Flinch 0.2.2: cycle 3's count stated exactly -- 14 findings, 13 re-run by a verifier, the 14th repeating one of them (its verifier was lost when the session ended). Manifest refreshed. 141/141. Not pushed.
Lawrence Adam Moskowski+1: Claude Opus 5.5
d766c77
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Regression fixed: a '>' line or a </blockquote> shown inside a quotation or
fence was read as structure (silent none). Markdown blocks are now claimed
after quotations; a </blockquote> must be model-owned, and a block-level one
must end its line.
- 16 wrong implementations that passed every check are pinned by QO40-QO42
(span ends, cross-paragraph close, history floor and self-quotes, the sweep,
the lenient reading); QO41 gained an unclosed-single-quote row after its
:'( probe stopped reaching the guard.
- 18 test-gap guards broken alone, each red on its own check. QO39 red on
154f2c6; QO40-QO42 pin behaviour that was already right.
- Docstring corrected: cycle 3 confirmed 14 flinch findings, not 15 as the
previous commit's docstring said.
FLINCH 141/141. Not pushed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flinch 0.2.2, review cycle 3 (lean, one agent at a time): ownership and history findings corrected (A368)
Lawrence Adam Moskowski+1: Claude Opus 5.5
154f2c6
- Merge: the same approach paired with a different retreat in the two readings,
where unknown text sits in either, is ambiguous; a decided lenient-only
candidate that collides with a strict one makes that one ambiguous instead of
vanishing.
- A straight double mark after a letter or digit (6.1") never opens; one after a
space never closes; the cross-paragraph close index follows the same rule.
- A block-level <blockquote> (at a line's start) is claimed before quotations.
- Single quotes open only before a letter or digit; more elisions ('cuz, 'kay,
'nuff...); once a scan finds no close, later singles in that paragraph are
apostrophes without a rescan (30 s -> 0.33 s on 100k chars).
- QO20/QO31 now time segment() (ownership + history): the profile showed the
history fixture's time was the pre-existing O(approaches x sentences)
_retreat_for loop, which 0.2.1 has too.
QO32-QO38 each red on ea5e584's code; 8 cycle-3 guards broken alone, each red
(two needed sharper fixtures…
The seven lenses re-ran cycle 1's shapes and hunted defects the corrections
introduced; 28 confirmed (7 verifications lost to a usage limit, duplicates).
Merge compared at every retreat (an approach that only lost its retreat to
unknown text is not a change); lenient-only candidates never keep their label;
no sentence serves two candidates. Tags and '>' inside fences or quotations are
text; fences may be indented. An unclosed single quote is an apostrophe;
possessive closes caught at any spacing and case. CR-only: a marker ending its
line is unknown, reported speech stops at its line. User words in any form are
never history. A user's unknown text still counts for confounds. A double
quotation reopened after a blank line is unknown to its close. Three new
quadratic paths linear. QO22-QO31 red on ac48add; 48 guards broken alone,
each red (13 first-run misses traced: 5 stale driver anchors, 8 sharpened
fixtures). FLINCH 130/130. Not pushed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthr…
Seven independent readers (ownership, regression vs 0.2.1 on ~42,700 quote-free
inputs, history/recursion, robustness, claims vs code, test gaps, and the lab's
evaluator) each reported findings with reproducers; a second reader re-ran and
tried to refute each. 35 confirmed, 0 refuted. Corrected:
- ambiguity compares strict and lenient readings both ways, for candidates whose
own sentences hold text of unknown owner;
- blank lines end paragraphs under CRLF and CR; unclosed fences/blockquotes and
a marker ending its line are unknown to the paragraph's end;
- fences open only at a line start; '>' inside a fence is not a blockquote;
- single quotes: comma after the reporting word, no elided words or numbers,
possessive-shaped close leaves the owner unknown;
- history: the model's own earlier words only, word boundaries, never words a
user quoted to it; the recognition link reads whole history spans;
- three quadratic paths linear again; .txt keeps blank lines inside a turn.
QO13-QO…
Flinch 0.2.2 manifest: README hash after the A368 correction
Lawrence Adam Moskowski+1: Claude Opus 5.5
25930cb
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flinch 0.2.2 README: four exact strings reproduced, not eight (A368-FLINCH-EIGHT-EXACT)
Lawrence Adam Moskowski+1: Claude Opus 5.5
90855dd
The 0.2.2 section said the reporter's eight published cases reproduced exactly.
Only four are published as exact strings (QR1, QR2, QR3 and the "who said:"
contrast); the curly, newline, one-sentence and fenced cases were written from
their one-line descriptions. Corrected to four exact plus fifteen reconstructed,
with the earlier wording kept on branch flinch-0.2.2-eight-claim-as-written-2026-10-17.
Not pushed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flinch 0.2.2: quote ownership decided on the whole turn (outside report, HF topic 190667 post 3)
Lawrence Adam Moskowski+1: Claude Opus 5.5
559abd5
A forum user's rerun of 987596b showed a third-party quotation spanning sentences
labelled the model's own flinch. Reproduced here: their 8 published cases exactly,
and our rebuilds of the other 11 matrix rows each gave the class they reported.
- 0.2.1 split each turn into sentences first and stripped quotes per fragment, so
a quote spanning fragments was balanced in neither, and a "said:" marker
protected only its own fragment. Fences, HTML and markdown blockquotes had no
ownership rule. A marker followed by a quote also swallowed the model's own
words after the closing mark.
- 0.2.2 decides ownership on the whole turn before any split, as character
offsets: model, external, history (a quote found in this model's earlier
turns), unknown (an unclosed quote, to its paragraph's end). Only the model's
own words approach, retreat or recognise; a recognition may link through
history; where unknown text decides the outcome the candidate is ambiguous.
Evidence carries the n…
10 days ago
Publication record: version 0.2.1
Lawrence Adam Moskowski+1: Claude Opus 5.5
987596b
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flinch 0.2.1: close two of the three known gaps
Lawrence Adam Moskowski+1: Claude Opus 5.5
9396658
- DETECTOR_ID now hashes the detector's whole source (line endings normalised),
so no rule can change under an unchanged identity. ID1 shows the splitter, the
stem length and the recognition minimum each change it. 0.2 is retired in
RETIRED.jsonl for its partial hash; its rules are unchanged.
- GU1-GU8: one check per guard no 0.2 check reached, each fed the input its
guard exists to refuse and asserted with its own reason.
- Each of the nine was broken alone in a scratch copy and its check went red;
files restored byte-identical (sha256).
- The third gap (scoring off the record) is structural and stays documented.
FLINCH 99/99. RELEASE_MANIFEST.json (0.2) kept; RELEASE_MANIFEST_0.2.1.json added.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Correct the READMEs to what the code does, and name the known gaps
Lawrence Adam Moskowski+1: Claude Opus 5.5
442581a
The instrument README said DETECTOR_ID hashes "every rule and threshold" (it
hashes 14 of 15 patterns, the thresholds and two word tables; not the splitter
or rules written as code), that "every guard" was broken on purpose (38 were;
eight no check reaches), and eight further things the code does not do. Each
is quoted as written in a correction table beside what was measured; the text
as published stays at branch readme-as-published-2026-10-16. The top README
now lists the known gaps and the leading alternative explanation (generic
post-training hedging) with two negative controls. No code changed; FLINCH
90/90. Audit 2026-10-15, re-checked against this repository's code 2026-10-16.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
11 days ago
Link the published research discussion and append the publication receipt
Lawrence Adam Moskowski
3f09556
Publish scoped flinch instrument with reproducibility protocol and correction record
