Corrective Action Report — Parallel PRD coherence gap (Compass Owner timing + third Compass sense)

← Back to Corrective Actions 2026-05-11

Corrective Action Report — Parallel PRD coherence gap (Compass Owner timing + third Compass sense)

CAR ID: CAR-2026-05-11-C Filed: 2026-05-11 Filer: Q (instruction-optimizer / Quality System) Companion audit: docs/quality/audits/2026-05-11-abs-company-compass-launch-audit.md § 8 CAR-C Status: Proposed (pending V acceptance) Authority: Q proactive audit authority granted 2026-04-12 (QMS establishment)

1. Incident

Two parallel PRDs operated simultaneously during the Abs Company AI-Native Shift launch workstream:

  1. Abs Company AI-Native Shift Launch PRD (docs/plans/abs-company-ai-native-shift-launch-prd.md, V-led, 26 FRs)
  2. Compass End-State PRD (docs/plans/compass-end-state-prd.md, Architect-led, 35 execution steps, 18 success criteria)

Each PRD named the other as canonical for blocked work (Abs Co PRD's FR-1/12/20 reference the V2 platform; Compass End-State PRD blocks FR-C3 and per-client migration protocol on "abs-company 5P deliverable").

The two PRDs disagreed at the FR level on Compass Owner Agent timing:

  • Abs Co PRD FR-3 / FR-16 / FR-24: Compass Owner Agent created in Foundation layer (no upstream dependency), ratified and pushing first content before kickoff, operating post-kickoff mode.
  • Compass End-State PRD v0.2 S17 / decisions 1+4 / execution steps 28-30: Compass Owner is named LATE — one of the LAST performance gates, after platform behavior validates P2.5 criteria. Architect leads BUILD end-to-end through Step 27.

The two PRDs also did not resolve the third-Compass-sense scope question (bias-check caveat C13). The shipped Compass surface is a relationship-narrative + project-management-communication surface that is neither a walkthrough rendering mode nor an asset/artifact library mode (the two modes the handoff prompt's Conflict 2 reconciliation addressed).

The conflicts were resolved by neither PRD. Default behavior (no Compass Owner agent created) aligned with End-State PRD's posture by inaction. The audit's ls .claude/agents/ | grep -iE 'compass|owner' at 2026-05-11 03:52Z returns no matches.

2. Timeline

Time (CT) Event
2026-05-10 ~12:00 Compass End-State 5P plan drafted (canvas-end-state-5p-plan.md, commit 9ef54c0d4 initial)
2026-05-10 ~13:00 Compass End-State PRD generated from 5P (commit c08c81a0a)
2026-05-10 ~14:00 Abs Co 5P plan drafted (commit e84788956)
2026-05-10 ~14:30 Handoff prompt for parallel execution committed (e79eaf5df)
2026-05-10 ~14:45 Handoff addendum identifies Conflict 1 (Compass Owner timing) and Conflict 2 (Compass scope) (commit 8484296c2)
2026-05-10 ~15:00 Abs Co PRD generated from 5P (commit e94b59b1c)
2026-05-10 ~16:00 Q bias-check audits on both 5P + PRD (caveats C1-C19 surfaced)
2026-05-10 ~17:00 Convergence map between PRDs (7a6aced7f) — identifies but does not resolve conflicts
2026-05-10 ~18:00 Compass End-State PRD v0.2 revision incorporates Chris's four decisions, including S17 Compass Owner LATE gate (6ae882399)
2026-05-10 22:17 Late-stage architectural reframe makes the third-Compass-sense question concrete (see CAR-A)
2026-05-11 03:52 Audit captures: no Compass Owner agent exists; methodology-content-as-surface assumption from Abs Co PRD did not hold

3. Root cause analysis

3.1 Primary root cause: no single reconciliation document at FR-level fidelity

When two PRDs run in parallel naming each other as dependencies, the PRDs need a single reconciliation document at the same fidelity as the individual PRDs to resolve cross-references and conflicts. The artifacts that existed were insufficient:

  • Handoff prompt addendum (commit 8484296c2) names two conflicts but resolves them at framing level, not at FR level. The Abs Co PRD's FR-3 was not updated to reflect the reconciliation's outcome.
  • Bias-check audits identify caveats (C1-C19) but do not gate either PRD's FR commitments on resolving them. Caveat C13 (third Compass sense) is explicitly flagged as unresolved and the PRDs ship anyway.
  • Convergence map (commit 7a6aced7f) describes the relationship between the PRDs descriptively, not resolutorily.

No artifact says: "Compass Owner Agent FR-3 in Abs Co PRD is superseded by S17 in End-State PRD; Abs Co PRD FR-3 is hereby deprecated." The deprecation, if intended, was implicit.

3.2 Contributing causes

3.2.1 Same agent authored both PRDs in the same window. V wrote both PRDs in the same multi-hour window. Both PRDs reflect V's intent at write-time; differences between them reflect V's evolving intent during the window. No cross-PRD reconciliation pass was performed before either PRD was committed.

3.2.2 Bias-check audits are PRD-scoped. Q's bias-check audit protocol treats each PRD separately. Caveat C13 is the only caveat that crosses PRDs; it identifies the third-Compass-sense gap but does not gate either PRD's FR commitments. The bias-check protocol does not have a cross-PRD pass.

3.2.3 Compass End-State PRD v0.2 revision was post-bias-check. Chris's four decisions (including the LATE Compass Owner gate in S17) were applied to the End-State PRD AFTER Q's bias-check completed. The Abs Co PRD was not synchronously updated. Two PRDs went out of sync mid-day.

3.2.4 Cross-references in PRDs are unidirectional in form, bidirectional in fact. Each PRD's "Dependencies & Risks" section names items in the other PRD. The PRD format does not require the named items to point to a specific resolved-decision-line in the other PRD. When the named item changes (S17 introduced in End-State PRD v0.2), the citing PRD (Abs Co's FR-3) does not get notified.

4. Immediate fix

V's recommendation 5 in the companion audit (§ 9): explicitly accept or reject the default (no Compass Owner agent). Two paths:

  • Accept default (LATE gate posture wins): Update Abs Co PRD FR-3 / FR-16 / FR-24 to align with End-State S17. Compass Owner created late; pre-kickoff work specified to be performed inline by Architect + Showcase + Chris is honored as the operational reality.
  • Reject default: Spawn Aegis to create the Compass Owner agent pre-kickoff. Update End-State S17 to specify the agent already exists. Compass Owner ratifies and pushes first content per Abs Co FR-16.

Either path resolves the conflict explicitly. The current state — implicit resolution by inaction — leaves the spec corpus contradicting itself.

5. Permanent prevention

5.1 New process: parallel PRD reconciliation document

When two PRD/5P pairs run in parallel and name each other as dependencies, require a single reconciliation document at the same fidelity as the individual PRDs:

Path: docs/plans/{date}-cross-prd-reconciliation-{topic}.md Content: Every cross-reference and conflict resolved at FR-level. Each FR in either PRD that names the other PRD as a dependency must point to a specific resolved-decision-line in the reconciliation document. Authority: Both PRDs are gated on the reconciliation document's existence. Q audits.

5.2 PRD template update

Hone updates the PRD template generator (/prd-generate command) to include:

Cross-PRD reconciliation block (required when this PRD names another PRD as a dependency). Each cross-reference cites a specific decision-line in the reconciliation doc, or an explicit "this PRD operates independently of {other PRD}" declaration.

5.3 Bias-check protocol extension

Q's bias-check protocol gains a fifth box (companion to CAR-A's Box 4):

Box 5 — NOT cross-PRD spec drift. When this PRD's FRs reference another parallel PRD, do they cite specific resolved decisions in a reconciliation document? Have all conflicts named in handoff prompts and bias-check caveats been resolved at FR-level in a reconciliation document? If no reconciliation document exists, the PRD is incomplete.

5.4 Echo pattern memory

Echo records: parallel-PRD coherence gap. First documented instance. No prior occurrences in this form. Threshold for mandatory enforcement: one more recurrence within 90 days.

6. Process register impact

New process to add to docs/quality/process-register.md:

Process Owner Risk Tier Verification Frequency Method
Parallel PRD reconciliation Q Tier 2 (High) Per-occurrence When two parallel PRDs name each other as dependencies, verify reconciliation document exists at FR-level fidelity before either PRD's FRs ship.

7. Recurrence prevention check

The Abs Company / Compass parallel session is the first formally-named instance of this pattern. No prior occurrences to verify against. If a second instance occurs without reconciliation document being created, the prevention failed and CAR escalates.

A related precedent — the Apr 22 mass rename → Apr 25 walk-back — involved cross-component conflation (canvas / portal / compass merged into a single substrate decision). That incident had only one spec document, so it is not a parallel-PRD pattern; but it shares the "merge / unify decisions made before reconciling" failure mode. The Compass End-State 5P P6 Box 3 explicitly notes Apr 25 walk-back as the closest-analog institutional lesson.

8. References

  • Companion audit: docs/quality/audits/2026-05-11-abs-company-compass-launch-audit.md § 4 + § 8 CAR-C
  • Abs Co PRD: docs/plans/abs-company-ai-native-shift-launch-prd.md
  • Compass End-State PRD: docs/plans/compass-end-state-prd.md
  • Handoff prompt + addendum: e79eaf5df, 8484296c2
  • Convergence map: 7a6aced7f "docs(plans): map convergence between Compass End-State PRD and Abs Company AI-Native Shift Launch PRD"
  • Bias-check audits: docs/quality/audits/2026-05-10-abs-company-5p-bias-check.md + docs/quality/audits/2026-05-10-canvas-5p-bias-check.md (caveat C13 specifically)
  • Apr 25 walk-back precedent: docs/quality/cars/2026-04-25-corrective-dewey-registrar-activation.md

Filed by Q under proactive audit authority. Awaiting V acceptance for permanent-prevention deployment.