Divergence Review Workflow¶
A divergence-routing workflow — instantiates Correspondence Validation
Routes each already-detected old-versus-new mismatch through a standing triage that classifies it, logs the accepted exceptions, and records the disposition — so differences are governed, not quietly ignored.
A Divergence Review Workflow is the standing triage that takes a pile of mismatches already surfaced by a comparison — outputs, decisions, or side effects where new and old disagree — and gives each one a disposition and a home in the record. Its defining feature is where in the pipeline it sits: it begins after the differences are found and ends before anyone decides the migration's fate. It does not detect divergences and it does not pull the migration's go/no-go lever; it sorts. Each mismatch is sentenced into a fixed taxonomy — intended reform, harmless change, tolerable approximation, unresolved anomaly, unacceptable regression, safety violation — routed accordingly, and written down. Its whole reason to exist is to stop a replacement's differences from collapsing into an all-or-nothing argument, and to keep "this is an improvement" from being confused with "this is a regression."
Example¶
A state agency replaces the twenty-year-old ruleset that decides childcare-subsidy eligibility with a new, simplified one. Running both engines over roughly 80,000 historical applications produces about 3,400 cases where the new engine reaches a different determination than the old. Left as a lump, that number is useless — nobody can say whether the replacement is safe. The Divergence Review Workflow processes the 3,400 one at a time. Most are the intended effect of the reform (a broadened income band now admits families the old rule excluded) — classified as intended change and logged as accepted exceptions with the reform ticket as rationale. A few hundred are harmless (a rounding presentation difference in the award letter). A dozen are unresolved anomalies routed to a caseworker panel. And nine are unacceptable regressions — families with a disabled dependent who lost eligibility the statute still guarantees — sent straight back to engineering as blockers. The output is not a verdict on the migration; it is 3,400 sentenced, recorded dispositions that a later gate can act on with confidence.
How it works¶
- Intake, don't detect. The workflow receives mismatches from a comparison run; producing them is another mechanism's job.
- Classify against a fixed taxonomy. Every mismatch is assigned exactly one class, so improvement and regression can never share a bucket.
- Route by class. Accepted differences are logged; anomalies escalate to a named owner or review panel; regressions are sent back as blockers.
- Record every disposition. Each sentence — accepted, revised, escalated, rejected — plus its rationale is written to the durable record, so a later reader knows which divergences were blessed and why.
- Run as a standing cadence. As each new comparison batch arrives, the same triage runs again; the taxonomy and the register persist across rounds.
Tuning parameters¶
- Taxonomy granularity — how many divergence classes exist. Finer classes separate subtle cases but slow every review and invite miscategorization.
- Auto-accept threshold — which classes clear without human sign-off versus which escalate. Loose auto-accept clears volume fast but lets regressions slip through relabeled as "intended."
- Escalation authority — who may bless a regression as intentional: one owner, or a quorum panel. Higher authority resists rubber-stamping but adds latency.
- Exception expiry — whether a logged exception is permanent or must be revisited on a schedule, trading register stability against accumulating unreviewed debt.
- Batch cadence — per-commit, nightly, or per-release triage, trading freshness of the record against review overhead.
When it helps, and when it misleads¶
Its strength is that it refuses to let a replacement's differences stay undifferentiated: it forces every mismatch to be named as improvement, tolerable, or regression, which is exactly the distinction that keeps a migration honest. It is the mechanism that prevents compatibility theater — differences noticed but never governed.
Its failure mode is quiet rubber-stamping: when the auto-accept threshold is loose or reviewers are overloaded, the register becomes a dumping ground where real regressions are relabeled "accepted divergence" to clear the queue, and the taxonomy's discipline evaporates. The classic misuse is a Material Review Board[n1] that dispositions everything "use-as-is" under schedule pressure — the paperwork exists, but nothing is actually held back. The guarding discipline is to require a written rationale and a named owner for every accepted divergence, to give regressions and safety violations a class that cannot be auto-accepted, and to re-review the exception register on a cadence rather than letting it ossify.
How it implements the components¶
The classify-and-log core of the archetype — the components a routing workflow actually owns:
divergence_classification— its central act: each surfaced mismatch is sentenced into exactly one class of the taxonomy, keeping improvement and regression from being confused.exception_register— every accepted or known-and-tolerated divergence is logged with its rationale and owner, so it is tracked forever rather than forgotten at cutover.transition_record— each disposition, escalation, and rejection is written to the durable record that later users consult to learn where equivalence was granted and where it was withheld.
The workflow dispositions each mismatch, but the authority to let the whole migration proceed, pause, or roll back — acceptance_or_limitation_decision, held behind a migration_guardrail — belongs to Migration Acceptance Test, its nearest twin; and running the behavior_comparison over a comparison_case_set that surfaces the mismatches in the first place is Protocol Conformance Test and shadow runs.
Related¶
- Instantiates: Correspondence Validation — it is the governance step that turns a list of raw differences into classified, recorded dispositions.
- Consumes: Shadow Run or Parallel Run and Protocol Conformance Test supply the detected mismatches it routes.
- Sibling mechanisms: Migration Acceptance Test · Model-Limit Validation · Protocol Conformance Test · Backward Compatibility Test · Golden Case Benchmark · Regression Test Suite · Shadow Run or Parallel Run
Editorial Notes¶
Form Classification¶
Form family: Decision, Gate & Allocation
Rationale: For each detected mismatch, the mechanism makes a bounded classification and disposition—accepted difference, anomaly, regression, or escalation—and routes it accordingly.
Nearest alternative: Protocol, Workflow & Routine — Intake and recording form a repeatable workflow, but its defining output is the case-specific disposition that determines the mismatch's route.
Review outcome: Adjudicated after independent review; medium confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Quality and configuration engineering cohered standing review boards that classify each nonconformance and record use-as-is, rework, repair, scrap, or rejection.
Related originating lineages:
- Organizational & Management Science — Governance workflow practice generalized exception triage, ownership, and recorded disposition beyond manufactured material.
Review resolution: Both current reviews place divergence_review_workflow primarily in engineering_design; the reconciled classification retains only lineages that materially shaped the mechanism and keeps breadth of origin separate from reach.
Attribution caveat: The generic old-versus-new mismatch workflow directly generalizes the material-review-board lineage.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The workflow deliberately separates sorting differences from deciding the migration. Keeping the triage distinct from the acceptance gate is what lets an agency improve its classification — tighten the taxonomy, add reviewers — without re-litigating cutover, and lets the gate raise or lower its bar without disturbing how individual mismatches are sentenced.
[n1] A Material Review Board (MRB) is the standing quality-control body, standard in aerospace and regulated manufacturing, that dispositions nonconforming material into use-as-is, rework, repair, scrap, or reject. It is the physical-goods analogue of routing each divergence to a disposition — and its well-known failure, waving items through as "use-as-is" under deadline pressure, is exactly the rubber-stamp risk here. ↩