This console runs the canonical chain end to end. You seal a forecast before execution, run the mission, and watch the outcome either pair into calibration or get refused as hindsight. Only paired observations calibrate. Nothing here learns silently: every behavior change passes a promotion court and a governance decision, in the open, with receipts.
One mission, full chain: neuruh.outcome-record.v1 → Outcome Calibration Ledger → Learning Update Proposal → Promotion Gate → Governance → promoted behavior. Seal the forecast first. The hindsight button is there so you can watch the starvation defect fire.
Demo court rule (thresholds chosen for this console; the spec itself requires paired pre-outcome forecasts, an independent court, and a governance decision, and leaves the numbers to the implementation): a proposal passes with at least 3 calibration pairs behind it, mean Brier at or under 0.25, and no governance-boundary item inside it. Failing proposals return as evidence, not as silent edits.
Each bar pair is one sealed forecast matched to its observed outcome. The gap between the two bars is the calibration error the ledger records. Hindsight outcomes never appear here: there is nothing truthful to plot.
| Measure | Live value from this session |
|---|---|
| Recipe calibration | awaiting pairs |
| Confidence calibration (stated vs observed) | awaiting pairs |
| False positive rate (predicted success, observed failure) | awaiting pairs |
| False negative rate (predicted failure, observed success) | awaiting pairs |
| Mean time-to-outcome | awaiting pairs |
| Mean cost vs value | awaiting pairs |
| Failure concentration (top archetype) | awaiting pairs |
The Sep-5 court mapped roughly 14 outcome producers and required no new outcome schema. These adapters are the bridges. Toggle A3 off and try to seal a forecast: the starvation defect is what a missing bridge feels like from inside the loop.
Does changing the computational organization around the same mission change measured results? Run the comparison protocol: same mission class, five organizations, repeated trials. One successful run proves nothing; this bench refuses to crown a winner under 10 trials per variant.
| Variant | Quality | Calib. error | Completion | Interventions | Cost / run | Latency | Policy fails | Econ. value |
|---|---|---|---|---|---|---|---|---|
| No trials run yet. Post-outcome record per trial captures actual outcome, time-to-outcome, value, cost, errors, reversals, human interventions, FP/FN class, causal caveats, acceptance, and a promote / demote / retest recommendation. | ||||||||
Automatic learning may propose the left column on its own. The right column stays a versioned proposal with tests and an independent court. Try forcing one:
OutcomeRecord teaches whether executed work was correct. Demand telemetry teaches what buyers repeatedly ask for that the organism cannot yet serve. Submit demand signals; unmatched demand becomes a MISSING_INGREDIENT, gets ranked by RICHIE, and repeated pull commissions a build instead of another idea backlog.
No raw sensitive prospect text is retained for learning: problem class, requested outcome, buyer class, match state, missing ingredient, band, conversion state, and reason lost. That is the whole record.
Existing offer registry: Property Opportunity Scan Property Intelligence Dossier
No unserved demand yet. Submit a signal the registry cannot match.
Searches, questions, likes and "that would be cool" are weak demand. Productization rises with this ladder, not with applause.
Repeated successful reasoning compiles downward into cheaper logic. Run the same classification task repeatedly: each clean run is evidence. At the thresholds, promote the pattern down a rung. A novel or ambiguous case routes straight back to frontier reasoning. Ambiguity is never compiled away.
Objective: lowest reliable cost for the required evidence quality and consequence level, not minimum token spend. Independent court and outcome calibration stay necessary at every rung; cost optimization never outranks truth.
Six moves, about four minutes. Every claim below is something the console actually does on screen.