Skip to content

Commit 77d8087

Browse files
committed
fix(the-professor): make the cross-document pointer true (block#2137)
Round 4's only finding, and the most useful one of the sequence. §2c ended "The full history is in the redesign doc." That was false: the redesign doc contained no trace of it and was never touched by the commit that made the claim. Verified -- zero hits for the distinctive terms, and the doc's last edit predates that commit. Fixed by writing the history rather than deleting the promise. The redesign doc gains decision 15 -- both failed §2c enumerations, what each broke in which direction, and the general rule that a section defining a rule is the only place that states it. §2c's pointer now names decision 15 and the file by path, matching the convention the rest of the skill already uses for cross-document references and which this one sentence alone ignored. WHY THIS ONE MATTERS BEYOND ITS SIZE. The previous commit was the first written with a claim-check applied, and it still shipped a false claim -- which was the test the reviewer was asked to run. The claim-check tested claims the file makes about ITSELF: its own line counts, its own rules, its own sections. Every one of those passed and was independently confirmed passing this round. It never tested a claim about ANOTHER document, and that was the single outward-pointing claim in the commit. So the root cause holds -- nothing tested the document's assertions -- but its first implementation drew the scope too narrowly. A cross-document pointer is a claim, and it is the cheapest kind to check: does the target exist, and does it contain what you said it contains. The check now covers both, and was mutation-tested rather than asserted: restoring the vague "full history is in the redesign doc" pointer fails it, and pointing at a decision number the target does not contain fails it. Running that check against the real file immediately found two defects in the CHECK itself, both the same class: it matched line by line against a file hard-wrapped at 88 columns, so "exactly two / constraints" read as absent and "decision 15 in / `path`" silently never fired at all. A check that skips without saying so is worse than no check. It now matches against a whitespace-flattened copy. Signed-off-by: Serina Mcfall <[email protected]>
1 parent 0105600 commit 77d8087

2 files changed

Lines changed: 35 additions & 1 deletion

File tree

‎launchpad/Research/the-professor-skill-suite-redesign.md‎

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1614,3 +1614,35 @@ the last came out of measurement during the build on 2026-09-16.
16141614
and a rule blocking a verdict whose own reason names a different literal is retained even
16151615
though the reordering removed its cause, because the contract must hold for any configured
16161616
`$PROFESSOR_VERIFIER_CMD`, not only the one this was measured on.
1617+
1618+
15. **`verify-claims` §2c points at §2b rather than enumerating it — decided 2026-09-17, by
1619+
Serina, after two enumerations failed measurement.** §2c defines what a parse failure is.
1620+
It carried a list of the ways a response can fail to match §2b's grammar, and both
1621+
versions of that list were wrong in opposite directions.
1622+
1623+
The first listed "a response with no reason", which §2b has no test that can detect — its
1624+
line-one constraints are non-emptiness and the competing-literal rule, neither of which
1625+
judges reason quality. A parser built from the list had no rule to write for that clause.
1626+
1627+
The replacement was worse. It listed "a response whose line one is empty", a state §2b's
1628+
own normalisation makes unreachable because blank lines are discarded before the two-line
1629+
count is taken — and a parser applying the list in its written order **rejected three of
1630+
four leading-blank responses §2b accepts**, reintroducing the intermittent false-rejection
1631+
§2b calls a measured necessity. The same list silently omitted §2b's competing-literal
1632+
constraint, so it **accepted two shapes §2b and §2d block**. One list failed toward
1633+
over-rejection, the other toward acceptance.
1634+
1635+
**Decision: a section that defines a rule is the only place that states it.** §2c now says
1636+
the response did not match §2b's grammar and stops. A restatement of a grammar is a second
1637+
copy, two copies drift, and both drifts above were introduced by commits whose purpose was
1638+
to fix the previous drift. A pointer cannot drift from its target.
1639+
1640+
**This generalises beyond §2c, and that is why it is recorded as a decision rather than a
1641+
fix.** The same review sequence found five High defects in this pack, none of them a wrong
1642+
rule — every one was prose describing a rule correctly stated elsewhere. The root cause was
1643+
named in a problem record at
1644+
`launchpad/docs/audits/2026/09/17/prose-describing-correct-rules-shipped-unverified.md`:
1645+
no build step enumerated a document's own claims and tested them, so rules were
1646+
mutation-tested while the prose around them shipped unchecked. Prefer a pointer to a
1647+
restatement anywhere in this pack, and keep history in this document rather than in the
1648+
normative sections of a skill.

‎launchpad/agents/the-professor/skills/verify-claims/SKILL.md‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -330,7 +330,9 @@ the omission is the point.** It carried such a list until 2026-09-17. Both versi
330330
that list were measurably wrong — one naming a failure §2b cannot detect, the next naming
331331
a state §2b's normalisation makes unreachable while dropping a constraint §2b does
332332
enforce. A restatement of a grammar is a second copy of that grammar, and two copies
333-
drift. Pointing at §2b cannot drift from §2b. The full history is in the redesign doc.
333+
drift. Pointing at §2b cannot drift from §2b. Both failed lists, and the decision to stop
334+
enumerating, are recorded as decision 15 in
335+
`launchpad/Research/the-professor-skill-suite-redesign.md`.
334336

335337
**Non-completion — judged independently of anything stdout contained.** A response can
336338
be perfectly well-formed and still not count, because how the command ended is part of

0 commit comments

Comments
 (0)