Fairway Common-Path Pilot

Date: 2026-07-11

Task: FW-295

Decision

Keep the common fairway work lifecycle and continue the intent-to-diff classifier as advisory measurement. Do not promote unexplained reversible-work deviations into a blocking work close gate from this pilot.

The common cohort completed materially faster with fewer task-state CLI transitions and no loss of passing evidence, active checkpoints, sessions, or consequential review enforcement. The pilot does not yet contain the precision, false-positive, or rollback-confidence data required to justify a new blocking default. Review and notification records per task also increased.

Boundary

This assessment changes no review, merge, release, deploy, credential, public, live-operation, dashboard, provider-send, or task-state authority. It does not use provider transcripts as measurement or provenance. The Fairway database, Git commits, evidence, reviews, sessions, checkpoints, and lifecycle rows are the source facts.

Cohorts

The historical baseline is the four completed small-team tooling tasks FW-274, FW-278, FW-279, and FW-280. They predate the completed common path implementation.

The common-path cohort is the nine v0.1.12 train tasks completed through the compact lifecycle: FW-303, FW-293, FW-306, FW-294, FW-299, FW-297, FW-298, FW-300, and FW-301.

This is an observational pilot, not a randomized comparison. Task complexity, review timing, and provider availability differ. Results support bounded product decisions but not a claim that the command surface alone caused every timing improvement.

Results

Measure	Historical baseline	Common path	Interpretation
Tasks	4	9	Cohorts above
Task-state CLI transitions per task	3.75	3.00	Common close/start removed extra task-state transitions
Checkpoint rows per task	3.25	1.22	Atomic start produced one clear active checkpoint
Tasks with active checkpoint	4/4	9/9	No visibility regression
Tasks with provider session	4/4	9/9	No attachment regression
Start to first evidence, average	2,222.1 s	581.7 s	Authoring/proof proxy improved
Active to done, average	6,194.2 s	713.5 s	Elapsed lifecycle proxy improved
First evidence to done, average	3,972.1 s	131.8 s	Closeout delay materially lower
Last validation to done, average	3,965.0 s	131.8 s	Closeout handback materially lower
Blocked seconds	6,559	0	No common-cohort blocked interval
Review-wait seconds projected	0	0	Reviews completed inside the acknowledgement window
Tasks with passing evidence	4/4	9/9	Evidence completeness preserved
Review records per task	3.25	4.00	Review ceremony increased
Changes/reject findings	0	6	Two common tasks had three-domain changes requests
Notifications per task	1.75	3.33	Coordination overhead increased
Reopen/retry count	0	0	No observed regression
Accepted material decisions	0/0	2/2	No hollow or insufficient decision in measured rows
Stale memory findings	0	0	Resume packet remained fresh
Promotion-debt findings	0	0	No unresolved promotion row
Rollback-referencing evidence rows	0	1	Too sparse and task-dependent for comparison

The six review findings came from FW-303 and FW-293; both were corrected before commit. This is useful defect discovery, not a reason to require the same review weight for every reversible task. The other seven common tasks had approval-only review records.

Measurement Limits

Fairway currently records durable primitives, not a complete command-invocation ledger. Exact manual commands and human/provider authoring effort therefore cannot be reconstructed without using transcripts, which this pilot rejects. Task-state transitions and start-to-first-evidence time are bounded proxies.

Intent-to-diff classification has not yet run across a labeled set, so the pilot has no measured precision, recall, or false-positive rate for deviations. Rollback confidence is also not a normalized evidence field and varies by task risk. These gaps prevent responsible promotion to a blocking reversible-work gate.

Reproducible Readback

The assessment used:

fairway --json delivery report --since 336h
fairway --json memory reconcile --older-than 24h
fairway --json decision list FW-303
fairway --json decision list FW-293

SQLite read-only queries grouped the named task IDs over task_state_history, task_evidence, task_checkpoints, agent_sessions, task_reviews, task_notifications, task_decisions, and task_decision_assessments. No row was inserted, updated, deleted, or backfilled for the analysis.

The exact cohort queries are preserved in fairway-common-path-pilot-2026-07-11.sql:

sqlite3 .fairway/state.db < docs/assessment/fairway-common-path-pilot-2026-07-11.sql

Task-level blocked/review-wait/reopen and defect-source rows came from the recorded delivery report. The exact cohort selection is reproducible with:

fairway --json delivery report --since 336h |
  jq '[.rows[] | select(.task_id as $id | ["FW-274","FW-278","FW-279","FW-280","FW-303","FW-293","FW-306","FW-294","FW-299","FW-297","FW-298","FW-300","FW-301"] | index($id))]'

Follow-Through

1. Implement FW-304 as an advisory/report-only intent-to-diff classifier for reversible work. Preserve existing blocking consequential boundaries.
2. Record classifier findings, accepted explanations, and false positives over a labeled pilot before considering promotion.
3. Keep review routing risk-scaled. Group approval-only reversible reviews when policy allows; retain direct review where a real defect class or consequential boundary justifies it.
4. Do not add a default rollback-confidence field or command ledger until a separate measured need establishes that the extra authoring cost improves safety or resume quality.

Promotion criteria for a later decision are: demonstrated classifier precision, bounded false positives, lower common-path effort, no weaker consequential boundary enforcement, and evidence that blocking catches material defects that advisory reporting would otherwise miss.