AIが行動する権利を、証拠によって獲得する場所。

Decision Lab protocol · definition only

From observation to operated-product evidence.

Planned · illustrative · no results collected

This protocol defines future records and measures; no run results have been collected for it. The physical Foundry remains planned and illustrative. Apply the protocol at the relevant Product System gate through all six lenses. Future observations belong in the existing Product System records and Evidence Passport, with configuration, exposure, authority, source, limitations, review and expiry. No stage establishes evidence maturity or authorises release automatically.

Five Product System gates and six lenses frame the decision. The five planned Foundry stations describe a facility concept. The six protocol stages below define what to record and measure.

1. 01

   Observation

   What is true now, how was it observed, and what reference establishes validity?

   Required record

   Timestamped sensor or data observation; origin and rights; configuration and envelope; evaluator or ground truth; missing, invalid and stale-state flags.

   Metric definitions · no measured values

   Valid observation rate

   Valid observations divided by required observation opportunities. Define validity before collection; retain missing observations in the opportunity count.

   Denominator: Required observation opportunities in the declared window.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Observation-to-state latency

   Distribution of elapsed time from a relevant physical event to a usable system state. Declare clock alignment and report censored or missing transitions.

   Denominator: Matched event-to-state pairs; disclose unmatched event count separately.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

2. 02

   Future

   Which future state is predicted, at what horizon, with what uncertainty and alternatives?

   Required record

   Source observation, forecast horizon, model and configuration, predicted state, uncertainty, alternatives and abstention. Pair predictions with later observations before assessing them.

   Metric definitions · no measured values

   Forecast error by horizon

   Apply a declared loss function between each prediction and its later reference observation. Report a distribution by horizon and slice; unobserved futures remain unassessed.

   Denominator: Matched prediction/reference pairs at each declared horizon; disclose unmatched predictions.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Forecast calibration

   Compare stated probabilities with observed frequencies within predeclared probability bins. Report bin sizes, uncertainty and excluded cases.

   Denominator: Resolved predictions within each declared probability bin and outcome definition.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

3. 03

   Action

   What was proposed, authorised, executed or rejected, and whose authority governed it?

   Required record

   Typed proposal, applicable constraints, human or deterministic authorisation, bounded command or abstention, execution acknowledgement and actual resulting state. Separate model proposal from control and specialist acceptance.

   Metric definitions · no measured values

   Accepted action transition rate

   Accepted resulting state transitions divided by authorised action attempts. Use predeclared acceptance criteria and retain failed or aborted attempts.

   Denominator: Authorised action attempts within the same configuration and authority boundary.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Authority boundary event rate

   Report blocked, escalated and unauthorised proposals separately by reason, divided by proposal opportunities. A blocked proposal is not inherently a system failure or success.

   Denominator: Proposal opportunities under the declared authority policy.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

4. 04

   Recovery

   After degradation or failure, was the condition contained and valid task state restored or the run stopped?

   Required record

   Trigger and failed transition, severity, retry/replan/rollback/handback/stop path, recovery owner, restored state, recurrence and outcome. Preserve failed recoveries and human interventions.

   Metric definitions · no measured values

   Accepted recovery rate

   Accepted recoveries divided by recovery attempts, using a declared restored-state criterion. Report degraded, failed, aborted and escalated outcomes separately.

   Denominator: All recovery attempts for the declared failure population and window.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Return-to-duty time

   Distribution of elapsed time from detection to accepted resumed useful work. Report unrecovered and censored episodes separately.

   Denominator: Recovery episodes; disclose which resumed duty and which remained stopped.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

5. 05

   Useful work

   Did the complete human–machine workflow deliver an accepted customer outcome, with what operator burden?

   Required record

   Customer acceptance definition and human baseline; attempted and accepted outcomes; exposure, scheduled and continuous duty; rejects, aborts, interventions, attention and recovery work.

   Metric definitions · no measured values

   Accepted useful outcome rate

   Accepted useful outcomes divided by attempted outcomes, under the declared acceptance definition. Report throughput and longest continuous useful duty alongside the ratio.

   Denominator: Attempted customer outcomes in the declared window; do not substitute easy subtasks.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Operator burden

   Operator attention, intervention and recovery minutes divided by duty hours. Distinguish elapsed supervision from person-minutes across multiple operators; compare with the human baseline.

   Denominator: Duty hours for the same system, workflow and observation window.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

6. 06

   Operated-product evidence

   Does this exact configuration sustain value, controls, service and economics for the declared population and window?

   Required record

   Evidence Passport and traceable Product System records: configuration, envelope, authority, population, window, exposure, evidence origin and maturity, source, limitations, review, expiry and human recommendation. Classify gaps; do not promote a planned protocol or simulation into field evidence.

   Metric definitions · no measured values

   Availability by cause

   Available duty time divided by scheduled duty time. Classify unavailable time by cause and disclose planned exclusions and overlapping causes.

   Denominator: Scheduled duty time for the same population and window.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Severity event rate

   Events divided by declared exposure, reported separately for each severity class. Preserve zero events, missing counts and absent exposure as distinct states.

   Denominator: Declared exposure such as attempts, operating hours or distance; specify the unit and never combine incompatible denominators.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

   Cost per successful outcome

   Total ownership and operating cost divided by accepted useful outcomes over the same window. Include integration, infrastructure, service, operator and recovery cost; disclose assumptions and allocation. An absent or zero denominator yields no ratio.

   Denominator: Accepted useful outcomes matching the cost population and accounting window.

   Window and breakdowns

   Declare the observation window, population, configuration, envelope and authority before collecting data. Report missing and excluded cases.

   configuration · operating-envelope slice · authority and supervision · site and task · failure or exclusion reason

These definitions support the existing Product Management metrics tree. Future observations must use the Product System and Evidence Passport. Version: decision-lab-protocol.v1.