Full Improvement Pass: Bottom-up structure inference needs capture at the decision surface, not the state
Target: kb/notes/structure-inference-needs-capture-at-the-decision-surface.md
Reports used: compression bundle, critique-note, composition-friction-gate, connect (semantic bundle skipped — exploratory pass, not a promotion decision)
Strongest retained claim
Bottom-up inference of rationale-bearing structure (the entities and relations that name why a decision went one way) from traces is feasible only when capture is decision-shaped rather than state-shaped, because the "why" is cheap to record at decision time and decays toward unrecoverable once only the resulting state remains. The italicized scope qualifier is new — see Body edits — and narrows the claim to defend it against critique-note's strongest attack, rather than leaving it stated as "structure" in general.
Body edits
| Location | Source method(s) | Finding | Action | Rationale |
|---|---|---|---|---|
| "Capture position..." §, final sentence of paragraph 1 ("not an optimization... learnable at all") | critique-note, composition-friction-gate (joint 2) | Two independent methods converged on the same sentence: critique-note calls it a categorical overclaim the note's own Open Questions undercut; friction-gate flags it as UNSUPPORTED for the same reason (tension with the continuum admission). Independent corroboration, not a hedge. | reword to scope the categorical claim to "past whatever threshold," matching the note's own continuum admission | Keeps the qualitative-shift intuition without asserting a hard binary the note elsewhere disclaims |
| "Capture position..." §, paragraph 2 (wikiwiki-principle paragraph) | compression/branch-bloat, compression/marginal-value-redundancy | Narrates this note's relation to another note's claim; duplicates the Relevant Notes annotation for the wikiwiki link | replace with a new paragraph (below), keeping one clause ("deferring structure is safe... why does not wait") | Compression's suggested revision explicitly asked to preserve that one clause while cutting the rest |
| (new paragraph, same location) | critique-note (primary attack + constructive finding) | Strongest attack: decision/process mining recovers structural regularities from state-shaped event logs with no decision receipts — a working counterexample to "only if," at the population level. Critique's own fix: state explicitly which scope the claim targets. | add a paragraph distinguishing rationale-bearing structure (this note's target) from predictive regularities over an already-imposed state schema (what decision mining recovers), and note the distinction sharpens rather than resolves the note's own continuum open question | This is not a hedge against a possible objection — it is critique-note's constructive finding, independently reinforced by friction-gate's joint 2 on the adjacent sentence. Left un-engaged, the note's "only if" claim is contradicted by a real, cited research program. |
| "Boundary" §, paragraph 2 (create-memory-directly / trace-derived-extraction paragraph) | compression/branch-bloat, compression/marginal-value-redundancy | Duplicates the Relevant Notes annotations for create-memory-directly.md and use-trace-derived-extraction.md almost verbatim; doesn't bound this note's own claim the way paragraph 1 does |
cut | Deletion test passes per the compression report — nothing in the note's support route depends on it |
| "The 'why' is cheap..." §, sentence 3 | compression/detail-overhang (INFO) | Re-lists the intro's decision-shaped/state-shaped item inventory in near-synonymous form | compress to a callback ("which of the inputs, constraints, or exceptions named above") | Lower priority than the WARNs above, but a safe, mechanical trim consistent with the INFO finding |
Not adopted: friction-gate's joints 1, 3, 4, 5 (the "nearly free" claim, the unsupported locality-of-intent mechanism, the "same gap" identity claim, and the "independent preconditions" claim) are real findings but not independently corroborated by another method, so per this instruction's rule they are not converted into edits here — see Routed attention below. Also not adopted: critique's secondary objections (the capture-schema-is-itself-upfront-imposition point, and the correlated-multi-stream recovery point) — both are real but read as hedges against edge cases the note's existing boundary language already gestures at; adding them would reopen the additive-apparatus failure mode compression is biased against.
Routed attention (composition-friction-gate — not auto-resolved)
Filter verdict: SURVIVES — the apparent binary-vs-continuum tension resolves because the note's own Open Questions already supply the reconciling move (a threshold on a continuum is still a necessary condition).
Thinnest joints (not converted into edits by this pass): 1. "Recording those 'decision receipts'... is nearly free" — UNSUPPORTED — conflates the information being available at decision time with capturing it being costless; authoring/maintaining consistent capture across every decision path (including exceptions) is not argued to be free, only that the raw material is in hand. 2. (Corroborated by critique-note — resolved above, not left routed.) 3. "The reason is a locality property of intent..." — THIN — the note's single load-bearing mechanism, asserted once with no worked example or corroboration. 4. "This is the same gap [raw accumulation]... names as the ingress problem" — THIN — claims identity between this note's mechanism and the other note's, where the text only demonstrates analogy. 5. "...two independent preconditions on trace-derived memory... satisfying one says nothing about the other" — THIN — independence is asserted by architectural parallelism; a decision-shaped record (e.g., an approval-and-signoff receipt) plausibly correlates with easier verification, which the note doesn't rule out.
Connection candidates (from connect report)
agent-memory-requirements/preserve-evidence-without-loading-history.md— extends (add). That note requires capture to keep enough for later extraction and separate capture from activation; this note supplies the positional precondition it leaves open.agent-memory-systems/lightweight/trajectory-informed-memory-generation.md— evidence (add). A trace-extraction system whose tip categories are tied to outcome/exception shape — corroborating evidence for the claim, from a different collection.distilled-artifacts-need-source-tracking.md— contrasts (add, outbound only). Same locality-of-provenance principle applied to a different provenance kind (source-dependency vs. decision rationale). Connect flagged the reverse edge as the source-note author's judgment call; out of scope for this pass, which only edits the target note.
Proposed revision shape
Same overall structure and length, net roughly neutral: the "Capture position" section's second paragraph is replaced (not just cut) with a scope-narrowing paragraph that does real argumentative work against critique-note's strongest attack; the "Boundary" section loses its second paragraph outright; the "why is cheap" section's redundant item-list sentence is trimmed to a callback. No section is removed, split, or rehomed — connect's own "not a split candidate" flag agrees. Footer gains three links.
Open items
- Friction-gate's joints 1, 3, 4, and 5 remain genuinely open — they were deliberately not resolved by this pass (see "Routed attention"). A future editor should judge whether any deserves a real fix; this instruction does not decide that for them.
- Whether the rationale-vs-regularity distinction added to the "Capture position" section actually survives a second adversarial pass (i.e., whether a sufficiently rich state-shaped log could still be argued to name new rationale entities, not just fit a rule over given fields) is left as the sharpened form of the note's own second Open Question, not resolved here.