Proven — Promoted from Emerging May 8, 2026 (CIO promotion authority per methodology-audit-policy-updates-2026-03-16.md). Trial-application evidence: diagnostic operated at origin (Apr 26 C-axis incident); branch-or-anchor rule shipped to two surfaces (methodology-24 + CT v2.3) without subsequent silent-extension recurrence; Architect May 4 architectural-soundness review identified parallel-authoring instance at code layer (legacy/refactored boundary_enforcer.py coexistence) using pattern’s framing. Three independent surface validations. Originally identified April 26, 2026 (CXO + PPM rubric C-axis reconciliation incident); filed Emerging Apr 27 under CIO self-approval authority + PM concurrence on slot allocation. Promotion analysis at dev/active/cio-pattern-promotion-analysis-2026-05-08.md.
Portable — Any team maintaining canonical specifications (rubrics, schemas, contracts, vocabulary, configuration) will encounter this when extending those specs in parallel without anchoring. The pattern operates at the specification layer, which means it cuts across code, documentation, evaluation instruments, and process artifacts.
When a canonical reference (rubric, schema, term, principle, slot reservation) is extended in parallel by two or more authors, each acting in good faith and producing individually correct work, the extensions can diverge silently at the semantic layer while appearing convergent at the surface layer. The failure becomes visible only when the parallel artifacts are applied to the same case and produce divergent results.
This pattern is a sibling sub-pattern of Pattern-062 (Assembly Assumption) at the specification layer, distinct from Pattern-064 (Extension Without Integration) which operates at the integration-contract layer. Both are sub-patterns of 062; both involve individually-correct components composing into incomplete or wrong outcomes; the difference is in the failure mechanism.
| Pattern | Layer | Failure mechanism |
|---|---|---|
| 062 (parent) | General | Individually correct components, unverified composition |
| 063 (this) | Specification | Two parallel extensions of shared semantics drift; verdicts converge while methodology diverges |
| 064 (Architect, planned) | Integration contract | One extension passes own tests but breaks on realistic composition |
Pattern-063 is about parallel authoring producing semantic drift. Pattern-064 is about single-path extension without integration verification. Different mechanisms, both children of 062.
Two authors extend the same canonical reference at roughly the same time, working independently:
Canonical doc v1 ─┬─→ Author A: extends with adaptation X
└─→ Author B: extends with adaptation Y
Both extensions: individually correct.
Both extensions: use shared label/identifier from v1.
Surface: converged (same label, similar verdict on test cases).
Semantic: diverged (different criteria, different consequences).
The failure becomes detectable only when the parallel artifacts are applied to the same case and produce different scores, decisions, or actions — by which point both have propagated to downstream consumers.
Three forces converge:
Canonical references are extension surfaces by design. Rubrics, schemas, and decision frameworks are written to be extended for local needs. The act of extension is encouraged; the discipline of how to extend is rarely formalized.
Parallel work is the operating mode. In multi-agent or multi-author systems, the assumption is that authors do not need to coordinate at every extension. Each author reads the canonical doc, drafts an adaptation, ships it. No structural moment forces consultation between authors.
Surface convergence masks semantic divergence. When two parallel extensions produce the same verdict on a test case (both score “PASS”, both call a query “high-risk”), the agreement reads as validation. The validation is illusory: the two methods may agree in the trivial cases while diverging arbitrarily in the non-trivial ones.
“If I asked the two authors to score each other’s work using the other’s rubric, would they get the same answer?”
If the answer is “I’m not sure,” the parallel-authoring drift pattern is operating. The agreement on the surface tells you nothing; the divergence is at the semantic layer where the two authors haven’t yet checked each other’s work.
A second diagnostic phrasing (cross-cuts to Pattern-045 + Pattern-062): “What would have to be true for these to be wrong in the same direction?” When two divergent methods produce the same answer, the answer is robust to method variation in the trivial cases — which tells you nothing about the non-trivial cases.
Two scoring instruments shared the canonical reference’s letter “C” with materially different criteria:
| Rubric | Author | Date | C-axis label | C=2 criterion | C=3 criterion |
|---|---|---|---|---|---|
| Phase E rubric | Lead Dev | Apr 23 | Clarity | “States decline but no reason” | “Offers user a constructive path forward” |
| Colleague Test v2 | CXO (Code) | Apr 25 | Context | “Generic LLM competence” | “Project-context injection visible” |
Both rubrics were authored in good faith. Both were individually correct. Both extended the same canonical predecessor (Colleague Test v1) without explicit anchoring or branching. The drift surfaced Apr 26 when the same Phase E transcripts were scored under both rubrics — verdicts converged at PASS, but C-axis scores diverged systematically.
Verdict-convergence as the masking signal: the same transcripts received PASS under both rubrics. The agreement on output level masked the divergence on methodology level. Without the cross-application that surfaced the divergence, both rubrics could have continued diverging through subsequent applications, with quality-threshold decisions silently depending on which rubric an author happened to consult.
When extending a canonical reference (rubric, schema, term, principle, slot allocation), the author must explicitly choose anchor or branch:
Silent extension — using the canonical’s label while changing the criteria — is the failure mode this rule prevents.
For the canonical example above (CT v2 C=Context vs Phase E C=Clarity):
Both options resolve the drift. Silent extension cannot.
The rule lives in two complementary surfaces (per CXO Apr 26 belt-and-suspenders concurrence):
Two surfaces because the populations of authors don’t fully overlap.
grep for similar axis labels in adjacent rubrics. Cheap, catches accidental collisions.These are useful infrastructure but secondary to the underlying discipline. The discipline is: do not extend a canonical reference silently.
| Don’t Do This | Why | Do This Instead |
|---|---|---|
| Extend a canonical’s label with new criteria silently | Surface convergence masks semantic divergence | Branch with explicit rename |
| Defer reconciliation to a “v2.x” patch when drift is noticed | Silent drift compounds during the deferral window | Treat the moment of noticing as the discipline issue; reconcile immediately |
| Trust verdict-convergence as validation | Agreement on outputs ≠ agreement on methodology | Apply the diagnostic question: would each author score the other’s work the same? |
| Reserve catalog slots informally in handoffs | Soft reservations fail under parallel authoring (this catalog itself, Apr 26 Pattern-063 slot conflict, is a meta-instance) | Use formal slot reservation with visible claim, or take the next available number when you formalize |
| Add only a registry/version-stamp without the underlying rule | Diagnostic aids don’t prevent the failure mode | Implement the branch-or-anchor rule first; add registry/stamps as defense in depth |
PPM noticed during Phase E scoring exchange that PPM (using CT v2) and CXO (using Phase E draft rubric) had diverged on the C-axis with different criteria but converging verdicts. Filed memo-ppm-to-cxo-lead-cio-cc-pm-pa-arch-exec-rubric-c-axis-reconciliation-2026-04-26.md framing the drift as a discipline issue per PM’s Apr 26 directive on terminology drift.
CIO responded with memo-cio-to-ppm-cc-cxo-lead-pm-pa-arch-exec-rubric-drift-methodology-2026-04-26.md naming the pattern as “Parallel-Authoring Drift” and proposing it as Pattern-062 at the methodology layer. CXO concurred on substance + belt-and-suspenders rule embedding.
Predecessor Architect’s handoff (Apr 25) had reserved Pattern-063 for a related but distinct sub-pattern of 062, “Extension Without Integration” (single-path extension, integration-contract layer). Architect filed a slot-conflict memo Apr 26 noting both patterns were legitimate sub-instances of 062 and recommending CIO take Pattern-063 (fully drafted) while predecessor’s claim shifts to Pattern-064 (to be formalized). PM concurred Apr 27.
The slot conflict was itself a meta-instance of Pattern-063 at the catalog-management layer: parallel-authoring drift around the catalog’s next-available slot reservation.
Filed under CIO self-approval authority per methodology-audit-policy-updates-2026-03-16.md with PM concurrence on slot allocation.
mailboxes/cio/sent/memo-cio-to-ppm-cc-cxo-lead-pm-pa-arch-exec-rubric-drift-methodology-2026-04-26.mdmailboxes/cio/read/memo-ppm-to-cxo-lead-cio-cc-pm-pa-arch-exec-rubric-c-axis-reconciliation-2026-04-26.mdmailboxes/cio/read/memo-cxo-to-cio-cc-ppm-lead-pm-pa-arch-exec-pattern-063-and-rule-embedding-2026-04-26.mdmailboxes/cio/read/memo-arch-to-cio-pm-cc-cxo-lead-ppm-pa-exec-pattern-063-slot-conflict-2026-04-26.mdmailboxes/cio/sent/memo-cio-to-arch-cxo-cc-pm-ppm-lead-pa-exec-pattern-063-slot-and-rule-2026-04-27.mdcreate-omnibus skill Step 7 (canonical-term verification) is the analog at the omnibus-synthesis layer of what branch-or-anchor is at the rubric/specification layer.dev/2026/04/17/methodology-audit-2026-04-17.md): the audit that formalized “Audit the Composition” as the Flywheel’s 5th practice and prefigured this pattern’s emergence.Pattern created: April 27, 2026 Origin: Apr 26 Phase E rubric C-axis reconciliation incident Author: CIO, with PPM + CXO + Architect cross-input Status: Emerging (CIO self-approval, PM concurrence on slot) Promotion criterion: one full cycle of trial application surfacing whether the diagnostic question and branch-or-anchor rule produce reliable catches in practice. Phase F+ scoring is the natural test environment.