work denominator, error and correction outcomes, and independent assessment.

Machine-readable sources:

• research-metrics-policy.json defines the prospective rule and terms;
• work-ledger.json contains all attempts and their metric receipts;
• each release's paper.json exposes the linked metric records.