Hiro development journal

Ranked autonomy candidate: intake, durable outcomes and parent qualification

2026.09.20 at 13:34:46 America/Los_Angeles Partial candidate; frozen pipeline passed; full suite timed out; not activated Machine-readable JSON

Executive summary

Extended the existing continuous queue with GitHub handoff intake, highest initial priority, revision provenance and retryable receipts. No new queue, database or daemon was added.

Routed the legacy promoter into the canonical queue and replaced briefing approval prompts with post-action queue-event oversight.

Added a bounded parent-owned qualification protocol for internal proposal-surface policy changes. Broader meta changes still require mechanism-specific independent evidence.

Work completed

Live architecture reconciliation

Inspected

Used the active local checkout, not the older GitHub default branch. Traced queue, scheduler, candidate, governor, transaction, probation and rollback paths.
Found ordinary-only selection, protected-path/meta-release stops and a still-callable legacy risk-based promoter. Historical/manual labs remain distinct from the canonical scheduler.

Durable handoff intake and processing

Candidate implemented

Reused the GitHub issue schema, author allowlist, continuous runner lease, queue ideas/events and runtime state.
Both conversational clients receive existing make_next semantics; exact duplicates retain lineage without permanent priority. Edited requests preserve prior revisions.
Reconciliation retrieves a verified derived memory view. A bounded planner can propose an existing code-owned experiment but cannot assert novelty, select authority or certify implementation. Missing evidence/planners remain UNKNOWN and observing.
Added idempotent research seeds for ranking, lineage exploration, memory interpretation and modification-envelope questions; no superiority or novelty is assumed.
Every accepted issue revision retains its complete source body and hash. Retrieved current-state text accompanies its manifest as bounded evidence.

Parent qualification and oversight

Bounded candidate implemented

The parent governor reads immutable policy blobs and independently checks control cases. Candidate-provided success claims do not qualify a policy change.
Ordinary protected paths remain unavailable. The first meta protocol covers internal proposal eligibility, not arbitrary evaluator/governor rewrites or evidence-free capability claims.
High risk has a longer canary schedule; existing rollback, circuit breakers, runtime identity and probation requirements remain.
Legacy experiment results enter the same queue; briefing content derives from queue events and retains item identifiers.
Consolidation includes the legacy caller before patch construction/application, not only its final promotion function.

Decisions and reasoning

Preserve one queue, one promotion authority and existing evaluation records.
Do not activate broad autonomy on branch tests alone or equate a policy scope value with demonstrated general meta capability.
Keep raw evidence, Mission Statement, source authority, discovery query and production memory behavior unchanged.

Validation and evidence

Check	Status	Result
Focused regression runs	passed	112 queue/governor/durability tests passed on an intermediate candidate; later build/loop/parent-protocol run passed 102 tests with three existing unknown-marker warnings.
Real local model probe	failed	The real candidate-model preflight failed with CandidateModelInfrastructureError: ReadTimeout in 5.25 seconds. A separate minimal completion probe timed out at 20 seconds. Live readiness was not established.
OS candidate executor	passed	Actual WSL/Bubblewrap readiness probe reported OS isolation, network denial and live eligibility.
Frozen pipeline qualification	passed	Exact candidate c41af1195948340066986d7480b53af0253bf55c passed all five scenarios, 46 tests. Packet SHA-256 cfe4c07d557b9fb3c21a987cba9589618ee0ad2348d2bd1a7399e0b8595166dd. This bounded result is not production probation or a full-target acceptance result.
Journal tests/build	passed	npm run test:hiro passed; npm run build passed, including validation of 240 journal entries, timestamp/alias/feed/noindex checks and frontend compilation.
Full regression suite	incomplete	The full suite timed out at 1,200 seconds after reaching 73% and starting the next group. No assertion failure had printed, but no final summary exists. This is not a passing full-suite result. The existing 900-second governor budget remains unchanged.
Runtime provision and launcher	passed	An earlier qualification failed for a missing pinned runtime in the isolated checkout. The existing provisioner supplied it and real launcher validate-only passed. Superseded runs interrupted for further corrections are not counted as passed.

Current state

Candidate remains isolated. No promotion, service restart, probation completion or production activation has been claimed.
General evaluator/governor self-modification and semantic status interpretation remain beyond the demonstrated bounded protocol.
Research seeds exist in candidate code and isolated tests; the production queue has not received them because this loop has not been activated.
No live end-to-end handoff-to-promoted-outcome trace exists for this candidate. General meta protocols and comprehensive briefing qualification remain incomplete.
Candidate c41af1195948340066986d7480b53af0253bf55c is pushed to codex/ranked-autonomy. Live checkout remains 20cd497cf6dff8a27e4fa3eccc8f0d85356d842d. Full-target completion is not claimed.

Next steps

Resolve model/evaluator readiness and finish the full regression gate.
Complete missing independent meta and end-to-end qualification before claiming the full architecture.
Integrate, activate and record production probation only when the applicable evidence gates pass.