Skip to main content

GPUaaS Control-Effectiveness Pilot

Date: 2026-08-02

Fairway task: FW-387

Fairway source: 5b5d8ddd741ba745af30ddd252100dea425bce22

GPUaaS source: 2bdba92f3cf1b870cd34ac47dd89eaa21e71e00e

Scope

This pilot ran the implemented fairway.control-effectiveness.v1 advisory model against the GPUaaS platform-foundation Fairway store and Git history. It asked whether the current records support interpretation of representative review, evidence, preflight, security, and process controls.

The pilot is an observational data-quality assessment. It does not establish causal control impact, change policy, or authorize removal of a control.

Method

The same committed Fairway binary produced JSON reports for contemporaneous 14-, 30-, and 90-day windows:

fairway --config .fairway/platform-foundation-config.toml \
control report --since <336h|720h|2160h> --format json

The project had no control-effectiveness override. The run therefore used the versioned product defaults:

  • minimum sample size: 5;
  • minimum coverage ratio: 0.80;
  • material outcome delta: 0.10;
  • high-friction p90: 900 seconds;
  • configuration revision: unversioned;
  • 90-day configuration digest: sha256:4cfee86de7bd5ffe9493e2ac60e422dac4f114d3fc2208f1953229dcdf842234.

Representative 14-day outcome horizons were inspected inside deterministic small, medium, large, and unknown diff-size bands. The pilot did not combine those bands into one effectiveness ranking.

Coverage Result

Coverage is the first result and the interpretation gate.

Observation windowEligible commitsCovered commitsCommit coverageEligible filesCovered filesFile coverage
14 days2254821.3%3,1542,96393.9%
30 days45415734.6%7,3487,28299.1%
90 days2,42543417.9%22,49222,36599.4%

Commit-to-task coverage is below the configured 80% threshold in every window. The high changed-file coverage does not compensate for this gap. Broad task source/target path ownership can cover files even when individual commits are not durably associated with the task that produced them.

All 195 14-day, 363 30-day, and 555 90-day cohorts were therefore classified insufficient_coverage. This suppression is correct product behavior.

Current Maintenance Facts

The 30-day report contains:

  • 257 promoted tasks;
  • 12 unique configured control IDs across 363 risk/size/horizon cohorts;
  • 1,569 observed control facts;
  • 186 unknown control-state facts;
  • no explicit attributable bypass facts;
  • no attributable friction facts;
  • no structured incident, rollback, reopen, corrective, or superseding-task outcome links;
  • Git-derived post_promotion_touch as the only available outcome.

The Git-touch proxy is too common to stand alone as a defect signal. Among the 367 applicable, mature, outcome-known facts where the control was observed at the 14-day horizon, 335 contain post_promotion_touch. This is still a task/control-fact count: a task may appear once per applicable control, and a same-file touch may represent planned follow-up, adjacent feature work, or refactoring. It must not be called escaped rework.

Representative Controls

The table below describes execution signal at the 30-day window and 14-day outcome horizon. It does not compare outcome rates because the coverage and bypass denominators are insufficient.

ControlApplicableObservedUnknownTriggeredPassedFriction samples
review:architecture115108712960
review:backend107931415780
review:ops105911416750
review:product-quality474256360
review:security10395814810
gate:platform-foundation:boundary-guard-report651050
gate:platform-foundation:ownership-map-evidence1082080

Review controls produced changes-requested signals in every represented size band. Architecture review triggered on 4 of 61 observed large tasks, 3 of 16 medium tasks, and 5 of 31 small tasks. Backend review triggered on 6 of 45 large, 3 of 19 medium, and 5 of 27 small tasks. Security review triggered on 6 of 53 large, 2 of 15 medium, and 5 of 26 small tasks. These are execution-yield facts, not proof of incremental effectiveness.

The two evidence gates have observable pass coverage but no retained trigger semantics or attributable friction. No stable preflight-family control appears in the current report. review:security is represented as quality_gate, not as a configured mandatory security_invariant. Those are instrumentation and configuration gaps; they are not evidence that the controls are ineffective.

Decisions

DecisionBoundaryRationale
KeepCoverage-first suppressionIt prevented low-linkage records from producing persuasive but invalid rankings.
KeepArchitecture, backend, ops, product-quality, and security review controlsThey have observable execution and trigger yield. This pilot cannot estimate incremental effect or justify removal.
KeepMandatory safety invariantsSparse incidents and absent bypasses can never authorize removing security, credential, release, deploy, or live-operation safeguards.
NarrowCurrent product claimControl analytics is implemented and validated for coverage/data-quality readback. GPUaaS has not yet validated causal or incremental control effectiveness.
RedesignGPUaaS commit associationMake task identity durable in normal commit/promotion flow so commit coverage does not depend mainly on message convention or broad path ownership.
InstrumentStructured outcomesLink incidents, rollbacks, reopens, corrective work, and superseding tasks before relying on same-file touches.
InstrumentControl frictionRecord attributable control start/resolution or review-wait facts; unavailable must remain distinct from zero.
InstrumentPreflight and security semanticsAdd stable control IDs/families and reviewed mandatory-invariant metadata rather than inferring them from names.
InstrumentEvidence-gate trigger semanticsDistinguish a gate that found a defect from one that merely recorded a passing artifact.
DeferOutcome-delta rankings and control-removal proposalsResume only after coverage and outcome-integrity gates are met; do not manufacture bypass cohorts for measurement.

Next Measurement Gate

The next GPUaaS effectiveness interpretation should require all of the following:

  1. commit-to-task and changed-file coverage at or above 80% for two consecutive 30-day windows;
  2. per-cohort control-state coverage at or above 80% for any cohort being interpreted;
  3. retained structured outcomes for incidents, rollbacks, reopens, and corrective/superseding work where those events occur;
  4. attributable friction samples for controls whose cost is being discussed;
  5. stable preflight and security-family metadata, including reviewed mandatory invariants;
  6. natural, explicitly authorized bypass/defer facts for non-mandatory controls when they occur, without weakening a control to create an experiment;
  7. nonzero, naturally occurring observed and bypassed cohorts, each meeting the configured minimum sample, before any observed-versus-bypassed outcome delta is interpreted; otherwise the comparison remains insufficient_sample;
  8. contemporaneous size/risk cohorts that preserve those separate observed and bypassed denominators.

Until then, Fairway should continue reporting coverage, missing instrumentation, execution yield, and uncertainty without ranking controls.

Evidence

The project-local raw outputs are retained outside version control at:

GPUasService/tmp-ux/fairway-control-effectiveness-pilot-2026-08-02/

Their SHA-256 digests, derivation notes, classification counts, and the bounded representative-control extracts used by this assessment are committed in evidence/fw-387/gpuaas-control-pilot-summary.json.