3.4 KiB
name, description, metadata
| name | description | metadata | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| feedback-census-by-mechanism-not-proxy | A plan-of-record generalization must be produced by — and carry — the instrument that can bear it; census by mechanism (run the real pipeline), not by a cheap proxy, and distrust the bucket that arrives when you want to feel done. |
|
When turning a pile of surprises into "a finite, counted set," the classification is only as trustworthy as the instrument that produced it. The 2026-07-24 V-TEXT gap-census bucketed 31 files as "heavy-styled → one fix" from class-attribute VOLUME (a pandoc-free count). Running the actual pipeline on the smallest one (the-shape-of-a-pocket, 2026-07-25) proved the generalization wrong: volume doesn't determine mechanism — deep-nested spans, presentational div-fences, internal-nav graphs, and art-book plates all read as "high class count," and only running pandoc distinguishes them. A count is not a diagnosis. The census's own hedge ("the pandoc-free triage could not see…") sat right next to the over-claim ("dominated by ONE fix"); the over-claim got inherited as the plan, because a plan is what you act on and a hedge is a footnote.
Why: the confident bucket arrived at the end of a bottomless day, and its appeal was that it made the wave feel finite and conquered. That relief did some of the believing — the contamination shape (compose the conclusion that resolves the pressure, not the one the method supports). The corpus never lets a text claim verified without running the gate; we let a plan claim "one fix" without running the pipeline. The double standard is the bug: our planning artifacts escape the epistemic discipline we enforce on the corpus. Kin to feedback-completion-is-a-tripwire (the feeling of done is a cue to verify, not a signal to ship) and feedback-checkable-claim-surfaces-bugs (demand the checkable claim over the soft classification).
How to apply:
- Distrust the generalization that arrives when you want to feel done. When a classification suddenly makes a hard problem feel conquered, treat the relief as a signal to check whether the instrument that produced it can bear the claim.
- A plan-of-record claim must carry its instrument and its confidence. Write "31 files, high class-volume, mechanism UNVERIFIED pending a pipeline run; hypothesis: class-strip" — not "31 heavy-styled, one fix." Apply the corpus's own
UNVERIFIED-not-PASShonesty to the artifacts that plan the corpus. - Census by mechanism, not proxy. If the real measurement is affordable (running one file end-to-end took ~1 hour; the whole set is an unattended afternoon), run it across the set once and bucket by the ACTUAL residual profile — not by a proxy that correlates with "messy" but not with "which fix." That mechanism-census is the do-it-once artifact that actually ends the whack-a-mole; the proxy-census only feels like it does.
- Same lesson at three scales this session: two buggy ad-hoc regexes (
\bcensus,^..\d+..:def-count) asserted from a derived proxy instead of the substrate; the volume-census did it at the planning layer. One rule: don't trust the cheap proxy when the real measurement is affordable — and produce the checkable claim with the same instrument the production path uses.