What text diff misses in controlled-document change

A redline can be exact and still be incomplete

In the synthetic QMS-SOP-014 revision, a text comparison precisely identifies two edits: may becomes must, and 30 calendar days becomes 15. That evidence matters. It still does not answer whether the edits change an obligation, which downstream material depends on the old rule, or who may accept the resulting actions.

• Text evidence answers: which characters differ?
• Semantic normalization asks: which actor, action, modality, condition, or timing changed?
• Impact evaluation asks: which declared objects have an eligible evidence-bearing path from that change?

The synthetic meaning change

The manually authored Semantic Delta records the Quality Investigator as actor, initiate investigation as action, OPTIONAL → MANDATORY as modality, and WITHIN 30 CALENDAR DAYS → WITHIN 15 CALENDAR DAYS as timing. Both fields remain bound to exact old and new excerpts.

The normalized fields are interpretations bound to evidence. The exact excerpt remains the source, and an accountable reviewer remains responsible for accepting or correcting the interpretation.

• This is a fictional explanatory record, not a customer result.
• One semantic delta records two material fields without turning either into an assurance claim.

Impact needs a dependency basis

WI-08 is potentially affected because a confirmed exact-reference and structural relationship shows that it cites QMS-SOP-014 and repeats the old timing. TR-22 requires review because its semantic relationship is inferred rather than confirmed. The same textual change therefore leads to different downstream states because the relationship evidence and lifecycle states differ.

A useful finding must retain endpoints, direction, relationship type, evidence, origin, method, version, and review state. Without those fields, a plausible target list cannot be reconstructed or challenged.

A bounded negative finding is not a universal conclusion

FRM-11 was successfully processed and no eligible path was retained under the declared synthetic graph and policy. SOP-21 has a stale source version and is not evaluated. The first can receive the bounded phrase no relevant impact found within evaluated scope; the second cannot silently receive it.

The distinction matters operationally: a processed object with no path and an object excluded by a coverage gate require different follow-up, even if neither appears on the positive-impact list.

Reproducibility has a boundary too

The current prototype basis preserves frozen M3–M4 comparison and graph provenance so recorded deterministic and model-free projections can be replayed against their pinned inputs. That is narrower than reproducing a probabilistic model call or independently verifying a complete future release package.

Reconstruction makes the declared basis inspectable. It does not convert machine inference into proof, repair missing sources, or authorize closure.

Questions for empirical work

The next evidence step is not a larger synthetic story. It is a bounded historical replay with an adjudicated review set, agreed measures, documented misses and disagreements, and permission rules established before any public result.

• Does separating text, meaning, dependency, impact, and decision improve reviewer comprehension?
• Which dependency channels contribute useful candidates, and where do they create review burden?
• Can evaluated-scope wording prevent false reassurance when coverage is incomplete?
• What evidence must remain visible for an accountable reviewer to accept or reject a finding?