Split and Rehome Critique: Prose has no reliable dereference, so a declared fact must be reinforced where it applies
Target: kb/work/agent-note-improvement/case-02-prose-dereference/baseline-working-tree.md
Main note to preserve
Claim: Because LLM-read prose does not mechanically dereference declarations into later uses, facts that must govern downstream interpretation often need local reinforcement, with an external check keeping the repeated copies aligned.
Minimum argument:
- In code and other codified forms, a named declaration can be resolved mechanically wherever it is referenced.
- LLM-read prose does not have that operation; applying a declaration later is an interpretive act.
- Interpretive reach decays with distance, non-obviousness, and conditional applicability.
- Therefore prose sometimes needs denormalized reinforcement at the point of use.
- Redundant prose copies can drift, so the safe pattern is denormalized reader-facing copy plus normalized verification.
- The rule is graded by representational form: codified artifacts need less reinforcement, prose-like artifacts need more.
Material that must stay: the code/frontmatter contrast, the status: seedling example, the denormalize-copy/normalize-check rule, the costs of bulk/branching/guard work, the representational-form boundary, and the falsifiable ablation.
Branch inventory
| Branch | Current location | Destination | Why |
|---|---|---|---|
| Code dereference analogy | Opening | Keep | It is the main contrast that makes the prose failure concrete. |
status: seedling point-of-use example |
Opening | Keep | It grounds the claim in an actual KB behavior: a status label should change how downstream claims are treated. |
| "Single-source-of-truth, correct for code, is unsafe for prose" | Main argument | Keep, but narrow | Strong and useful, but should be phrased as safe where dereference is mechanical rather than universally correct for all code. |
| Denormalized copies with normalized check | Main argument | Keep | This is the note's operational payoff and distinguishes useful redundancy from uncontrolled duplication. |
| Conditional applicability pushes branching somewhere | Costs | Keep, but compress | This is a real cost of reinforcement, but the paragraph currently carries more detail than needed for the central claim. |
| Guard work | Costs | Keep | It prevents the note from recommending unchecked duplication. |
| Representational-form scope | Scope | Keep | It prevents overgeneralization and links the claim to the KB's vocabulary. |
| Testing it | Final section | Keep, but tighten | The note is seedling; the falsifiable form is useful. The section should be shorter and point directly to the ablation. |
Rehoming candidates
None. The baseline does not contain a branch like the hallucination section in case 01. Its secondary material mostly supports the same operational rule.
Possible future note, not a split from this one:
- Conditional facts make prose denormalization pay a branching cost — Claim: denormalizing unconditional facts is cheap, but denormalizing value-dependent facts pushes branching into templates, process constraints, or validators. Required support: examples across note status banners, type-specific instructions, and generated indexes. It should not be split now because the current note needs this cost as a caveat.
Deletion candidates
- Delete or compress wording that invites a literal claim that all code uses single-source-of-truth safely. The relevant property is mechanical dereference, not code as a domain.
- Compress the conditional-applicability explanation so it states the branching tradeoff without walking through too many alternatives.
- Compress the "Testing it" section so the weak point and ablation do not feel like a second mini-proposal.
Revision target
The revised local copy should stay as one note. It should make the central chain more explicit: formal dereference makes single declarations travel; prose has interpretation instead; interpretation decays with distance and conditionality; reliable use therefore requires local reinforcement; safe reinforcement requires a normalized external check. The revision should tighten costs and testing, not split the note.