Both are the one-hop failure the amendment itself is largely about -- the
correction reached the top of each file and stopped.
CONTRACT_P13S16_PROJECTION.md
The status block was moved to LANDED / RESOLVED, but two screens below it the
ratification-time framing still read "not the fourth", "nothing is
implemented", "every gate, test and mutation is specified and none has been
run", and "4a's landing obligation is outstanding by construction". A reader
scrolling past the status line met the pre-execution contract presented as
current.
Everything from that point to 7 is now marked as a DATED RECORD, S27-style:
retained rather than rewritten, because the pins are frozen and the review
records are the review state. The four "what ratification does NOT settle"
bullets keep their text and gain their outcomes -- t9 confirmed, t8d
falsified, all gates and mutations run, 4a honoured then discharged at
3328b2c, and the predicted post-execution amendments being this one.
EVIDENCE_P13S16_EXECUTION.md
Its header still said it was "deliberately not staged", absent from the touch
table, and that committing it would need an amendment adding a touch row --
describing, in the present tense, the state the commit it arrived in had just
ended. Recast to past tense, pointing at 7's touch row 12.
The distinction it was drawing survives and is kept: the fourteen paths in
aee4ff9 are the rung, this is the evidence the rung was reviewed against. It
documents aee4ff9; it is not part of it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ps1szk2mSfgp4Cz21eVH9x
Five corrections against CONTRACT_P13S16_PROJECTION.md, all against the
document rather than the code -- aee4ff9 is accepted and unchanged. Filed as
one amendment on the owner's decision because the findings are not
independent: two pairs share a root cause, and separate rounds would have
recorded symptoms while hiding the causes.
Root A -- the contract reasoned about MECHANISM where the question was what
the test ASSERTS. Findings 1 and 3. Both cells are right about the production
code and wrong about the observation, because a test that verifies its own
fixture is meaningful fails for a second reason no mechanism argument reaches.
Root B -- enumeration where completeness was required. Findings 2 and 4. Both
lists were built by reading one file. The contract already carries derive-do-
not-enumerate under pin 10a and 6 item 1, and broke it twice in its own
apparatus.
1. Invariant 21 does not detect the undo residue (0.6, pin 5a, pin 6b). It
abstains on dangling members by design; M5 shows invariant 10 firing
instead. The hole was already covered before this rung. u5's direct
members assertion is what signs pin 5 -- the fallback pin 6b itself names
-- so nothing is unsigned.
2. 3's M6a row names six failing tests; seven occur. The fifth all()
consumer is in tests/score_graph.rs, outside the four-in-generators.rs
list.
3. 3's t8d-under-M2 survivor cell is falsified -- it fails on non-vacuity,
not idempotence. The assertion must not be weakened: under M2 t8d is
genuinely vacuous. The other falsifiable cell, t9 under M1, was CONFIRMED.
4. Pin 10 cites four of the nine sites its instruction covers, and gate 6
cannot see core_spec.tex's "exactly 20 invariants" sentence.
5. cargo test --workspace truncates the failure set at the first failing
suite -- exactly the condition every mutation creates. 3's uniform rule
and gate 1 now specify --no-fail-fast.
Finding 6 is WITHDRAWN as false and retained as record. Pin 8's line numbers
were exact on the ratified tree 34232dc -- all four verified -- and only
stopped matching because this rung inserted ~1,500 lines above them. It had
been reported without the cited locations ever being measured. Writing the
amendment is what forced the check, which is an argument for consolidating
rather than filing six rounds.
Touch row 12 added, so spec/EVIDENCE_P13S16_EXECUTION.md is committed rather
than left untracked: the findings are conclusions, and the annex is the
evidence they were derived from, including the runs that falsified two of this
contract's own predictions. Its trailing whitespace is stripped so gate 4
passes, and it says so.
CLAUDE.md, the handoff and the ledger row are corrected in step -- all three
said "six findings", which is now the count before review rather than after.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ps1szk2mSfgp4Cz21eVH9x
Levi Neuwirth