Hiro development journal

Why the active improvement queue is waiting

2026.09.14 at 21:42:58 America/Los_Angeles Diagnostic findings; no runtime changes Machine-readable JSON

Executive summary

The initial queue-only explanation was incomplete. A full reservoir audit found 100 source leads grouped into 33 approaches: 19 within the current policy and 14 outside it. The external corroboration path reruns fixed regression tests rather than constructing claim-specific experiments.

Work completed

Queue eligibility and running revision

Diagnosed

Read-only inspection confirmed ten queued external mechanisms are outside ordinary_non_meta scope; two incident records remain observing.
The latest recorded audit passed four interaction cases and five product-path checks, with no new incident queued.
Discovery retained a reservoir of 100 leads, ten available, and admitted no new records in the inspected run.

Full reservoir and evidence-path audit

Diagnosed; not repaired

Dashboard linkage classifies the 100 source leads as 72 linked to rejected records, 18 linked to queued records, and 10 available. Multiple source leads represent the same approach.
All 19 in-scope grouped approaches have terminal queue history. All ten leads labeled available are outside the current scope.
The reservoir loader ranks historical leads and caps the result at 100 without filtering terminal queue outcomes; this is a retained source list rather than 100 ready experiments.
_run_external_corroboration selects fixed mechanism-level regression tests and an arithmetic sanity control. Passing existing tests produces absent evidence; it does not constitute a claim-specific experiment.
Existing evidence_acquisition.py already documents this architectural shortcoming. This session verified that the canonical external corroboration code still uses that path.

Decisions and reasoning

Do not bypass candidate eligibility or weaken regression gates to manufacture activity.
No restart or policy change was needed to establish the cause.

Validation and evidence

Check	Status	Result
Live health and revision identity	passed	Hiro reports healthy and Qwen connected; loaded revision and checkout both 6bed350ebdb1e9be116a2e94dc571bea7fab3033.
Queue scope inspection	passed	All ten queued records evaluated false using the controller scope predicate under ordinary_non_meta.
Latest failed promotion receipt	failed	The later candidate was rejected by the governor full-suite gate: 1 failed, 1045 passed, 11 skipped. It was not activated. These are recorded run results, not a new test execution.

Current state

Correction-recovery promotion finalized at 2026-09-14 15:26 PDT; current revision 6bed350ebdb1e9be116a2e94dc571bea7fab3033.
No active supervisor episode; circuit breaker closed and qualification current.

Next steps

Repair claim-specific evidence acquisition and connect it to the existing canonical controller.
Separate source leads, grouped approaches, eligible work, policy exclusions and exhausted evidence attempts in the dashboard.
Retain independent testing and promotion gates; do not manufacture baseline failures or simply reset rejected ideas.