Hiro development journal

Phase 3A stops at missing source claims before reproduction

Source quality insufficient; the fixed sixteen-lineage reservoir contains no operationalizable retained claim Machine-readable JSON

Executive summary

Accepted Phase 3 as not demonstrated and qualified only the evidence-acquisition boundary upstream of candidate construction. Promotion, activation, probation, rollback, thresholds, and meta-improvement were outside scope and unchanged.

Traced the same sixteen ordinary, non-meta Phase 3 lineages from selection through corroboration. Their retained records contained source identifiers, URLs, content hashes, controlled-vocabulary mechanism labels, and code-owned experiment suggestions, but none retained the external source's actual claim.

The prior corroboration implementation did not build claim-specific experiments. It reran fixed mechanism-level regression files and correctly treated healthy generic controls as absent evidence.

Five records had been called structurally complete under the legacy schema because they carried a claim ID, source hash, expected direction, protocol label, and keyword tuple. Phase 3A established that those fields still do not state a falsifiable source claim.

Added an explicit evidence-acquisition qualification layer, a preregistered evidence-package contract, append-only per-lineage assessments, and granular funnel milestones without creating a second state authority.

All sixteen lineages classified as CLAIM_EXTRACTION_INCOMPLETE. Zero were testable, zero hypotheses or reproduction packages were ready, and zero reproductions were executed. All sixteen Phase 3A corroboration decisions were DEFER.

The exact completion disposition is Source quality insufficient. The retained discovery reservoir does not contain enough operationalizable claims to establish whether autonomous reproduction construction works. Phase 3A is not demonstrated.

Work completed

Pre-change evidence path

Traced
  • External discovery reduced untrusted text to prompt-safe metadata and controlled-vocabulary labels. Raw source text was intentionally not persisted or sent to the model.
  • Selection moved a queue record into investigating. External corroboration then chose fixed regression files based on the broad mechanism, such as routing or memory, rather than constructing a source-claim-specific experiment.
  • The corroboration gate expected a baseline assertion failure plus a healthy negative control. Because the broad suites passed, evidence was classified absent and the candidate deferred.
  • BUILT in the existing funnel means construction and testing of the proposed Hiro modification. It does not mean construction of a reproduction experiment.
  • No component between selection and corroboration was responsible for extracting an exact source claim, preregistering a hypothesis, building an isolated experiment, or executing that experiment.

Fixed sixteen-lineage audit

Completed
  • The qualification used an immutable manifest containing exactly the sixteen Phase 3 lineages. It did not fetch easier replacement ideas, seed candidates, select a preferred idea, or authorize implementation.
  • Eleven queue records failed even the legacy structural-claim check. Five passed that legacy check but contained only keyword tuples and generalized metric directions, not an exact source claim.
  • All available historical source packets were indexed by their retained source-reference identifiers. The audit recovered metadata for most references but confirmed that exact external claims were deliberately absent from every retained projection.
  • Because the source content itself was not retained, the qualification cannot honestly distinguish a source that had no testable claim from a source whose claim extractor discarded the needed semantics. It therefore uses CLAIM_EXTRACTION_INCOMPLETE rather than claiming the original source was false or useless.

Evidence-package contract

Implemented
  • A valid package requires a source claim, exact hypothesis, expected observable, baseline or control, experimental treatment, measurement, acceptance criterion, reproduction specification, source provenance, and preregistration timestamp.
  • Execution results, raw artifact references, and interpretation are result fields and cannot be present in the preregistered package.
  • The preregistered portion receives a deterministic hash so an acceptance criterion cannot be silently changed after the result is known.
  • Negative results remain valid evidence when the experiment infrastructure is valid. Infrastructure failures remain inconclusive rather than evidence against a claim.

Authoritative funnel extension

Implemented
  • The continuous queue and append-only event ledger remain the only lifecycle authority. The Phase 3A report is a projection, not a second workflow database.
  • New observable milestones are TESTABLE_CLAIM, HYPOTHESIS_READY, REPRODUCTION_READY, REPRODUCTION_EXECUTED, REPRODUCTION_VALID, CORROBORATION_PASS, CORROBORATION_FAIL, CORROBORATION_DEFER, and CANDIDATE_BUILT.
  • Each fixed lineage received one idempotent evidence_acquisition_assessed event containing its granular reason and stage flags. Rerunning the qualification does not duplicate those events.
  • Future discovery records now explicitly declare claim extraction incomplete rather than presenting a keyword tuple as a fully extracted source claim.

Phase 3A empirical result

Source quality insufficient
  • Sixteen lineages assessed; zero testable claims; zero hypotheses ready; zero reproductions ready; zero reproductions executed; zero valid reproductions.
  • Corroboration decisions were zero PASS, zero FAIL, and sixteen DEFER. The normalized evidence-acquisition reason was CLAIM_EXTRACTION_INCOMPLETE for all sixteen.
  • No local experiment was fabricated from a generic topic label. No candidate was built and no promotion or meta-improvement was attempted.
  • The first unresolved boundary is claim extraction, before hypothesis operationalization and reproduction construction.

Decisions and reasoning

Validation and evidence

CheckStatusResult
Focused evidence-acquisition and corroboration tests passed Eleven focused tests passed after the initial implementation; the complete relevant three-file group then passed ninety-two tests.
Full repository regression suite passed 801 tests passed in 434.56 seconds. Six non-failing warnings concerned pre-existing unregistered test marks and an inaccessible pytest cache directory.
Fixed-reservoir qualification completed 16 assessed, 0 testable, 0 hypothesis-ready, 0 reproduction-ready, 0 executed, 0 valid, 0 corroboration passes, 0 corroboration failures, and 16 corroboration defers.
Append-only evidence integrity passed Exactly sixteen Phase 3A assessment events were recorded, one per frozen lineage. The assessment operation is idempotent.
Active runtime identity passed Hiro restarted healthy with its model connected and loaded revision equal to checkout revision ef510823ff7c72dc3e1d970e42bb12d478207694.
Public journal tests and build passed npm run test:hiro passed. npm run build generated and validated 170 journal pages, compiled TypeScript, and completed the Vite production bundle.

Current state

Next steps