David F GliddenandClaude Opus 4.8
d7824cffe8
Session file + ledger returns + 7 KG appends (2 drift-patterns, 3 good-direction, 2 facts) + feedback-derive-the-rule-from-the-consumer-not-from-the-survivor. MEMORY.md: prior Active Session demoted to MEMORY-reference.md, new promoted. Skill-harvest register: /wake-up backlog substrate-check patch + /field-divergence-sweep create, both PROPOSED for steward authorization. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Xefg5EXwcpd9RMAr63dWrD
3.2 KiB
name, description, metadata
| name | description | metadata | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| feedback-derive-the-rule-from-the-consumer-not-from-the-survivor | When two implementations of one field disagree, do NOT pick the survivor — derive the rule from what a CONSUMER of the value must do, then test the shapes where each candidate happens to be right. Comparison is selection, not derivation. |
|
When two implementations compute the same field differently, the reflex is to decide which one is correct by comparing them. That is selection, not derivation — and it silently inherits the assumption that one of the two is right.
Why: verifying that A ≠ B tells you nothing about whether A or B is correct. The correct rule comes from what a consumer of the value must be able to do with it — and it may be neither candidate.
Earned hard 2026-07-27 (chamber-library source_lines), in two stages:
- Two formulas disagreed:
len(splitlines())(engine-side) vscount("\n") + 1(chamber-sidewrite_sidecar). I ratifiedsplitlines()on the strength of the engine-side convention — without checking whatsplitlines()splits on. It also breaks on\v \f \x1c \x1d \x1e \x85 U+2028 U+2029, which no consumer of a line coordinate treats as a break (editor ·wc -l·sed -n 'Np'· the engine's chunker). Four canonicals carry them; onlintottthe disagreement is 308 lines — a coordinate off by 308 is a citation pointing at the wrong page. - Both candidates were wrong, on complementary subsets:
count+1on the 1,261 newline-terminated files,splitlines()on the 4 exotic-bearing ones. The correct rule (count("\n") + (0 if trailing newline else 1)) was a third thing.
Two corollaries, both load-bearing:
- Complementary-correctness defeats sampling. Because each wrong candidate is right on some subset, no single sample distinguishes all three — the steward's "confirm the rule, not the sample" (which caught a live corruption risk: 36/1,297 files where a blanket
-= 1migration would have corrupted data). Test the shapes where each candidate happens to be right, not a convenience sample from one producer. - Fix the root cause: ONE implementation, imported. Two formulas for one field is how the divergence arose; a third arrives the same way. Put the definition where the semantics live (the gate that owns the coordinate) and have producers import it.
How to apply: when a field is computed in more than one place — (1) ask what a consumer must be able to DO with the value, and derive the rule from that; (2) enumerate the shapes where each existing candidate is right, and test all of them; (3) collapse to one implementation; (4) make any migration RECOMPUTE, never do arithmetic on the stored value.
Kin: feedback-checkable-claim-surfaces-bugs (the demand for verifiability as defect-detector), feedback-census-by-mechanism-not-proxy (the instrument must bear the claim), feedback-trust-prior-pass-frame (an existing implementation's correctness is not inherited by comparison). A new face of assert-from-derived-not-substrate: the derived artifact here is the other implementation.