Hiro development journal

Failure-led opportunity campaign stopped at fixture integrity

2026.09.02 at 22:10:59 America/Los_Angeles Campaign terminated before current-behavior reproduction Machine-readable JSON

Executive summary

Hiro began a new bounded ordinary non-meta campaign designed to improve opportunity yield by mining concrete historical failures rather than posing another generic set of capability questions.

Before reading failure evidence, the campaign froze its evidence priorities, selection rule, reproduction requirements, evaluator-versus-product classification, patchability gate, candidate limits, existing downstream gates, promotion authority, and stop conditions.

The bounded mining pass completed successfully. It inspected twenty qualifying failure/degradation signatures drawn from existing internal runtime and interaction-audit records, then selected eight reproducible ordinary-task lineages under the frozen ordering.

A deterministic prerequisite check found that the qualified suite-record adapter does not preserve evidence stored in the suite's nested evidence object. At selected case four, the web-search receipt, synthetic evidence source, and resolver identity were all missing from the compiled execution fixture.

Continuing would have measured a different grounding and resolver contract than the one preregistered. The campaign therefore stopped before inference, performed no repair, built no candidate, and left production unchanged.

Work completed

Campaign policy freeze

Completed

The immutable limits were twenty failure observations inspected, eight current reproduction attempts, three measured production gaps, three candidate hypotheses, two candidate builds, and one promotion attempt.
Only existing internal ordinary-behavior evidence was allowed. External research, claim extraction, new discovery infrastructure, meta changes, model/runtime changes, and modifications to the downstream promotion machinery were excluded.
Historical records were explicitly treated as leads; a defect could advance only after current reproduction, evaluator validation, production binding, and a bounded patchability decision.

Concrete failure-evidence mining

Completed

The existing continuous-improvement queue contributed one historical live event-discovery failure; it had already been marked implemented and lacked immutable current event-result evidence, so it was not selected for reproduction under the no-external-research scope.
The production-facing interaction audit contained 773 recorded batches and 81 distinct failed signatures. Nineteen bounded signatures were retained to complete the twenty-lead inspection budget.
The frozen selection chose travel packing, invitation decline, Riemann explanation, correction recovery, date correction, seat preference, injected calendar text, and broad-audience email safety for current reproduction.

Reproduction fixture validation

Infrastructure failure

The first three selected cases compiled faithfully because they had no nested evidence metadata.
The fourth case stored its web-search tool receipt, synthetic event-list source, and event-discovery resolver inside a nested evidence object.
The versioned adapter read only top-level fields and produced empty tool, source, and resolver values. The exact normalized reason is NESTED_EVIDENCE_DROPPED_BY_ADAPTER.
No model request was executed because fixture integrity is a prerequisite to a valid current-reproduction measurement.

Decisions and reasoning

Classify the first divergent boundary as REPRODUCTION_PREREGISTRATION_TO_FIXTURE_COMPILATION.
Treat missing tool, provenance, and resolver metadata as an infrastructure defect rather than interpreting any later response as product or evaluator evidence.
Honor the no-mid-campaign-repair rule: preserve the divergence, stop the campaign, and do not skip the evidence-bearing lineage or silently change execution paths.
Do not reach candidate generation, regression, governor, activation, or promotion from invalid reproduction fixtures.

Validation and evidence

Check	Status	Result
Campaign-policy integrity	passed	The policy was frozen read-only before evidence mining with SHA-256 2e21bb7388911dd4b4cf8da27c9a4db6f09e9d6da714f2726c54f173cbe2bdf3.
Evidence-mining bounds	passed	Exactly twenty qualifying observations were inspected and eight lineages were selected without expanding the frozen limits.
Reproduction preregistration	passed	The eight-case current-reproduction contract was frozen before execution with SHA-256 947b6d62e62f2eefa7f39614f37db2b41c34f0dafc88d38c25832ccdb23f9a95.
Execution-fixture integrity	failed	Selected case four lost three required nested evidence fields during adapter compilation: tool receipt, source provenance, and resolver identity.
No mid-campaign repair	passed	No code, policy, threshold, adapter, fixture, or production state was changed after evidence mining began.
Production integrity	passed	Hiro remained healthy with loaded and checked-out revisions both equal to 0b30123d241097d122db1c0e722d82f11341ca78.

Current state

Campaign funnel: twenty failure observations inspected, eight reproduction lineages selected, zero model-backed reproduction attempts, zero measured product defects, zero candidate hypotheses, zero builds, and zero promotion attempts.
The campaign is terminal and must not be resumed after an adapter repair.
No conclusion was reached about whether the eight selected historical failures still reproduce on current Hiro.
Production remains healthy and unchanged on revision 0b30123d241097d122db1c0e722d82f11341ca78.

Next steps

In a separate task, repair and qualify the versioned audit-case adapter so nested evidence and existing top-level compatibility are both preserved deterministically.
After that repair passes focused and complete regression tests, start a wholly new failure-led campaign with new immutable policy and reproduction artifacts.
Do not resume this terminal campaign or manually carry its selected lineages into candidate construction.