OpenXR Policy Recompute Uses Latest Registry While Adjudication Rows Stay Versioned - Quest Fix

Your adjudication close captured reason_code_version = 12 and passed preview checks. After the team publishes registry v13, nightly policy recompute recomputes eligibility and waiver posture using the latest effect tables. Suddenly rows look non-compliant even though operators never reopened them; dashboards spike policy drift without any new closes.

In 2026, Quest-heavy adjudication lanes hit this more often because taxonomy cadence accelerated while batch jobs still defaulted to always-latest interpretability.

That mismatch is not reviewer noise; it is a version join bug between immutable adjudication snapshots and mutable registry readers.

Problem

Typical symptoms:

historical disputes flip promotion labels after a registry publish even though close payloads never changed
recompute logs show registry_version_read newer than reason_code_version stored on the row
signer packets disagree with automated exception summaries generated post-recompute
migration stabilization phases report rising drift counts despite flat reopen volume
QA notices identical evidence packets yielding different policy classes between preview UI and batch exports

If any batch path resolves codes through “whatever registry is live today,” audit replay breaks whenever taxonomy evolves.

Direct answer

Bind policy recompute to each adjudication row’s frozen semantic tuple: read registry artifacts versioned to reason_code_version (plus remap lineage when applicable), or explicitly migrate rows forward with logged remap steps—never silently reinterpret legacy closes against the newest registry.

Root cause summary

Latest-registry reader — batch SQL or worker loads current taxonomy tables without joining stored version IDs.
Partial migrations — registry bumped while adjudication rows remain on older versions during phased rollout.
Dual-effect definitions — preview UI reads snapshotted policies; batch jobs read live tables.
Missing lineage on recompute — jobs ignore migration_window_id, mapping_rule_id, and policy_hash captured at close.
Cache keyed wrong — memoization caches policy outcomes by dispute ID only, not registry tuple.

Fastest safe fix path

Identify one disputed production row and compare reason_code_version on the adjudication record against registry_version_used in recompute logs.
Freeze risky batch lanes until reads use version-targeted registry snapshots or explicit remap joins.
Patch recompute query paths to join registry_release_id compatible with each row’s version tuple.
Replay a sample set before next registry publication.
Add a CI assertion that batch preview matches preview UI for frozen fixtures.

Step-by-step fix

Step 1: Instrument recompute inputs

Emit structured fields on every policy recompute execution:

adjudication_reason_code_version
registry_schema_hash_read
remap_chain_id if reopen migrated codes
policy_effect_engine_build

If registry_schema_hash_read cannot be tied to the adjudication version, stop and fix reads before arguing about reviewer quality.

Verification checkpoint: sampled logs show matching version tuples between adjudication export and batch worker.

Step 2: Replace latest-table joins with versioned reads

Patterns that usually fail:

SELECT * FROM reason_code_effects WHERE code = ? without version predicate
denormalized JSON blobs rebuilt from current CMS draft instead of published artifact at close time

Preferred patterns:

materialized registry bundles tagged release_id aligned with semver bump
loader reads effects@v12 when row declares version 12

Verification checkpoint: forcing registry v13 publish does not alter outputs for rows pinned at v12 unless migration rules explicitly remap them.

Step 3: Align preview UI and batch semantics

Ensure preview surfaces use the same resolver module as batch jobs. If preview snapshots policy from adjudication payload while batch recomputes from tables, split stacks guarantee drift.

Verification checkpoint: frozen dispute fixtures produce identical policy classes across preview modal and nightly export.

Step 4: Handle stabilization-phase mixed inventories

During migration you may legally hold:

rows closed under v12
newly closed rows under v13

Recompute must branch:

pinned path for untouched legacy inventory
forward-migrate path only when compatibility rules and approvals exist

Never collapse both into one latest-registry sweep.

Verification checkpoint: migration dashboards segment counts by version cohort without unexplained cross-cohort flips.

Step 5: Fix caches and idempotency keys

If caches exist:

key by (dispute_id, reason_code_version, registry_bundle_hash)
invalidate when intentional migration advances a row’s version after logged remap

Verification checkpoint: rerunning identical recompute twice yields bitwise-stable summaries.

Step 6: Governance gate before registry releases

Block registry publishes unless:

regression harness proves historical cohort outputs unchanged or
companion migration patch explicitly targets affected IDs with lineage

Verification checkpoint: release checklist references artifact hashes consumed by batch workers.

Verification checklist

[ ] Batch logs prove registry reads tied to adjudication version or authorized remap chain.
[ ] Preview UI and batch exports match on frozen regression disputes spanning at least two registry versions.
[ ] No unexplained policy-class motion when publishing registry docs-only releases without schema migration.
[ ] Idempotent recompute reruns produce identical outputs for stable inputs.
[ ] Signer-ready packets reconcile with automated summaries after recompute fixes ship.

Alternative fixes and prevention

Shadow recompute mode — compute both pinned and latest interpretations during migration; surface deltas instead of silently overwriting labels until governance accepts migration scope.
Read-through overlay — temporary overlay tables mapping legacy codes for dual-window teams; retire after stabilization with expiry dates.
Append-only adjudication amendments — when policy must change retroactively, issue explicit amendment rows rather than reinterpretation through newest registry alone.
Quarterly registry readiness drill — rehearse publish plus batch replay with synthetic fixtures before peak release weeks.

FAQ

Should recompute ever reinterpret legacy closes with the newest registry

Only through an explicit migration program that writes new lineage and approvals. Silent reinterpretation breaks audit defensibility.

Do doc-only registry edits require code changes

If documentation changes alter operator-facing semantics tied to automation, treat them as schema-touching until replay proves otherwise.

What if we cannot store full registry snapshots per version

Store effect_bundle_hash per release plus retrieval URL or artifact pointer; hashes must resolve reproducibly for compliance reviews.

Escalation criteria

Escalate to platform governance when:

batch outputs diverge from signer-facing summaries after routine registry publishes
multiple cohorts show unexplained label churn within one release train
migration stabilization criteria cannot be met because recompute drift masks real quality signals

Bookmark this fix beside your registry publish checklist and batch worker deployment notes so version skew regressions are caught before promotion gates widen.