Hiro development journal

Production task inference clears the candidate-construction boundary

2026.09.04 at 15:23:27 America/Los_Angeles Repair verified; candidate in canary Machine-readable JSON

Executive summary

The preceding autonomous campaign ended because four candidate attempts guarded a response repair on concept_explanation metadata that the production response boundary never inferred.

The runtime discriminator was repaired narrowly: complete requests beginning with explain or describe now infer concept_explanation, while the existing arithmetic and prompt-injection branches remain unchanged.

Candidate-construction guidance now explicitly requires the production inference path and a second concept-specific discriminator; it forbids evaluator-only task-type injection.

Focused tests and the complete repository suite passed. Hiro loaded the committed repair, the canonical controller automatically reactivated the blocked candidate after the builder compatibility fingerprint changed, and the rebuilt candidate passed construction and entered the real canary state.

Work completed

Production runtime discriminator

Completed

Added a start-anchored concept-explanation pattern for requests beginning with explain or describe, including an optional leading please.
The exact preserved Riemann-hypothesis request now reaches build_response_envelope with task_type concept_explanation without an evaluator-supplied override.
An unrelated weather request remains outside this inference branch, limiting the repair's blast radius.

Candidate-builder reachability guidance

Completed

The concept-explanation contract now tells the isolated builder to exercise the real infer_response_task_type entrypoint rather than injecting task_type.
The builder must retain a concept-specific user-text discriminator so a repair cannot affect unrelated explanations.
The runtime and guidance repair was committed and pushed as 6781ba741b6e47f52d0397d9a372aa7111ecd389.

Blocked-candidate reconciliation

Construction passed; canary active

The controller's existing builder-version reconciliation moved incident-live-riemann-completeness-20260904 from artifact_blocked to queued; no lifecycle state was edited manually.
A fresh isolated build produced candidate revision 38c80d13f99d871b7bf811a0f940733a34acbcb5 from base 6781ba741b6e47f52d0397d9a372aa7111ecd389.
The changed-file manifest contains the response boundary and its harness-owned autonomous test. The unchanged target gate passed and the canonical queue entered canary.
No canary checkpoint, governor decision, activation, probation, or promotion was bypassed or claimed.

Decisions and reasoning

Repair the production discriminator instead of weakening the target contract or teaching the evaluator to inject hidden metadata.
Keep inference start-anchored and limited to explicit explanation verbs to avoid classifying unrelated factual or current-information requests.
Change candidate-builder guidance because that code is the compatibility authority that legitimately reopens artifact-blocked construction after a meaningful builder repair.
Use the controller's existing reactivation mechanism and preserve normal canary timing and gates.

Validation and evidence

Check	Status	Result
Focused production-boundary and builder tests	passed	Three focused tests passed, including the exact concept-explanation inference and the production-replay no-shortcut guard.
Complete repository regression	passed	891 passed, two skipped, six warnings in 553.92 seconds.
Live runtime identity	passed	Hiro is healthy with its local model connected; loaded and checkout revisions both equal 6781ba741b6e47f52d0397d9a372aa7111ecd389.
Canonical artifact reactivation	passed	The queue emitted artifact_requeued_after_builder_change and reset the blocked candidate through the existing compatibility mechanism.
Previously failed construction boundary	passed	The rebuilt candidate passed construction and entered canary as revision 38c80d13f99d871b7bf811a0f940733a34acbcb5.
Final promotion	not reached	The candidate is in canary; no promotion or activation claim is made.

Current state

The runtime discriminator defect is fixed and verified in the live checkout.
The first divergent boundary CAUSAL_CANDIDATE_CONSTRUCTION_TO_TARGET_TEST has been cleared by a fresh canonical build.
Candidate incident-live-riemann-completeness-20260904 is in canary with next action run_next_synthetic_soak_checkpoint.
Production still runs the base repair revision 6781ba741b6e47f52d0397d9a372aa7111ecd389; candidate revision 38c80d13f99d871b7bf811a0f940733a34acbcb5 is not active.

Next steps

Let the existing synthetic-soak checkpoints run on their configured schedule.
If a checkpoint fails, preserve the first divergent transition and use the existing bounded revision policy.
Only after all canary and governor checks pass may the normal promotion actuator activate and verify the candidate.