GPUaaS Quality Record Pilot

Date: 2026-08-05
Fairway task: FW-395
Fairway source: 1582601bf02bc38142468736dff1b44aafc9e555
GPUaaS source: 8000199e21b968c7ba3370d876acc1ae67472302

Purpose

This second GPUaaS quality-system pilot tests whether the versioned fairway.quality-record.v1 projection can reconstruct a bounded, cited record from a large consumer's existing Fairway history. It also establishes the post-instrumentation baseline for task-to-commit links, structured outcomes, and attributable control friction.

The pilot does not score task quality, qualify verifiers or reviewers, infer missing facts, establish causal control effectiveness, or backfill historical records.

Method

The committed Fairway binary was run against the GPUaaS platform-foundation store after a consistent SQLite backup. The population was every task that was still done and had a done transition in the preceding 30 days: 288 tasks. Each task was projected independently with:

fairway --config .fairway/platform-foundation-config.toml \
  quality-record <task-id> --format json

Six recent cross-domain tasks were retained as drill-down examples. The same binary also generated a 30-day control-effectiveness report. Generated raw records remain in the GPUaaS tmp-ux workspace; only bounded aggregates and digests are committed here.

Population Result

Quality Record stage	Present	Other state	Interpretation
Intent	218 (75.7%)	70 missing	Titles and roles existed, but 70 records lacked at least one required intent element, usually acceptance checks.
Material decisions	5 (1.7%)	283 missing	Decision capture was exceptional rather than normal.
Production context	246 (85.4%)	42 missing	Sessions, checkpoints, or commit context existed for most tasks.
Collected evidence	288 (100%)	0 missing	The configured completion evidence gate is visible in the historical population.
Automatic verification	287 present	1 conflicting	The projection retained one blocked-then-passed verifier history as conflicting instead of silently selecting a winner.
Human judgment	195 (67.7%)	93 missing	Attributable review rows existed for two thirds of the population. Missing review is not treated as failed review.
Promotion decision	0 owned	288 externally owned	Fairway reports readiness and completion facts but does not claim Git, CI/CD, deployment, or accountable-person authority.
Operational outcomes	0 present	288 unavailable	The new structured outcome table had no historical rows and was not backfilled.
Lessons	0 present	288 missing	No explicit corrective/superseding outcome or lesson handoff met the bounded projection rule.

Across 2,592 stage records, the projection reported 1,239 present, 776 missing, 288 unavailable, one conflicting, and 288 externally_owned. The distinction is useful: missing configuration, unavailable outcome instrumentation, contradictory verifier history, and external promotion authority are not collapsed into one generic incomplete state.

Conflict Drill-Down

UAT-HANDOFF-END-USER-LIFECYCLE-001 retained the same Playwright invocation as blocked twice and later passing. The projection reports the verification stage as conflicting and cites all three evidence rows. It does not infer that the later pass supersedes the earlier environmental or product failure. A durable supersession or interpretation record is still required.

This is evidence that the projection can expose a review question. It is not evidence that the verifier was qualified or that the final artifact was correct.

Control Measurement Baseline

The contemporaneous 30-day control report contained 449 eligible commits and 154 covered commits: 34.3% commit-to-task coverage. Changed-file coverage was 99.0%, but broad task path ownership cannot substitute for durable commit association. The report found zero explicit task_commits links, zero structured outcomes, and zero attributable friction samples in GPUaaS history.

The small difference from the 34.6% commit coverage measured on 2026-08-02 is not interpreted as a trend because the windows have different endpoints. Both measurements remain below the 80% interpretation threshold.

The new Fairway features establish a forward measurement boundary:

• work start and work close can now retain baseline, work, and completion commit associations;
• structured outcomes can distinguish incidents, rollbacks, reopens, corrective work, and superseding work from Git touch proxies;
• attributable friction can distinguish measured, open, unavailable, and missing control cost;
• the Quality Record exposes these facts without filling historical gaps.

No historical association, outcome, or friction row was manufactured for this pilot.

Operational Characteristics

The complete 288-task projection took 7.49 seconds wall time on the local consumer environment, approximately 26 milliseconds per task when invoked as separate CLI processes. Repeating a representative projection produced byte-identical JSON after removing generated_at.

The projection remained read-only apart from applying the product's normal schema migrations. A consistent pre-migration database backup was retained in the project-local temporary evidence directory.

Decisions

1. Keep the Quality Record as a supported read-only projection. It makes evidence boundaries and missing instrumentation visible without creating a score or new authority.
2. Do not reposition Fairway as a complete AI Quality System yet. The pilot validates record reconstruction and data-quality diagnosis, not trustworthy outcomes or process capability.
3. Use forward instrumentation instead of historical backfill. New GPUaaS work should use normal work start/work close, structured outcomes when an outcome occurs, and friction records when control cost is discussed.
4. Retain coverage-first suppression. Control rankings remain invalid until representative commit association and outcome coverage meet the reviewed thresholds.
5. Add explicit verification supersession before interpreting conflicts. Chronology alone is insufficient to decide whether a later pass closes an earlier failure.

Next Measurement Gate

Run the next GPUaaS measurement only after at least one natural 30-day window uses the new association and outcome/friction commands. Interpret a control only when the existing coverage, sample, risk, size, and authority gates pass. Until then, report adoption coverage, missing data, conflict yield, and measured friction without effectiveness rankings.

Evidence

The bounded committed summary is evidence/fw-395/gpuaas-quality-record-pilot-summary.json. Raw project-local artifacts are retained outside version control under:

GPUasService/tmp-ux/fairway-quality-record-pilot-2026-08-05/