GPUaaS Harness-Record And Trajectory Pilot

Date: 2026-08-21

Fairway task: FW-413

Consumer task: FAIRWAY-TRAJECTORY-SUPERVISION-PILOT-001

Fairway implementation: d74d025

Decision

Keep and narrow. Keep the versioned harness records, task-local named cohort, evaluator-backed efficiency projection, and cited repeated-pattern advisory. Narrow the current claim to an experimental, manually integrated read-only capability. Do not yet add automatic provider redirects, cross-task or cross-provider comparisons, model routing, or product-availability claims.

The pilot demonstrates that a consumer can preserve a failed strategy, detect its repetition from durable facts, record a materially different redirect, and show the resulting evaluator-backed observation without granting the evaluator or advisory any Fairway workflow authority. One pilot is not evidence that the detector generally improves engineering outcomes.

Consumer Workflow

GPUaaS had already configured a trajectory-redirect packet and measured its fail-closed behavior. The public harness contract records that real workflow as three attempts:

1. render with objective-only context; deterministic evaluator fails because current_strategy is missing;
2. deliberately repeat the under-specified strategy as a detector-calibration control; the exact command again fails closed on missing current_strategy;
3. redirect to complete context containing the old strategy, observations, new bounded experiment, expected observation, stop condition, and authority boundary; the evaluator passes.

The fixture is gpuaas-trajectory-supervision-harness-records.json. It uses safe summaries and artifact references, not prompts, private reasoning, transcripts, raw tool bodies, generated-content dumps, or credentials.

Measured Readback

The exact batch was ingested into the GPUaaS Fairway store through the public harness ingest command. The first ingest inserted three runs, three observations, and three evaluator results. An identical second ingest inserted nothing and reported all nine source-scoped identities as existing, confirming canonical replay behavior.

harness report then returned:

Measure	Result
External-run attempts	3
Explicit recorded actions	3
Evaluator results	3
Evaluator-backed outcomes	3
Named cohort	configured-packet-render@1.0/configured-packet:trajectory-redirect/GPUasService-local/complete
Attempts per evaluator-backed outcome	1
Actions per evaluator-backed outcome	1
Usage/cost efficiency	unavailable; no task usage event or exact comparable cost denominator
Repeated-action finding	two rejected/inconclusive objective-only actions; recommend change_execution_profile
Repeated-evaluator finding	two failed/inconclusive results without changed revision or hypothesis; recommend reframe_hypothesis

The second baseline is a controlled calibration input, not a claim that the consumer naturally became stuck twice. It provides a known repeated pattern so the detector's citations and false-positive warning can be inspected. The later passing result remains in the same named evaluator cohort and cites a new hypothesis and action fingerprint. The findings preserve the historical reason for redirect; they do not disappear merely because the next experiment passes.

False Positives And Missing Data

The detector cannot know whether two identical fingerprints hid materially different inputs or environments. It also cannot see a meaningful strategy change that the producer failed to express through revision, hypothesis, or action identity. A supervisor must inspect the cited rows before redirecting.

This pilot intentionally had no provider-usage event correlated through a session referenced by the external runs. Fairway therefore withheld token, elapsed-time, and cost ratios rather than assigning zero or borrowing unrelated task usage. That is the correct result and a concrete integration requirement for any future efficiency study.

The no-new-evidence advisory did not fire because the consumer task had no live provider session. A stale checkpoint alone is insufficient. This avoids presenting historical task metadata as current execution stagnation.

Authority Readback

Ingestion, replay, reporting, and the dashboard projection did not:

• change either task's state;
• create or accept Fairway evidence;
• create or satisfy a review;
• contact, cancel, pause, resume, or redirect a provider;
• approve a merge, deployment, release, or live operation; or
• make the GPUaaS packet or this integration generally available.

The stored recommendation is a next question for a supervisor. Normal task, review, source-control, environment, and promotion authorities remain external.

Next Validation

Before broadening the feature, run it on several naturally occurring stalled tasks with producer-emitted session correlation and exact or explicitly estimated usage. Measure whether a cited redirect produces new evaluator-backed information with less repeated work and acceptable reviewer effort. Add more cohort support only when separate populations can remain visibly separate; do not average incompatible tasks, evaluators, subjects, environments, or completeness levels.