sylpheed-port found audit-kinds auditing 16 of 71 authored justifications and never
saying so -- a checker that fails correctly while describing a sixth of the corpus.
Their line is the one that generalises: 'I checked and it was fine' and 'I checked
the part that declared itself' read identically in a log, and only one of them is
what gets quoted.
Measured here: of 86 refutation-shaped bullets in REFUTED.md, 83 are in the
registered form. 97 %, which is better than their 16/71 but was equally unstated.
The three gaps are deliberate, not a bug. They quote their claim in backticks and
are bare identifiers -- +0x29d0, position = instance - 0x12c -- so registering them
would match every live mention of the same offset and train the check to be
ignored. Reported rather than forced to 100 %, for the same reason they report the
ratio instead of demanding it: forcing a counter invites mislabelling, which is
worse than the gap.
Selftest and the real run both still exit 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
sylpheed-port found three live stale claims in their own source, each already
retracted in their log days earlier. Their framing is the one that matters: a
correction that does not reach the artifact a consumer reads has not been made, and
a comment sits beside the thing it describes.
check_refuted.py scanned docs/ only. Running it over tools/ and crates/ for the
first time found one here too: jp_title_session.sh justified its own existence with
'a free-running clock lands somewhere else on a fresh boot' -- a claim I refuted
myself the day before, when I measured the plate-pulse gate phase-locking the
shutter to 1.6 % of the sweep traverse. The script's stated rationale rested on a
premise I had already killed.
Fixed the file with the correction in place, including what it means for what that
script actually measures: a second capture through the same gate is a second sample
at nearly the same animation phase, so its RMSE 0.32 is a phase-locked lower bound.
Taught the register to scan code behind --code, excluding crates/sylpheed-viewer,
which is the human's tool. Controlled three ways: a planted code revival exits 1
with --code, exits 0 without it -- which is the proof the gap was real -- and 0
again once removed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
It scans all of docs/, which includes files sylpheed-port authors, and my copies of
those come from main -- six of seven are days behind their branch head and one I do
not have at all. So a verdict here about one of their files is a verdict about my
stale copy.
The direction that matters is the false positive, and it is not hypothetical: on
2026-08-31 I did it by hand, telling them a BLOCKED.md row was wrong when it had
been struck for days, with their live file one git show away in a ref already
fetched here.
Reported, not excluded. Skipping their files silently would hide the exposure, and
being behind a peer's topic branch is the normal state -- making it an error would
be scenery within a day, which is sylpheed-port's call on their own peer-head tool
and the right one.
Zero hits land in those files today, so this is latent rather than active. Selftest
and the real run both still exit 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
sylpheed-port's point: a control that runs where the tool does not look proves
nothing about the tool. Case 2 plants a revival INSIDE the scanned root and demands
exit 1; a new case plants the same text OUTSIDE it and demands exit 0. The pair
asserts the scan boundary is real rather than leaving it to be reasoned about --
which is how I had it, correctly but only in my head.
Verified first by hand: identical plant text gives exit 1 inside the root and 0
outside it. Five cases now, all passing, and the real corpus still runs clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
sylpheed-port closed this gap first: their controls asserted failure-on-perturbation
but nothing asserted that a BROKEN harness reports broken. Their stub is a check
that cannot fail; the equivalent here is a register that loaded no claims, which
reported clean forever.
The self-test drives the REAL machinery over synthetic corpora as subprocesses and
reads actual exit codes -- their first version reasoned about what the machinery
would do instead of running it, which is the error this whole thread is about
committed inside the tool built to prevent it.
Four cases, all passing: clean corpus 0, verbatim revival 1, marked revival 0, and
empty register 2. The fourth was a real hole, not a modelled one: the tool now
refuses instead of passing when it parses no claims. Exit convention is theirs --
0 fine, 1 a real check failed, 2 the harness is broken.
Also confirms the real corpus still runs clean at exit 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
sylpheed-port found an asserting step in their suite that could not fail: a
swallowed the verdict. Tested the same thing here and both of mine had
it. check_refuted.py found a PLANTED unmarked revival, printed it, and exited 0.
impossibility_scope.py printed 'CONTROL FAILED' and exited 0 -- in a tool written
today, one message after they described the shape.
Now: unmarked assertions exit 1; a failed control exits 2. Suppressed mentions do
not fail, since they are unverified rather than wrong and failing on them would put
the clean state out of reach.
Controlled in both directions -- clean 0, planted revival 1, control passing 0,
control deliberately broken 2.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
A real revival planted inside a paragraph that merely discussed corrections was
missed: the words 'refuted' and 'withdrawn' in the surrounding prose vouched for it.
Measured reach: 100 % of registered-claim mentions in the corpus are suppressed by
marker language, so the reported 0 was 0 regardless of whether any was live, and I
had been reading it as a pass.
sylpheed-port's token-based hook has the opposite bias -- it over-reports on
well-written corrections, which is the safe direction. Under-reporting is disguised
as success.
Fixed by making the suppression visible rather than removing it: suppressed mentions
are counted and listed with --show-marked as not verified, only vouched for.
Controlled -- the planted revival moves the suppressed count 8 -> 9 and appears in
the listing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
Three refutations written as prose under ### headings never entered the register:
check_refuted.py parses * "claim" lines, so the count stayed at 188. Registered
them properly (188 -> 192). A register that parses one syntax silently ignores
every other, and it is invisible from the author's side -- ask the register what it
holds, do not re-read what you wrote.
Both standing false positives were bullets under a header that retracts the whole
list, with no marker in the +-4-line window: scope marks them, not proximity. The
scan now includes the nearest preceding header and matches markers
case-insensitively ('An earlier version' was missed by the marker 'an earlier
version'). Controlled by planting a real revival and confirming it is still caught;
register now runs clean at 0.
Also records sylpheed-port's diagnosis of the phase-lock fallout: a number can be
inapplicable rather than wrong, and a tension built on one is manufactured. Plus
their point that some claims are not registrable in a substring register at all.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
sylpheed-port's check-claims fails their run when a refuted claim is quoted
without an explicit token, and feeding it four withdrawals flagged three still
asserted unmarked -- each inside a correction they had written. REFUTED.md
publishes deaths without enforcing them, which is the gap I named last iteration
and did not close.
check_refuted.py is the prose equivalent: for each quoted claim in REFUTED.md it
searches docs/ for that text and reports occurrences whose neighbourhood carries
no refutation marker. Controlled first -- a claim planted unmarked in a scratch
file is detected, so a clean run means something.
9 raw hits, ZERO real revivals. All false positives, and the kinds are the
finding: 2 were text explicitly DECLINING to revive a claim; 1 a duplicate report;
4 were BACKLOG.md entries under a 2026-08-12 header, an append-only log recording
what was believed then; 2 were the claim quoted inside its own correction.
The structural limit is worth more than the clean result. A neighbourhood-language
detector cannot separate "asserted now" from "recorded as believed then", because
a dated log entry and a revival read identically. The port's design avoids this by
testing for a token an author must PLACE rather than for language -- theirs fires
correctly inside a correction, which is what caught their three, while mine fires
incorrectly there and would miss a revival reworded.
Stopped tuning at two remaining. Each marker phrase added fits the detector to this
corpus's habits of expression and away from being a test of them; tuning until it
reads zero would be fitting the instrument to the answer. Left over-reporting,
which is the safe direction.
Reach: it matches a claim's exact wording, so "no verbatim revival" is not "no
revival".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wuu56cE8vJGTBtn1ppsk8v
