NODEDC_MISSION_CORE/docs/adr/0032-ravnoves00-methodology-correction-before-transfer.md

ADR 0032 — RAVNOVES00 methodology correction before transfer

Date: 2026-07-28

Status: accepted for implementation

Context

E37–E40 produced useful, immutable source-scoped engineering evidence, but the architecture report was written before the source implementation and artifacts were audited together. E41 reproduced the exact data lineage and found four methodological limits:

1. E37 contains 484 engineering-reviewed labels and 2 human exception labels; it explicitly declares independent_ground_truth=false.
2. The nominal 340 development / 146 validation split is not group-independent. Development and validation overlap by 14 exact frames, 38 track identities, 58 fixed 50-frame time blocks and 59 connected whole-track-or-scene groups.
3. The E40 worker package co-located prediction inputs with the visible acceptance rows and split assignments.
4. E38–E40 therefore measure performance against a historically evaluated, source-scoped engineering contract. They are not blind accuracy gates.

This does not invalidate their physical predictions, closed accounting or conservative authority. It changes the claim that those artifacts are allowed to support.

Decision

1. Historical E37–E40 artifacts remain immutable evidence. Their metrics are retained, but “sealed”, “blind” and independent-accuracy language is removed from current status.
2. The current 146-item slice is named historical evaluated visible validation. It may compare immutable methods against the source-scoped engineering contract; it may not prove independent perception accuracy or cross-route generalization.
3. Prediction and evaluation are physically separated. The E41 predictor package and result contain no acceptance rows, references, split assignments, severity or scoring state. The visible evaluator joins predictions to E37 only after inference.
4. Runtime identity records the environment that actually executed a result. A local run records the repository lock, Python and NumPy identities and is not represented as a Worker 006 container run.
5. Structural invariants are exercised independently of the visible labels. E42 verifies predictor identity, row-order and chunk-boundary invariance plus PointSlab row-order invariance and rigid-coordinate equivariance.
6. Native per-stage telemetry is a compute contract. Stages publish started/completed/failed, exact source/method/run identity, duration and accounting through an injected sink. A JSONL evidence sink and an already-connected MQTT sink are admitted; no network transport starts implicitly.
7. Current storage work starts with content-addressed references and exact-content deduplication. E44 measures 1.680146× amplification across the admitted E30–E40 package/result roots. It does not authorize a storage-format or database migration.
8. The later same-K1 transfer capture is preregistered before collection. E43 freezes minimum streams, route/control-bridge duration, connected scene/track/time partitioning, independent human review and truth reveal only after frozen prediction.
9. Product windows, page anatomy and new design work are outside this stabilization decision and remain a separate owner review.

Current immutable evidence

• methodology audit: e41-methodology-audit-703dac176f20843ad6fbb9a1060b199942fba05e3bb34d3e2264ee87c7bfdd96;
• truth-free predictor package: e41-predictor-package-fb35f42698013c63c6d417fd3b26986e9d3b89d12dce138bab18fe4203d593ad;
• truth-free prediction result: e41-predictions-adcc1671de44646145a396d5815e522d6b4257d6ead9ace995761751655b3aac;
• visible engineering evaluation: e41-visible-evaluation-57aa6c8569e8339630406e3c88cd539df13917fb08c1af57f0baae5cdd1069d2;
• metamorphic suite: e42-metamorphic-suite-c358204e0b9e10eea5871e90b7c4fbe2c132f1f6d87074b2c5ffec18eaeebb40;
• preregistered future-capture protocol: e43-future-capture-protocol-28f091b9648daffce988d44c183e21f56d77988061630de934f8003fb13701d8;
• data-amplification audit: e44-data-amplification-89791978894ec785009e5b76c9325de4d9e910237af57476c9331eb47a8ccca4.

Consequences

• RAVNOVES00 remains the active engineering reference source, but R0 is reopened as an independent-truth and leakage-free-partition gate.
• E40 does not pass R1: the visible evaluation is 84.2466% for presence, 84.2466% for geometry association and 94.5205% for freshness, with 11 high-severity failures, complete accounting and zero false-free claims.
• Further threshold or model tuning against the visible 146 items cannot close the blind gate. It may only create another explicitly visible engineering comparison.
• Work that does not require a new recording can proceed now: methodology enforcement, predictor/evaluator separation, invariants, telemetry, exact-content deduplication design and preregistration.
• A later new-route recording can close an independent transfer question only if the E43 protocol passes before labels are revealed. Until that capture exists, cross-route quality remains unknown rather than failed or inferred.