Bind allowed uses to records, purposes, and policy versions

A useful eligibility record identifies the data subject or governed unit, dataset or collection scope, permitted action and purpose, authoritative evidence, policy version, and applicable effective or revoked state. The organizational policy determines which inputs matter. Consent can be one relevant form of evidence, but not every permission or restriction in every setting is represented by the same mechanism. Distinguish an operational approval to deliver a requested service from an approval to build training features, evaluate an unrelated model, or export data to another organization. Reusing a payload does not automatically reuse the original permission.

A free-text purpose chosen by a client is not sufficient to establish an approved use. Bind the decision to a trusted workload identity and a concrete operation. Keep unknown, absent, conflicting, and expired evidence distinct from approved eligibility. If a policy denies training when evidence is unknown, implement that conservative branch rather than using a truthy default. Record the decision inputs and version needed for review while minimizing the sensitive information copied into the decision log. A policy representation such as ODRL can express permissions and constraints, but representation is not the enforcement mechanism or a complete legal assessment.

A training manifest should identify which policy evaluation and record membership produced it. A publication-time filter can prevent disallowed records from entering a release, but a long-lived job may execute after eligibility changes. Define whether the policy requires rechecking at execution, stopping an active job, withdrawing an artifact, or another documented response. The behavior must be explicit across queues, notebooks, feature stores, and exports. A downstream system that reads an unrestricted copy without consulting the required policy can bypass a correct upstream filter even when both teams believe the dataset was already approved.

Schema validity and quality validity are separate gates from use eligibility. A perfectly parsed, accurately labeled example can remain outside the training purpose set. Conversely, an eligible example may still fail quality or deduplication rules. Compute the intersection required by the release contract and record counts for each stage without adding overlapping exclusion counts blindly. Keep evaluation-only and training-approved cohorts separate so a later merge cannot erase their distinction. A reproducible data version explains what a prior run consumed; it does not grant permission to reuse those records indefinitely after the governing conditions change.

A changed approval can affect pending jobs, copied datasets, cached decisions, and derived artifacts. Use lineage and explicit ownership to discover affected candidates, then apply the policy's defined response and verify it. Deleting a training record does not automatically erase its influence from an existing model, so model handling needs its own documented decision and evidence. Preserve sufficient historical provenance to explain prior use without treating that history as current authorization. Restore and backfill workflows must also apply the current required use checks before re-exposing an old dataset, just as they must honor retention decisions.

The LRU cache lab highlights why a hit is not a fresh policy decision. For each separate view, Reset, set both capacity C and stride s, predict the initial 16-request counts, then inspect. With C=4, s=4, four distinct addresses fit and yield four misses plus 12 hits; with C=2, s=4, every one of those 16 requests misses. This deterministic cache has no policy version, revocation, expiry, or consent state. A real decision cache needs a justified invalidation or freshness mechanism; its replacement policy alone cannot establish that a previously cached allow decision still applies.