Change-Provenant Supervision: Governing Learned Artifacts Under Policy Change

Separating recorded lineage from independent current-contract revalidation for sound admission of learned artifacts

Jesus Salas· 2026· DOI 10.48550/arXiv.2609.28574

The core problem

Learned artifacts such as fine-tuned adapters are increasingly governed by recorded dependency graphs that link model bytes to training data, authority records, and downstream targets. A central problem motivates this work: a recorded dependency graph cannot certify that it contains no omitted edge. For learned artifacts, graph-scoped invalidation therefore cannot by itself justify admission after authority changes, especially when output testing misses a provenance-stale derivation. The paper addresses this gap by separating two concerns that are often conflated. First, recorded lineage proposes an impact scope and a dependency explanation. Second, independent current-contract revalidation of every retained target supplies soundness relative to the declared contract. The design is evaluated in two evidence layers: a set of four Qwen3-14B LoRA packages binding 1.028 GB of adapter bytes to training and authority records, and a controlled harness with six role-separated amendments. The stated contribution is a governance mechanism whose properties are established under controlled conditions, not an enterprise invalidation rate, unlearning result, operationally optimal workflow, or me

The results are reported across the two evidence layers. In the adapter layer, stale and full-fresh adapters trained on mutually exclusive versioned supervision emitted identical plans on all 80 tested authority-neutral inputs. Despite this behavioral equivalence, byte-derived reconstruction and a sealed current-contract ledger refused the stale and naive-append packages and admitted the full-fresh and lineage-selective packages. In the controlled harness with six role-separated amendments, transitive lineage recovered all 40 semantically invalidated amendment-target pairs while nominating 56 impact candidates, compared with 120 under document-wide invalidation. All 56 nominated candidates had a declared path to an amended clause, although 16 remained content-valid. Omitting one load-bearing edge per amendment reduced semantic recall to 34/40, while independent revalidation refused all six affected packages. These outcomes establish properties of a controlled governance mechanism rather than an enterprise invalidation rate, unlearning result, operationally optimal workflow, or measured total-cost reduction.

Learned artifacts such as fine-tuned adapters are increasingly governed by recorded dependency graphs that link model bytes to training data, authority records, and downstream targets. A central problem motivates this work: a recorded dependency graph cannot certify that it contains no omitted edge. For learned artifacts, graph-scoped invalidation therefore cannot by itself justify admission after authority changes, especially when output testing misses a provenance-stale derivation. The paper addresses this gap by separating two concerns that are often conflated. First, recorded lineage proposes an impact scope and a dependency explanation. Second, independent current-contract revalidation of every retained target supplies soundness relative to the declared contract. The design is evaluated in two evidence layers: a set of four Qwen3-14B LoRA packages binding 1.028 GB of adapter bytes to training and authority records, and a controlled harness with six role-separated amendments. The stated contribution is a governance mechanism whose properties are established under controlled conditions, not an enterprise invalidation rate, unlearning result, operationally optimal workflow, or measured total-cost reduction.

The methodology instantiates change-provenant supervision as a two-layer evidence design. The first layer comprises four Qwen3-14B LoRA packages that bind 1.028 GB of adapter bytes to training and authority records. Stale and full-fresh adapters are trained on mutually exclusive versioned supervision. The evaluation tests whether these adapters emit identical plans on all 80 tested authority-neutral inputs, and whether byte-derived reconstruction plus a sealed current-contract ledger can refuse stale and naive-append packages while admitting full-fresh and lineage-selective packages. The second layer is a controlled harness with six role-separated amendments. Transitive lineage is used to recover semantically invalidated amendment-target pairs and to nominate impact candidates, compared against document-wide invalidation. The harness also tests the effect of omitting one load-bearing edge per amendment on semantic recall, and whether independent revalidation refuses affected packages. Formally, let G = (V, E) be the recorded dependency graph and C the declared current contract. Recorded lineage yields an impact scope S \subseteq V and an explanation path set P . Independent revalidation is a predicate R : V \times C \to {admit, refuse} applied to every retained target v \in V ∖ S . Admission requires R (v, C) = admit for all retained v, while soundness is relative to C, not to G . The design is summarized in the following flow.

The findings support a clear separation of concerns. Recorded lineage is useful for proposing impact scope and providing a dependency explanation, but it is not sound on its own because a graph cannot certify the absence of omitted edges. Independent current-contract revalidation of every retained target supplies soundness relative to the declared contract, and it is this revalidation that refuses stale and naive-append packages even when output testing shows identical plans on authority-neutral inputs. The harness results quantify the trade-off: transitive lineage recovers all 40 semantically invalidated amendment-target pairs while nominating 56 candidates instead of 120 under document-wide invalidation, and all 56 have a declared path to an amended clause, though 16 remain content-valid. The sensitivity result is equally important: omitting one load-bearing edge per amendment reduces semantic recall to 34/40, and independent revalidation refuses all six affected packages. This indicates that lineage quality affects recall, while revalidation provides a safety net that does not depend on lineage completeness. The authors explicitly bound the claims: the results establish properties of a controlled governance mechanism, not an enterprise invalidation rate, unlearning result, operationally optimal workflow, or measured total-cost reduction. The architecture can be summarized as follows.

CS practitioners and researchers

Opening member content…