# Evaluation provenance

A result records an evidence-bounded outcome, not a developer's mastery. A delivery record states what the system confirmed or withheld, not whether the guidance was correct. Keep these dimensions separate in research and operational exports.

## Review provenance

Join a result to its review, practice version, item version and developer account. Review practice records retain the complete frozen method, including every model slot and epoch. Execution jobs and script usage describe attempts. They can expire independently of product history. Usage retains its review reference after job deletion.

Historical unknown proof remains evidence version 0. A missing field is not replaced with today's definition or a fabricated quote. A generated digest does not reassess the original work. Frozen empirical records retain their schema and criteria. Do not translate a positive behavior label into full-standard conformance without justified equivalence.

## Delivery evidence

Join feedback through its basis to the exact supporting results and problems. Copies retain physical appearances, write fences and provider receipts. Decisions and developer responses retain their own history. Resolution and usefulness are independent dimensions.

A started write is not proof of visibility. Unconfirmed records cannot be counted as delivered or automatically posted again. A confirmed sibling retains its appearance time even when the parent is Unconfirmed. A view-as read is an admin disclosure, not a developer's practice-page delivery receipt.

## Evaluation joins

Scan membership identifies the actual practice/file pairs, including reused results from earlier producer reviews. Close-time recorded counts remain separate from currently visible open-problem counts. An accepted mark changes open counts, not the measured outcome or proof of repair. An incorrect mark records error and excludes unsafe support without resurrecting an older claim.

Reports retain listing completeness and separate Withheld from Not reviewed. No-result operational failures never enter a decided-result denominator. The [standard mapping](/contributor/practice-review-glossary.md#standard-terms) explains conceptual correspondence with external formats.

## Interpretation limits

Source authorization and erasure can hide proof while safe outcome facts remain available. Unavailable evidence is not an empty problem list. A refusal to copy earlier feedback cannot establish that the developer never received advice. A complete scan proves only its recorded boundary, not the correctness of the repository. The [data map](/admin/dsms/personal-data-map.md) owns export and retention limits.