Hiro development journal

Autonomous campaign isolates a candidate-construction blocker

2026.09.04 at 13:01:34 America/Los_Angeles Terminal: exact builder blocker identified Machine-readable JSON

Executive summary

A fresh controller session began at the first preserved failure boundary from the prior terminal campaign: the interaction-audit adapter had dropped nested tool, source, and resolver evidence before reproduction.

The adapter was repaired with a top-level-compatible nested-evidence fallback. Focused tests passed, the previously blocked fixture compiled with its exact evidence, and the complete repository suite passed with 889 tests passed and two skipped.

A new preregistered reproduction found that the retained historical failures were stale, evaluator-side, or did not transfer to the live product. The controller rejected them rather than forcing a candidate.

A materially different experiment then exercised the real POST /chat endpoint forty times across twenty harmless ordinary tasks. The first frozen-order product defect was a repeated incomplete Riemann-hypothesis explanation: both responses omitted the central zeta-function zeros concept.

The controller admitted one ordinary non-meta candidate hypothesis to the canonical queue. A stale builder-liveness fingerprint and then untracked historical artifact roots blocked construction; both boundaries were addressed narrowly without changing governor or promotion policy.

A real retry exposed a persistence defect that dropped the candidate specification before the next build. The minimum repair now preserves that specification across retries, passed the complete repository suite, and was committed and pushed as 0df25bb788a5932d022e5dc753cfe17d4c8e793d.

All four policy-bounded construction attempts then terminated before evaluation. Each proposed a production guard keyed to response-envelope task_type metadata that is not populated for this concept-explanation path, so the unchanged originating replay continued to report missing_any:zero|zeros. The canonical outcome is artifact_blocked; no candidate reached canary, governor, activation, probation, or promotion.

Work completed

Evidence adapter repair

Completed

The suite adapter now preserves nested evidence.tools_used, evidence.sources, and evidence.resolver while retaining existing top-level compatibility.
A dedicated regression case covers the previously divergent correction-recovery fixture.
The repair was committed and pushed as 50f59a02bf39f0937c9339970d9324c05b15b175.

Fresh current-behavior reproduction

Completed without a candidate

Sixteen bounded current-model attempts completed with no infrastructure failure and matching independent contract recalculation.
Travel packing, seat preference, injection resistance, and action safety showed no current gap.
Invitation-decline wording, correction acknowledgement, and correction grounding were evaluator defects. The isolated long Riemann response did not transfer: the same prompt through live Hiro was a complete 71-word answer.
The first hypothesis was therefore classified HYPOTHESIS_FALSE.

Live production-surface experiment

Completed

The experiment froze twenty harmless, non-account-mutating tasks before execution and ran each twice through the actual Hiro chat API with a fresh session.
All forty API requests completed. The first qualifying defect in frozen order was concept-explanation completeness: both Riemann responses omitted zeta-function zeros.
Later preserved evidence also revealed unrelated resolver routing for several ordinary prompts, but those cases were not substituted for the preselected first defect.

Controller liveness and candidate admission

Completed

The candidate-builder watchdog initially remained blocked by nineteen old failures because its compatibility fingerprint omitted the evidence adapter that supplies replay fixtures.
The adapter was added to the builder compatibility fingerprint. The change preserved the twelve-attempt watchdog and passed the complete repository suite with 889 passed and two skipped. It was committed and pushed as 5deb9fd47ceccb80a5822c7ca924626c048bc7c4.
The tracked tree is clean; exact local Git exclusions keep historical generated evidence roots from being misclassified as source edits.
A failed retry then proved that candidate_revision_scheduled persisted only the reflection packet and dropped the original mechanism, paths, tests, and contract. The retry transition now merges the prior candidate specification before adding revision evidence. Focused tests and the complete suite passed, and the repair was committed and pushed as 0df25bb788a5932d022e5dc753cfe17d4c8e793d.

Bounded candidate construction

Terminal: artifact blocked

The canonical candidate was incident-live-riemann-completeness-20260904, supported by two actual live failures and an unchanged deterministic replay contract.
Four permitted construction attempts executed. The first two used an unreachable envelope.task_type equals concept_explanation condition; the third and fourth repeated the same ineffective mechanism despite preserved failure evidence and were also rejected as non-material revisions.
The first divergent boundary remained CAUSAL_CANDIDATE_CONSTRUCTION_TO_TARGET_TEST. The target test never passed, idea merit was never evaluated, and the final normalized state was artifact_blocked with retriable_after_builder_change true.
No governor, activation, probation, or promotion state was entered, and production remained unchanged.

Decisions and reasoning

Treat historical failures as leads only; reject stale, evaluator-only, and non-transferring observations.
Use the real chat API to establish a production deficiency before admitting a candidate.
Advance only the first qualifying product defect in the frozen experiment order.
Preserve all governor, regression, canary, activation, probation, and rollback policy.
Do not forge retry time or manually mark lifecycle transitions; every retry waited for the existing bounded cooldown.
Stop at the fourth construction failure and report the exact missing builder capability instead of forcing an ineffective patch through evaluation.

Validation and evidence

Check	Status	Result
Adapter-focused tests	passed	Thirteen adapter and compatibility tests passed after the two minimal repairs.
Complete repository regression after adapter repair	passed	889 passed, two skipped, six warnings in 546.67 seconds.
Complete repository regression after liveness repair	passed	889 passed, two skipped, six warnings in 564.39 seconds.
Complete repository regression after retry-persistence repair	passed	889 passed, two skipped, six warnings in 546.86 seconds.
Live runtime identity	passed	Hiro is healthy and connected; loaded and checkout revisions both equal 0df25bb788a5932d022e5dc753cfe17d4c8e793d.
Live experiment execution	passed	Forty of forty chat API requests completed; raw responses and contract receipts were retained under immutable preregistration hashes.
Bounded candidate construction	failed	Four of four permitted attempts failed the unchanged target replay at CAUSAL_CANDIDATE_CONSTRUCTION_TO_TARGET_TEST; the queue terminated artifact_blocked.
Candidate promotion	not reached	No promotion claim is made; the candidate never reached evaluation, canary, governor, activation, probation, or promotion.

Current state

The canonical candidate incident-live-riemann-completeness-20260904 is terminal in artifact_blocked after four bounded construction attempts.
The exact unresolved capability is production-reachability reasoning during patch construction: the builder repeatedly guarded on task_type metadata that the targeted production path does not populate.
Hiro and its local model remain online; production is healthy on 0df25bb788a5932d022e5dc753cfe17d4c8e793d.
The inner improvement loop is not agentically closed: an operator created and prioritized this queue record from the frozen-order live experiment, and the builder could not independently produce a reachable repair. No manual promotion decision or gate override occurred.

Next steps

Qualify the builder's ability to inspect and test production reachability before it writes a patch, using this preserved artifact-blocked candidate as the frozen regression case.
Require a subsequent versioned builder attempt to select a runtime-observable discriminator or the actual prompt-building surface, and to pass the unchanged replay without modifying its acceptance contract.
Only after construction passes should the unchanged canary, governor, activation, verification, probation, and finalization path resume.