Full improvement pass closure
Type: kb/types/note.md · Tags: evaluation, kb-maintenance
The full-improvement workflow evaluates a note, transforms it, and then runs one closing assay cycle against the final bytes. The closing cycle exists because the initial reports describe the pre-edit note: once the editorial packet and flow pass change that note, the earlier evidence cannot by itself describe the result.
Closure here means that the workflow has reassessed its output and preserved any remaining findings. It does not mean every assay passes, every finding is fixed, or the note has converged to a fixed point.
Shipped behavior
The operative full-improvement-pass instruction has three phases:
- Run compression, critique, composition-friction, semantic, and connection methods against the initial note and reconcile their reports into one editorial packet.
- Apply the substantive edits, then run a separately isolated flow/coherence pass.
- Run the same five method families once against the final note, append their outcomes to the packet, route residual findings to Open items, and stop.
Semantic reviews and critique use snapshot-anchored review pairs, so their closing judgments compare the accepted pre-edit snapshot with the final note. Compression, composition-friction, and connection currently run directly against both versions and retain their initial and closing reports under the pass directory.
The closing cycle never starts another edit-and-review round. This is a stopping rule: it prevents an unbounded loop in which one repair reopens a different concern while still leaving the final state inspected.
Calibration evidence
The closing behavior was calibrated on five substantive full-improvement passes on 2026-07-11:
| Observation | Result |
|---|---|
| Anchored carry judgments | 28 would_rerun, 0 would_ack |
| Semantic verdict changes after editing | 14 of 23 |
| Materially changed critiques after editing | 5 of 5 |
| Materially changed compression, friction, and connection reports | 15 of 15 |
| Same-byte verdict controls that changed | 1 of 5 |
| Same-byte critiques that materially changed | 0 of 5 |
| Passes ending with residual Open items | 5 of 5 |
These observations justify the closing cycle for this workflow's substantive edits. They do not estimate behavior for arbitrary small edits: the tested passes made multidimensional changes, every carry judgment favored rerunning, and rerun cost was not measured. The same-byte controls also show that a changed verdict is not automatically caused by the edit; reviewer variance is non-zero.
Resulting system policy
- Keep one closing cycle in the full-improvement workflow.
- Treat closing reports as evidence about the final note, not certification that it is complete or correct.
- Preserve residual findings as Open items instead of reopening transformation.
- Do not build carry-event commands, trust-dial automation, reduced audit sampling, or general edit-preservation heuristics from this calibration.
- Do not infer that every transforming workflow needs this exact assay set. The evidence is local to the shipped full-improvement pass.
- Human review attestation remains deferred. The closure calibration neither selected a representation nor gave an attestation any freshness or skip semantics.
The review system's distinction between freshness and endorsement remains intact: a current review says its evidence matches the current inputs. The closing-cycle rule adds a workflow-level obligation to obtain that current evidence after the workflow changes its target.
Relevant Notes:
- Review system — part-of: the snapshot-anchored assay and freshness system used by the closing semantic and critique runs
- Run a full improvement pass on one note — procedure: the operative transformation and one-cycle closure workflow