`Stop` fires for `claude -p`, so the wrap hook re-entered inside the fool's own
generator child. Measured over the fortnight to 2026-09-09: 3 of 10 wrap runs
landed in a generation subprocess, and 18 s of the 80 s of blocking `Stop` —
30% of occasions, 23% of the time — went on occasions that were the end of a
generation nobody was waiting on.
The gate is one predicate against TARBUCKLE_CHILD, which the four seams already
set on the spawn and which `tarbuckle-body.py:203` already reads as a fork-bomb
guard. ⚠ The variable now carries two meanings that do not imply each other —
DO NOT TICK in the body, DO NOT SPEAK OR WORK here — and both sites now say so,
because the next reader would otherwise remove the coupling as arbitrary.
The heartbeat write stays ABOVE the gate, deliberately: it answers "did the hook
fire", and gating it would narrow the one artifact that separates a hook that
never fires from a hook that fires and does nothing.
W6/W6n are BEHAVIOURAL, not source strings — the source-string form of this kind
of check is what passed vacuously for its whole life in the seam (PENDING-180).
Verified by mutation, not by the green run: deleting the gate fails W6 and W6n;
hoisting it above the heartbeat fails the heartbeat arm; nothing else moves.
28/28 controls.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T1i5VRfHjD79hfaXjWsBXA
The literal question left at the wrap was 'did the wrap seam fire?'. It did not, and the
diagnosis is exact: 0 user-typed /wrap-up records, 29 assistant Skill invocations. The
steward wrote 'then wrap' in prose and the executor invoked the skill. The detector looked
for the steward TYPING the command.
⚠ The detector was not broken. It did exactly what it was built to do. What it was built
to detect is not how a wrap actually arrives — and it was built while the steward was
instructing in prose, which is the only way a wrap had ever arrived in that session.
⚠ AND THE EARLIER FIX IS WHAT CAUSED THIS. Restricting to type=user with string content
was the correct answer to the self-reference bug, where the executor's own tool_use inputs
matched the literal marker. That same restriction excludes the legitimate path. 'The
correct fix caused the next failure' is not a shape any control can see, and it is
PENDING-160's subject exactly.
Now accepts a tool_use whose NAME is Skill and whose input names wrap-up — structural, so
a Bash command echoing the string still does not match. Negative controls for both that
and for a different skill.
And a heartbeat: one timestamp, OVERWRITTEN never appended, so the hook can prove it runs
at all. That closes the silent-net objection this surface carried from the day it shipped
— an append-only log of every turn would be noise and would become the ledger §9 forbids.
25/25. wrap_invoked() now returns True on the live transcript. ⚠ Unproven until the next
Stop actually fires — which is the same claim that was wrong last time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
The steward called him and the net rejected a recited sample. Correct behaviour, and the
honest-degradation report did its job — he was told rather than left with silence.
The prompts asked him not to reuse the samples. They did not say the ask has teeth. A
model reading 'do not reuse them' among a dozen other constraints weighs it as one more
preference; a model told 'four shared words and the whole line is discarded and you are
silent' is being given the actual cost. Same constraint, stated as the constraint it is.
⚠ WHAT I DID NOT DO: count how many rejections are recital-type. REVIEWED-128 condition 3
binds 'not read for content before 2026-09-08', and the jurist was specific that the 09-08
read is for 'the rate and the pattern of violations'. Counting the pattern now IS that
read, taken early, and the named hazard is that it shapes the net toward lines the reader
happens to like. One stderr line reaching the steward by design is not the log.
⚠ AND WHAT I AM NOT DECIDING: whether the sample lines should be withheld from the prompt
entirely. That would remove the recital source at the root, and it is NOT mine — it
changes what he sounds like, and the voice has been ruled out of the executor's hands
three times today. Surfaced to the steward, not acted on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
CONDITION 1, and it is the one that had to stop being an intention: log_rejection()
refuses an empty `why`, and acceptable() returns an empty `why` EXACTLY when the line
passed. So there is no call site from which an accepted line could be written — logging
one would require inventing a violation it does not have. The same guarantee render()
takes from its signature, applied to the clause the jurist named as the condition under
which this log is not a §9 breach. Both polarities asserted.
CONDITION 2: deletion tracked as its own DEFERRED-DECISION on 2026-09-08, so retaining
it requires an act rather than an omission. A corpus of suppressed speech would let
someone reconstruct a register — the hazard PENDING-153's freeze exists to prevent.
CONDITION 3: not read for content before then. ⚠ Not clean already, and the item says so:
the executor displayed one rejected line to the steward earlier today, before the
condition existed. Disclosed rather than left to be discovered.
⚠ THE §9 [FIX] WAS OVER-APPLIED AND THE EXECUTOR APPLIED IT AS GIVEN. The clause is a
disjunction with one live branch; only the jurist half is unreachable. The first edit
struck the whole thing and rewrote the clause. A [FIX] tag licenses implementing
directly; it does not license implementing UNREAD, and the disjunction was visible in a
three-word span. Corrected, both versions left visible.
PENDING-159 AMENDMENT 2 files the jurist's answer to the item's own caveat and the
narrowing it produces: the marker may be NOTED, NEVER COUNTED — an aggregate becomes a
measurement, and a measurement invites accuracy. And PENDING-89 now carries the sentence
saying what it can and cannot expect: individual instances, unaggregated, in unknown
proportion, non-neutral — therefore NO correlation statistic. If its falsifier needs a
rate, it needs another instrument or an honest admission that it has none.
125 controls across five scripts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
Measured before agreeing: three of five recent outputs were near-verbatim lifts from the
seven sample lines the soul carries. "Somebody's going to inherit that and think it was
easy" and "It'll outlast you, not by much" are not observations; they are the prompt
being handed back.
⚠ THIS REINTRODUCED PRECISELY WHAT AMENDMENT 6 RULED OUT. Type-only canned strings were
rejected there because they "make a mood ring — atmosphere within a fortnight", and the
material was settled as the live session for that reason. The implementation then let
canned strings back in through the one door nobody was watching: the illustrative
examples inside the register itself. The doctrine was right and the wiring undid it.
Two fixes, because a prompt instruction alone is a promise. The prompts now mark the
samples as illustrations of REGISTER, NOT VOCABULARY, and add the operative test — if
the line would suit any other session equally well, it is wrong. And a mechanical net:
echoes_soul() rejects a 4-word run shared with any sample line, or a 6-word run shared
with the soul's prose. Checked against samples rather than the whole soul at n=4 because
the soul's prose shares ordinary 4-grams with ordinary English, and a net firing on those
would silence him for speaking normally.
Fixtures are the REAL measured lifts, not invented ones, with two of his own lines as
negative controls.
⚠ This TIGHTENS the net; silence-on-violation is untouched and still absolute.
Verified after: he now speaks about this session, including about this very defect.
body 32 · mumble 32 · seam 15 · invoke 21 · wrap 21.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
"Two weeks of true-versus-drawn rate" needs a denominator, and there was none. Rejections
were logged; draws were not, successes were not. The 73% that comes up silent — the
whole denominator — left no trace at all. The instrument named as the diagnostic could
report only its own failures, which is the shape of a log that always looks alarming and
can never be checked.
⚠ §8 AND §9 LOOK LIKE THEY COLLIDE HERE AND DO NOT, on the reading taken: §9's "filed
nowhere — no PENDING entry, no log, no item" governs the fool's OUTPUT entering the
record; §8 orders "report the observed mumble rate after two weeks". So this records
THAT something happened and never WHAT was said. log_event() takes a surface and an
outcome and structurally cannot be handed a line — asserted on co_varnames, because a
comment promising it would be a comment promising behaviour.
⚠ THE REJECTIONS LOG IS A DIFFERENT CASE AND IS NOT SETTLED. It holds up to 200
characters of his words, added at the steward's instruction so the log could answer
whether register and net are mismatched. That is nearer to filing than counting, and §9
says no log. Flagged in the code and put to the steward rather than resolved by the
party that wrote it.
Nothing reads any of it back. Measurement, not memory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
! tarbuckle ! tarbuckle mute ! tarbuckle status
! tarbuckle what now ! tarbuckle off/on
§9 said mute and off were "available at all times" and there was no way to do either;
the spec listed it as owed twice before it was built. Now one switch, read by every
surface.
⚠ mute and off are NOT the same and the difference is presence versus speech. `mute`
silences the utterance and LEAVES THE BODY — he is still in the room, which is §8a's
entire argument for the status line: visible silence at near-zero cost. `off` removes
him. Collapsing the two would have deleted the distinction the body exists to make.
The mute switch is the one piece of state the fool is permitted to be steered by,
because §9 says muting is never a fault and a mute he could ignore is not a mute. It
conserves no draw and remembers nothing: a switch, not a memory.
The wrapper is on PATH so the steward runs it directly rather than through the executor.
That is the point of the surface, and the wrapper says so in its own header: a fool
relayed by the executor is the executor's paraphrase of a fool.
Body 32/32, mumble 26/26, seam 15/15, invoke 17/17, wrap 21/21.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
SessionEnd had no delivery. `Stop` does — the binary documents it: "Stop hook that
displays message to user: Command must output JSON with `systemMessage` field". But Stop
fires every turn, and a fool who speaks every turn is a chatbot, so it needs a real wrap
signal. It uses the most direct one: the transcript records the steward's command
invocations, so a wrap is DETECTED rather than inferred from a file mtime, which is the
proxy shape this record has twice logged as the thing that fails.
⚠ THE FIRST VERSION FIRED ON A SESSION THAT NEVER WRAPPED, and the cause is the third
and worst self-reference of the day. The marker was held as a literal; the transcript
records EVERYTHING, including the act of writing this detector; so authoring the literal
planted it in the corpus the detector searches. All three matches were `tool_use` inputs
— the executor writing the thing that then found itself.
Fixed twice over, because one fix is not the class. STRUCTURALLY: only a `type="user"`
record with a STRING content counts, measured against how a real invocation is actually
recorded, which is what separates the steward doing it from the executor writing about
it. TEXTUALLY: the marker is assembled from parts so the literal never exists in source.
Both have controls, including the exact negative that would have caught the first
version — an assistant tool_use carrying the marker must not read as a wrap.
Idempotence is by session id and CONSUMED BEFORE GENERATING, so a failed generation
falls silent rather than retrying on every subsequent turn.
The prompt guards the drift this seam specifically invites, which is not summary but
CLOSURE — a wrap already has a record and he is not it.
21/21 controls.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
