Full improvement pass over the recoverability/economics cluster

Goal

Run run-full-improvement-pass-on-note.md over the seven notes this cluster produced or altered on 2026-08-22/23, plus one older note the work implicated — eight in all. Both batches now have retained, completed passes. Batch 2's first pass overlapped Batch 1's edits, so all four Batch 2 notes were rerun in binding order against Batch 1's committed result. Every Batch 2 reframe follow-up is complete.

The notes were written fast, in one conversation, several of them revised or retitled hours after landing. At workshop start, none had been through a full pass. They validated clean and were densely linked, which is exactly the state where an unexamined claim survives on structural plausibility.

Batches

Two batches of four. The notes are topologically sorted across both batches, and order is binding within each — every note runs after the notes it cites. A pass can reframe a note's title and thesis, which invalidates citers — so the most-depended-on notes go first, and the second batch runs against a settled first batch.

  • Batch 1 — complete and committed.
  • Batch 2 — complete after an ordered rerun against the committed upstream notes.

Batch 2 depends on batch 1 six ways: addressability-grain extends documentation-generates and grounds on llm-recompute; artifacts-must-preserve-named-choice-scope contrasts superseded-choices and cites areas-exist as evidence; an-insufficient-summary extends llm-recompute and contrasts documentation-generates. If batch 1 reframes any of those, batch 2's premises move before its own pass reads them.

Two further edges fall inside batch 2 — addressability-grain contrasts opposed-recompute-factors, and an-insufficient-summary contrasts addressability-grain — which is why order is binding within a batch and not only across the boundary.

Excluded: artifact-analysis-README.md, document-system-README.md, context-engineering-README.md. They are type tag-readme, weight-gated, and carry index entries rather than claims; the pass targets notes.

What this pass is likely to find, and should not be shielded from

Recorded so a later session does not mistake a known weakness for a surprise, and does not pre-empt the pass's own judgment either. These are expectations, not instructions to the pass.

  • opposed-recompute-factors carried a claim that might not have been its own. Four consecutive "Consequence:" sections, one of which — segmentation strips the drift detector from the low-traffic layer — is a maintenance claim on an economics note, and more general than its host. A split candidate. Deciding it before the pass would also be legitimate; deciding it during is what the pass is for.
  • Two notes cite Commonplace as an existential witness. The collection-fit check should confirm that reads as witness rather than subject. Both were written under the earlier choice-scoping requirement, so this is a live test of whether the requirement was met, not a formality.
  • llm-recompute received a scope paragraph on 2026-08-23 naming the volume condition under which its default flips back. It is the oldest note here and the least examined against its new neighbours.
  • artifacts-must-preserve-named-choice-scope was revised by a second agent after its initial write, including a replaced section. Two authors, one pass.

Operating constraints

From the instruction, not restated in full — read it before starting.

  • One note at a time per pass. The concurrency precondition forbids any other actor editing {note-path} from step 1 until the pass stops or completes. Two passes over two different notes may overlap; two passes over one note may not.
  • Steps 8, 9, 10 are a strict pipeline, not a parallel batch. Closing jobs snapshot the note at pair-create, so a closing job queued while the copyedit runs pins freshness to pre-copyedit text.
  • Re-entrancy preflight first. A prior pending report for the same historical path blocks a new pass until resolved.
  • Cost is real. Six method families per note, each in a fresh single-use sub-agent, plus a closing cycle that reruns all of them. Budget per note, not per batch.
  • A reframe leaves the KB asserting the old claim. The required follow-up — rename via commonplace-relocate-note, then fix inbound link text, summaries, and any citer that used the old title as a premise — runs after the pass, not inside it. Within this cluster the citers are mostly each other, so a batch-1 reframe creates batch-2 work.

What closes this workshop

Every note has a retained packet. keep packets are applied and closed; delete, merge, and rehome packets are handed back with their dispositions unresolved, since executing those is the reader's call. Any reframe's follow-up operation is either executed or recorded as a named open item with concrete scope. Findings that generalize beyond these six notes are promoted; the rest is consumed and the workshop deleted.

Complete file listing (generated at build time)