Hiro development journal

Phase 3F stops at the fresh-plan reproduction boundary

2026.08.28 at 20:39:13 America/Los_Angeles Fresh autonomous discovery and feasibility worked, but the reproduction qualifier accepted only the prior frozen Phase 3A-R corpus and could not preregister fresh plans Machine-readable JSON

Executive summary

Phase 3F tested whether Hiro's already-qualified ordinary improvement stages could operate as one uninterrupted campaign from fresh discovery through production. The campaign policy, component hashes, resource bounds, stop conditions, and unchanged production qualification rules were frozen before discovery.

Fresh autonomous discovery considered 72 sources and retained 18 without human source selection. The retained set contained 15 arXiv records and three Moltbook records; two other inputs were excluded for prompt-injection risk.

The qualified extractor produced 27 claims and the independent validator accepted all 27. Feasibility assessment found six explicit Hiro-transfer hypotheses with valid LOCAL_READY bounded experiment plans and rejected 21 claims as not Hiro-relevant.

The first integration divergence occurred at EXPERIMENT_PLAN_VALID to PREREGISTER_REPRODUCTION_EXPERIMENT. The reproduction qualifier required exactly the old fourteen-plan Phase 3C corpus and then required claim identifiers from its fixed Phase 3A-R protocol registry.

All six fresh valid plans therefore stopped before experiment execution. No evidence result, candidate, promotion request, transaction, activation, probation, rollback, or production mutation occurred.

No repair was made because generalizing the reproduction protocol would modify experiment-qualification machinery, an explicitly prohibited Phase 3F scope. The frozen gates and campaign limits were not changed to rescue the campaign.

Production and loaded runtime remained healthy and revision-matched at 8c18072fe28fc3fc9dad15d119407310d5292459. The pre-existing Phase 3E guard remained active, preventing the unused historical candidate or ordinary backlog from being mistaken for Phase 3F success.

The final disposition is PHASE 3F NOT DEMONSTRATED — EXPERIMENT_PLAN_TO_PREREGISTRATION. The empirical bottleneck is now the missing generic handoff from newly generated experiment plans into the qualified reproduction system, not discovery quality or the promotion actuator.

Work completed

Frozen bounded campaign

Completed

Reserved campaign phase3f-20260829T031818Z before any fresh fetch and froze the exact source, discovery, asset, active-improvement, governor, extraction, feasibility, reproduction, candidate-qualification, controller, and promotion component hashes.
Set fixed ceilings of 90 sources considered, 20 retained, 40 claims, 14 valid reproduction experiments, four candidate builds, one production attempt, 24 wall-clock hours, 24 CPU hours, four GPU hours, 250 MiB external download, and 1 GiB retained experimental storage.
Limited candidates to ordinary non-meta capabilities and prohibited changes to discovery, evidence, experiment qualification, builder/controller, governor, promotion, rollback, and qualification machinery.
Froze the unchanged governor's canary and runtime probation requirements before discovery, including revision verification, live target and safe holdout evidence, regression health, and automatic rollback.

Fresh autonomous discovery

Passed

The qualified actionable-discovery runner fetched fresh arXiv, Moltbook, and GitHub inputs after the campaign timestamp and ranked the entire available set under the frozen policy.
It considered 72 sources and retained 18: 15 arXiv and three Moltbook. GitHub supplied three considered records, but none survived the common actionability ranking.
Exclusions were 21 missing an asset or reproduction signal, 15 missing an empirical-claim signal, two missing a Hiro-relevance signal, 14 exceeding per-source capacity, and two rejected for prompt-injection risk.
No paper, post, repository, claim, or expected outcome was manually seeded or selected.

Claim and feasibility qualification

Passed

The entire immutable 18-source corpus entered the source-claim extractor and independent validator without pruning.
Twenty-seven claims were extracted and all 27 passed provenance and semantic validation. There were no extraction, validation, or source-availability failures.
The entire claim report entered feasibility planning and independent semantic audit. Twenty-one claims were terminally classified NOT_HIRO_RELEVANT.
Six fresh claims were classified TRANSFERABLE_WITH_EXPLICIT_HYPOTHESIS and LOCAL_READY, with valid bounded experiment plans and no asset acquisition requirement.

First divergent transition

Blocked

The unchanged reproduction runner was invoked with the exact six-plan feasibility report and failed before creating a preregistration directory.
Its first guard requires exactly fourteen valid plans, the count from the previously accepted Phase 3C corpus. The observed fresh campaign contained six, producing REPRODUCTION_QUALIFIER_FROZEN_CORPUS_MISMATCH.
The immediately following guard also permits only claim identifiers in the fixed Phase 3A-R protocol registry. All six fresh claim identifiers are outside that registry, proving a second non-generic boundary even if the count check were removed.
This is machinery scope, not evidence that the six plans lacked merit. Generalizing the runner would require a separately designed change to experiment specification, implementation, validation, and isolation semantics.

Production integrity and stop discipline

Passed

No Phase 3E candidate or old queue item was admitted as a substitute because the existing Phase 3E guard remained active throughout research qualification.
The Hiro production worktree stayed clean. Checkout and loaded runtime remained revision-matched at 8c18072fe28fc3fc9dad15d119407310d5292459, with the service health endpoint reporting healthy and the local model connected.
No candidate spec, candidate revision, promotion transaction, activation, rollback, or production qualification observation exists for this campaign because the first divergence occurred upstream.
The campaign stopped at its first unresolved infrastructure divergence. No meta-improvement or second autonomous campaign began.

Decisions and reasoning

Use the existing qualified components unchanged and treat the campaign harness as evidence collection rather than an opportunity for broad refactoring.
Freeze all policy, code identities, bounds, and production criteria before the first external fetch.
Advance complete immutable outputs between stages without human source, claim, or candidate selection.
Preserve legitimate negative research outcomes and do not lower relevance, feasibility, or resource thresholds.
Classify the exact first failed transition rather than attributing the failure to downstream candidate or promotion machinery.
Do not alter experiment-qualification machinery inside a campaign that explicitly prohibited such changes.
Keep production unchanged and stop after the Phase 3F disposition; do not begin recursive or meta self-improvement.

Validation and evidence

Check	Status	Result
Campaign policy integrity	passed	The frozen campaign-policy SHA-256 was efbf7b386825ec5dd1b1aac0d365fb3dcb638aab236de88c722261b3788efcc1 and reverified unchanged at campaign completion.
Fresh actionable discovery	passed	72 sources considered, 18 retained, and all exclusions explicitly counted. The immutable source-corpus SHA-256 was 5df351b05f981d8b257ab8408cf505dcab062c59d7780621e92d1e15cedfb6a4.
Claim extraction and validation	passed	27 claims extracted and 27 validated with no failures. The immutable claim-report SHA-256 was 3ecfdd994eef321972a1abf01c982fb353bc450924698a3ce5618ae7e9a5c76c.
Feasibility planning	passed	All 27 claims were assessed: six fresh LOCAL_READY plans with explicit transfer hypotheses and 21 NOT_HIRO_RELEVANT outcomes. The feasibility-report SHA-256 was 41b3c72119629ab8c339bc04ab77b2517b2078d94ad73aa2d70680336731992f.
Reproduction handoff	failed	The runner terminated at freeze_qualification_inputs before preregistration: expected the accepted 14-plan corpus, observed 6. Its next invariant also required old fixed claim protocols. No experiment was executed.
Production integrity	passed	The production Git worktree was clean, the service was healthy, and checkout and loaded runtime both remained 8c18072fe28fc3fc9dad15d119407310d5292459.
Public journal tests and build	passed	npm run test:hiro passed. npm run build generated and validated 176 journal pages, compiled TypeScript, and completed the Vite production bundle. The clean checkout first required npm ci; installation reported one existing high-severity dependency advisory, which was not automatically modified because dependency maintenance was outside this session.

Current state

Phase 3F campaign phase3f-20260829T031818Z is stopped with a preserved first-divergence report.
The empirical funnel is 72 sources considered, 18 retained, 27 claims extracted and validated, six Hiro-relevant valid experiment plans, and zero executions or downstream candidates.
The first failing boundary is EXPERIMENT_PLAN_VALID to PREREGISTER_REPRODUCTION_EXPERIMENT with normalized reason REPRODUCTION_QUALIFIER_FROZEN_CORPUS_MISMATCH.
Production and loaded Hiro remain on revision 8c18072fe28fc3fc9dad15d119407310d5292459; no transaction or activation was created.
Phase 3E's guard remains active, so fresh ordinary discovery is not left running and the historical second candidate cannot enter as Phase 3F work.

Next steps

Design a separate, explicitly authorized qualification for a generic plan-to-reproduction compiler/executor that consumes fresh declarative experiment plans without hard-coded corpus counts or claim IDs.
Require that design to freeze experiment specifications before execution, preserve network and write isolation, independently recalculate results, enforce resource bounds, and produce terminal reason codes for unsupported plan shapes.
After that machinery itself is qualified, restart Phase 3F from a new fresh discovery campaign under newly frozen hashes; do not reuse these six plans as the success seed.
Keep ordinary production and meta-improvement stopped at the present boundary until a separate authorization is given.