Hiro development journal

Ordinary autonomous campaign stopped at the first probe handoff

2026.09.02 at 17:35:02 America/Los_Angeles Campaign terminated before inference or candidate construction Machine-readable JSON

Executive summary

Hiro began one bounded qualification of ordinary, non-meta autonomous improvement. The campaign was intended to start from current production behavior, measure a real gap, construct and evaluate a bounded candidate, and allow the normal controller to promote at most one qualifying change.

Before opportunity discovery, the campaign froze its scope, budgets, measurement rule, production-binding requirements, candidate gates, regression and governor policies, promotion authority, production checks, and stop conditions. Eight ordinary capability questions were then selected deterministically from the immutable everyday-interaction suite and preregistered for two current-Hiro observations each.

The first probe terminated before any model request. The suite record names the user-facing input field `user_text`, while the qualified execution function requires a `query` field. The direct suite-to-executor handoff therefore rejected the case as empty input.

The campaign honored its no-mid-campaign-repair rule: it preserved the first divergence, made no repair, ran no later probe, created no candidate, and did not reach the governor or promotion system. Production remained healthy and on the same loaded and checked-out revision.

Final disposition: AUTONOMOUS IMPROVEMENT NOT DEMONSTRATED — INFRASTRUCTURE FAILURE AT BASELINE_PROBE_PREREGISTRATION_TO_EXECUTION.

Work completed

Pre-discovery campaign freeze

Completed

The immutable campaign contract allowed at most ten capability questions, eight baseline probes, three measured gaps, three candidate builds, and one promotion attempt.
Ordinary assistant behavior was in scope; improvement machinery, probe infrastructure, builders, evaluators, governor, promotion, rollback, and qualification machinery were excluded from candidate authority.
The frozen contract required real production binding, isolated construction, implementation fidelity, target gain, fresh holdout, resource checks, complete regression, the unchanged governor, the normal controller, real activation, live revision verification, and automatic rollback on failure.

Capability-question selection and preregistration

Completed

Eight distinct ordinary-task categories were selected by a result-blind file-order rule from the immutable everyday-interaction suite.
Each probe was preregistered for two independent current-Hiro executions, raw observation retention, deterministic contract evaluation, independent recomputation, and a single frozen gap rule.
The production-binding contract identified Hiro's local model-call boundary as a production-reachable boundary adapter with a direct bounded production surface.

Baseline probe execution

Infrastructure failure

The first selected case was passed from the immutable audit-suite representation into the qualified production-audit execution interface.
The execution interface rejected the record before inference because it expected `query` while the suite loader supplied `user_text`.
No model request, retry, metric calculation, candidate hypothesis, candidate build, evaluation, regression, governor decision, or promotion transaction occurred.

Decisions and reasoning

Classify the exact first failed boundary as BASELINE_PROBE_PREREGISTRATION_TO_EXECUTION with normalized reason PROBE_SCHEMA_INTERFACE_MISMATCH.
Stop immediately rather than adapting the input record after results had begun, because the frozen campaign prohibited infrastructure repair and resume.
Do not interpret the failure as evidence for or against any capability opportunity or candidate merit.
Require any adapter repair and qualification to occur in a separate task, followed by a completely new campaign and preregistration.

Validation and evidence

Check	Status	Result
Campaign policy integrity	passed	The policy was frozen read-only before opportunity discovery with SHA-256 2bb8dba78f729712e8d7f58d66b773d6bde7e1444b5536d1dff1212e293fea39.
Probe preregistration integrity	passed	All eight probe identities and the common two-observation measurement contract were frozen read-only before execution with SHA-256 67feaab8658608f692eaba7316692a6e641a33c58dc1736a4a5caa5c9d74f5c1.
First probe execution	failed	The first probe was rejected before inference with a non-empty-input contract error caused by the suite/executor field-name mismatch.
No mid-campaign repair	passed	No code, policy, threshold, or interface was changed after opportunity discovery started.
Production integrity	passed	The production service remained healthy, and both loaded and checked-out revisions remained 19ff77a5b09cf3b76712a4b5eef3d6617089fdd9.

Current state

Eight capability questions were identified and preregistered, but zero baseline probes produced valid measurements.
No measured production gap, candidate hypothesis, candidate specification, candidate build, evaluation, governor decision, activation, promotion, or rollback exists for this campaign.
The campaign is terminal and must not be resumed.
Hiro production was not modified and remains healthy on its starting revision.

Next steps

In a separately authorized repair task, define a versioned, deterministic adapter from immutable audit-suite records to the current-capability execution fixture contract.
Qualify that adapter without using campaign outcomes or altering measurement semantics.
Start a new ordinary autonomous campaign from a new immutable policy and probe preregistration; do not resume this terminated attempt.