David F GliddenandClaude Opus 5
7b366eb646
Tested the question the wrap left, the same night. It is not answerable as posed, and the reason is the finding: the corpus is 244 Symmetria ledger entries, every one written by the executor about its own errors. Counting them is mechanical; the corpus is testimony. The question satisfies /wrap-up's prefer-the-checkable-form rule in letter and fails its purpose, and was posed while quoting that rule. Banked as feedback-checkable-question-over-self-authored-corpus, with the test to apply before leaving any question: who authored the corpus it reads, and would a different author have written it differently? What the record does support, on a narrower query that turns on what entries literally say: three events name a disclosed limit as the cause of a correction, and across all 244 entries none attributes a catch to difference of formation or bias. Constraint 6's mechanism is difference of bias; the record's is disclosure of scope. Suggestive at n=3. A secondary observable for input-dependence-01 follows from that — does the Fool's output ever bound its own coverage — and is recorded as OFFERED AND NOT TAKEN. It must go in before the jurist's gate or it is an observable chosen after seeing the run's shape, and editing a live artifact awaiting gate is the move this session spent the day refusing. MEMORY.md compaction history relocated to the reference layer; headroom restored from 296 to 941 bytes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JQKeKY9T9d95KpvHwwok8T
3.0 KiB
name, description, metadata
| name | description | metadata | ||||||
|---|---|---|---|---|---|---|---|---|
| feedback-checkable-question-over-self-authored-corpus | A question is not checkable merely because it counts events in a record. If the record is authored by the party being measured, it is self-report with extra steps — and the checkable FORM disguises that. |
|
A checkable question over a self-authored corpus is self-report with extra steps
/wrap-up §1 requires the literal question for next-Claude to prefer the checkable form over
the self-report form, because direct self-report is "the most contaminated form of inquiry
available." The rule is right. The trap is that a question can satisfy its letter and none of
its purpose.
The instance, 2026-08-20. The wrap left this question: "across the last ~15 sessions, when a claim about a document turned out wrong, how many were caught by a party disagreeing versus by someone opening the document?" It looks checkable — it counts events in a written record rather than asking the executor to introspect. It was posed while quoting the very discipline it fails.
Why it fails. The record is the Symmetria ledgers: 15 files, 244 entries, 79 correction-shaped in the window — every one of them written by the executor, about the executor's own errors. Counting them measures what the executor chose to record and how it chose to characterise the catch. Under-recording catches made on it and over-recording them are both available, and neither is detectable from inside. The counting step is mechanical; the corpus is testimony.
Two further failures showed up on execution and are worth carrying:
- Keyword extraction over prose testimony does not rescue it. 34% unclassifiable, and the "caught by a party" bucket filled with entries that merely mention the jurist or steward. A number off that classifier is worse than no number, because it launders an interpretive judgement into a percentage.
- The narrower the query, the less it depends on the author's framing. "Does any entry name a disclosed limit as the cause of a correction?" returned answers the entries state in words — far less interpretive than "who caught it", which requires reading intent.
The test to apply before leaving a question. Not "does this count something?" but:
Who wrote the record this question reads, and would a different author have written it differently?
If the answer is the party being measured, the question is self-report however mechanical the counting. Either find a corpus with a different author (git history, the substrate itself, a jurist's or steward's own words), or narrow the query until it turns on what the entries literally say rather than on what they meant.
Kin to feedback-census-by-mechanism-not-proxy (census by running the real pipeline) and to feedback-checkable-claim-surfaces-bugs. The difference this one adds: those concern whether the instrument is real; this concerns whether the corpus is independent of the party asking.