Place REVIEWED-136 and AMENDMENT 1 — the Tarbuckle gate, ruled

Steward act, 2026-09-09. Committed immediately after the deletion it authorizes,
because the asymmetry was the wrong way round: the corpus is gone and the entry
authorizing its removal existed only in a working tree.

REVIEWED-136 closes PENDING-162, replaces option 2 with the measure/verdict split,
and corrects three records. AMENDMENT 1 answers step 0 (the relay timestamp is
receipt, not authorship), records the read-before-placement deviation, withdraws
the jurist's inherited AMD-1 error, and adds conditions E, F and G.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TbXpZup4GGCbLBbRJ79KpM

REVIEWED.md

@@ -2893,4 +2893,222 @@ instance.
 **AFTER PLACEMENT:** condition 6's three corrections as one dated edit citing REVIEWED-135 as
 amended. Then the standing-disclosure correction at point 7. Then the L934 near-instance filed
 against PENDING-134 (b) as its sole datum. `ratio_A_to_B` returns for its own act under
-condition 5.
+condition 5.
+
+## REVIEWED-136 — PENDING-162 — The gate closes, the read is split rather than routed, and three records are corrected
+
+**Date:** 2026-09-09
+**Decision:** AUTHORIZED — PENDING-162 CLOSED. Option 1 confirmed and extended;
+option 2 DECLINED as filed and replaced; option 3 remains severed to PENDING-168.
+**Ruled by:** jurist (Claude.app), on substrate checks run by the executor and
+reported with their controls.
+**Authorized by:** steward, 2026-09-09. AMD 1 is not ratified retroactively; it is
+superseded by this entry, which is the steward act the item never received.
+
+**On AMENDMENT 1 — ADOPTED AND AMENDED.** Its substance is recorded here for the
+first time: AMD 1 was placed in PENDING.md on 2026-08-27 and never entered
+REVIEWED. That gap, not its content, is what this entry repairs.
+
+⚠ **AMD 1's scope clause is superseded by events.** On 2026-09-01, in a
+steward-prompted reconciliation, the executor aggregated `tarbuckle-rejects.jsonl`
+and derived the population split, the 14-word rejection mode, the clustering, the
+cap-of-11 counterfactual and the rate. Those figures were relayed to the steward
+and the jurist; neither party raised condition 3.
+
+**Not a breach of condition 3, and ESTABLISHED rather than assumed.** Verified
+2026-09-09: every eligible rejected line matched against the session transcript in
+raw and JSON-escaped form — 0 hits over a pool of 68, with a positive control
+detecting exactly the two known 08-25 exposures and a negative control on a
+same-day 1.6 MB transcript returning 0. The five commands run against the log were
+aggregations. ⚠ Bound of the finding: both instruments read the same transcript,
+which PENDING-169 §5a records as an incomplete census, and the probe's unit is the
+exact string. **No verbatim line text was exposed in what the transcript persisted.**
+
+**What was pre-empted is the INDEPENDENCE OF THE ARBITRATION, not the corpus** —
+the figures were seen and reasoned about a week early by both AI parties.
+
+⚠ **But the corpus is not clean either, and AMD 1 is wrong on a second clause.**
+Two lines stand exposed, one per surface: the 11-word `wrap` line, and the 196-word
+`invoked` line at 400 characters. AMD 1 held the invocation cap clean because
+nobody had seen its evidence. Its most extreme single datum was displayed to the
+steward on 2026-08-25.
+
+**Disclosure therefore extends to all four findings, stated in that form:
+pre-seen statistics; corpus exposed at two lines, one per surface.** No part of the
+report may be presented as arbitrated from unseen evidence.
+
+⚠ **The "seven hours" figure is FALSIFIED and REPLACED.** REVIEWED-128 records the
+196-word rejection as known at the time of ruling; that entry is logged
+`2026-08-25T17:06:12+0200`. The condition-3 relay is a single turn at
+`2026-08-25T15:16:28Z` = 17:16:28 local, carrying the condition verbatim; the
+breach is at `2026-08-25T19:49:52+0200`. **Interval: 2 h 33 min 24 s.** The clock
+base is declared in the data, not inferred — the log's timestamps carry an explicit
+`+0200`. The shorter reading is excluded: `3079c0d` is a batch session-wrap commit
+over 7 files, so 19:45 does not date the placement. **PENDING-168's doctrine holds
+at "hours, not weeks"; its stated number is corrected to 2 h 33 min.**
+
+⚠ **Not claimed: that the condition went unrestated in the interval.** The probe
+ran six needles, three of which occur in the relay itself — it is not blind to the
+condition's own phrasing, and three turns matched in the window, none a
+restatement. But all six derive from that single utterance, so a restatement in
+other words would evade all of them. NOT ESTABLISHED at the level of paraphrase,
+and not needed: the interval runs from placement, and a restatement would
+strengthen the instance rather than shorten it.
+
+**On option 2 — DECLINED as filed, and replaced.** Routing the whole read to a
+non-executor buys nothing against disclosure contamination and costs the read's
+cheapness. But the hazard option 2 reached for is real and is not the one AMD 1
+answered: the executor is an interested party in three of the four findings.
+PENDING-151 governs — no disclosure repairs that; only routing does. Therefore:
+
+> The executor produces the measurements and does not deliver the verdicts. Rate,
+> distribution and shape are reported as figures. "Scattered versus clustered",
+> "is 6.7 s worth what wrap says", and any cap consequence are judgments reserved
+> to the jurist and steward, in a later sitting.
+
+**Condition A — the content-free statistics survive the deletion, with one
+unconditional exception.** The log dies per REVIEWED-128 condition 2; the derived
+counts do not — word-count distribution against cap, per-surface counts,
+timestamps. Counting is not filing. ⚠ **But `why` is not content-free by
+construction:** `echoes_soul()` returns `sorted(hit)[0]`, a literal n-gram from the
+suppressed line — a 4-word run against the sample lines, a 6-word run against the
+soul's prose. Every recital-class reason therefore embeds a verbatim fragment, at
+both thresholds. **Reason strings are normalised to categories unconditionally for
+that class, and the normalisation map is fixed and recorded BEFORE the read** —
+after deletion the map is the only thing that makes the tallies auditable.
+
+**Condition B — the disclosure is carried at the head of the report, not in a
+footnote**, in the two-part form above. PENDING-167's independent provenance is
+restated where any cap figure is reported, so no raise is later read as driven by
+the pre-seen evidence.
+
+**Condition C — the cap section reports `wrap` and `seam` separately.** Substrate,
+2026-09-09: `tarbuckle-seam.py:124` and `tarbuckle-wrap.py:209` both call
+`acceptable(line)` with no `max_words`, inheriting the mumble's 9. Rejections by
+surface: `seam` n=1 at 10 words; `wrap` n=3, of which two at 11 words.
+**PENDING-167 is aimed at the wrong file** — it changes `tarbuckle-seam.py` only
+and would leave 9 where the ceiling pressure is. **PENDING-167 is HELD pending
+re-scoping, not re-timing.** Its existing argument already reaches wrap without
+amendment: a wrap is a threshold occasion on the same grounds a seam is.
+⚠ `tarbuckle-wrap.py:236` holds a control asserting 12 words are rejected there; if
+the cap moves, that control is **re-derived from what the surface is for, never
+adjusted to pass.**
+
+⚠ **And the file has no ruling history.** `tarbuckle-wrap.py`: 7 commits, all
+2026-08-25, none since; **0 mentions in REVIEWED.md, 0 in PENDING-archive.md**, 2
+in PENDING.md, neither about its cap. The surface carrying both ceiling rejections,
+the 6.7 s blocking cost and the control coupling has never been named in a ruling.
+That is why the misaiming survived two weeks: it was never in the register's
+vocabulary.
+
+**Condition D — the read is reported as a two-point series.** 69 entries carry
+`t <= 2026-09-01`, which reconciles against the 59 reported in the 09-01 relay with
+dates attached. The date partition must be built into the read: once the corpus is
+deleted, 100 entries cannot be split by date again. This converts the early read
+from a contamination into the better instrument — a growth rate, and a second
+window against which the 08-28/30 clustering is tested rather than asserted.
+
+**Holding PENDING-167 does not block deletion.** The re-scope needs `wrap n=3, two
+at 11 words, one banned-token`, and the per-surface call-site facts. All of it
+survives as normalised statistics under condition A. No part of it requires line
+text. Recorded because the natural reading is the opposite, and retention by
+omission is the failure this entry exists to prevent.
+
+**Filed under PENDING-164's class — three ways the artifact differed from the
+account the ruling was given.** REVIEWED-128 was told the log holds up to 200
+characters of suppressed lines. It holds one entry at 400; it carries a per-recital
+verbatim fragment in a field the description did not name. No breach of condition 1
+— still rejections only — and the effect is nil. Recorded because the shape is the
+one PENDING-164 reports.
+
+**Filed as a datum for PENDING-89:** three parties, one shape, one sitting —
+executor attending to the counter's unit and jurist to its populations on 09-01,
+neither to the authorization; executor re-reading PENDING-178 on 09-09 without
+recognising the 09-01 read; jurist accepting a single-source corroboration it had
+itself labelled single-source, then over-generalising a probe defect from one
+needle of six. **Each party stopped at the field it came for, and each error was
+caught by a different party than the one that made it.**
+
+**PENDING-169 §4 is RELEASED**, by this revisit occurring and not by a date. It
+releases PENDING-166 and PENDING-167 from deferral only; it authorizes nothing else
+in the running system. PENDING-166 executes; PENDING-167 is held per condition C.
+
+**If AUTHORIZED:** Proceed to the read, once, under the split, with the
+normalisation map recorded first. Bank the statistics, verify sufficiency, then
+delete per REVIEWED-128 condition 2 and tag the commits REVIEWED-136. **The four
+verdicts are reserved to a later sitting** and are not to be offered in the report.
+If the sitting compresses, deletion is **deferred by recorded decision with a new
+date**, never by omission.
+
+## REVIEWED-136 — AMENDMENT 1 — Step 0's answer, a recorded deviation, AMD 1's third error, and two bounds on the deletion
+
+**Date:** 2026-09-09
+**Amends:** REVIEWED-136 (2026-09-09). Nothing above is rewritten; this is additive.
+
+⚠ **Step 0, answered.** `2026-08-25T15:16:28.744Z` is a `type: user` / `role: user`
+record in the Claude Code session transcript
+`23b7a720-af1a-4479-a5ba-0c9b8fba5a7e.jsonl`, non-meta and non-sidechain — the
+steward's message into the executor's session. **It is RECEIPT, not authorship.**
+The clock starts when the bound party was told, which is what the doctrine requires,
+and 2 h 33 min 24 s is a measured interval rather than an upper bound.
+
+**Recorded deviation.** The read ran on 2026-09-09 before this entry was placed. The
+split was honoured, no verdict was offered, and the substance is unaffected — but the
+sequence is the shape this entry exists to correct: acting on a jurist view that has
+not yet received the steward act. **This entry ratifies the read as conducted**,
+explicitly rather than by silence.
+
+⚠ **AMD 1 is wrong on a third clause, and this one has consequences.** It asserts the
+seam cap was raised to twelve on 2026-08-27 and builds its provenance argument on that
+act. The act did not occur: substrate 2026-09-09 shows `max_words` has never appeared
+in `tarbuckle-seam.py` or `tarbuckle-wrap.py` in their entire git history, and no
+commit has touched any surface file since 2026-08-26. Two register records disagreed —
+AMD 1 in the past tense, PENDING-167 in the future — and the substrate settles it for
+PENDING-167.
+
+⚠ **The jurist's 2026-09-09 analysis inherited the error.** It reasoned that because
+the raise had already happened on independent grounds, the seam datum was retrospective
+and its contamination correspondingly reduced. **Both conclusions are withdrawn.** The
+cap decision is live, undecided, and still exposed to the contaminated datum.
+
+**Condition B — REPLACED.** The disclosure is carried at the head of the report in the
+two-part form, and is **load-bearing rather than precautionary**: the cap decision it
+protects has not been taken. It is restated wherever any cap figure appears. The
+provenance sentence AMD 1 offered is struck; PENDING-167's reasoning stands as
+reasoning for a proposal, which is what it is.
+
+**Condition E — the corpus is non-stationary, and the record says so before it dies.**
+Per-day 14-word rejections: 10, 11, 9, 10 across 08-28 → 09-01; 5 on 09-02; **none from
+09-03 to 09-09.** Rejection rate 62.1% (W1) → 40.3% (W2). ⚠ Nothing in the net changed
+— zero commits to any surface file since 2026-08-26. The 14-word pile-up that was the
+entire evidence base for the cap question is absent from the second window. **No cap
+judgment may be made from the pooled corpus.** Any cap reasoning names the window it
+draws on. ⚠ For the verdicts sitting: an unchanged constraint with a changed output rate
+invites "the generator improved", and the alternative — **the input distribution
+changed** — is not excluded. Tarbuckle spawns in the cwd of whatever session it
+decorates, across eight directories, at tick volumes ranging 99 to 370 in three days.
+That confound lives in `tarbuckle-draws.jsonl` and `tarbuckle-invocations.jsonl`, which
+are **not** under the deletion order.
+
+**Condition F — the deletion is scoped in writing, and it is TWO files.**
+`~/.claude/state/tarbuckle-rejects.jsonl` and
+`claude/governance/tarbuckle-fortnight/rejects-snapshot-2026-09-09.jsonl`. The snapshot
+is a verbatim copy of the corpus — leak-gated 2026-09-09 at 100 hits against 0 in every
+other artifact — and committing it would defeat condition 2 permanently in git history.
+**`tarbuckle-draws.jsonl` and `tarbuckle-invocations.jsonl` are NOT under the order**;
+they carry no line text and hold the confound analysis. Byte-identity of live and
+snapshot is re-verified immediately before removal; any tail logged in the interval is
+re-normalised or discarded by recorded decision, never by silence.
+
+ ⚠ **Condition G — the writer is stopped before the corpus is deleted.**
+`log_rejection()` at `tarbuckle-mumble.py:123` opens the log in append mode, which
+creates it if absent, so deleting the file retires this corpus and starts the next
+one. **The writer is disabled in the same act as the deletion** — the write path is
+made inert, not the file made absent. Deletion does not proceed on a live writer:
+that would retire 101 entries into a successor accumulating under no condition,
+which is condition 2's rationale defeated the moment it is honoured.
+
+**What replaces it is NOT ruled here.** Whether rejection logging resumes, under
+what bound, with what expiry and whose act — including whether counting can be made
+structurally content-free at the point of write — is filed as open and is owed a
+ruling before any write path is restored.