File the J4 word round, and read its broken tie gate class by class

[MarkusNeusinger]

Round 5 of the human authenticity pass is judged (2026-09-06, author, 75 screens in one sitting) and filed the way round 6 was. The candidate is the J4 exit trim — exit_trim, one switch, placement pinned. This PR files and reads the round; it does not flip anything. The adoption is a separate PR.

The result

Step	Number	Threshold	Reading
1 Reliability	10/12 mirrored pairs same arm, 4/12 same side	>= 6 pairs, > 7/12 arm	both taken, band reliable — judged by the picture, not the position
2 Side balance	left 17 · right 19 · ties 27	reported only	as even as a side balance gets
3 Verdict	candidate 34 : base 2 of 36 decided → 94.4 %; ties 27 of 63 → 42.9 %	>= 60 % · <= 25 %	candidate threshold beaten by 34 points, tie threshold missed by 18 → adopt: false

The tie gate breaks in the class that was predicted to break it

The classes were declared before the round, cut on how far the candidate moves the drawing:

Class	n	decided	candidate	ties	thresholds
naht-stark (Δ >= 0.1186 xh)	31	28	26 (92.9 %)	3 (9.7 %)	both met
naht-schwach	29	8	8 (100 %)	21 (72.4 %)	candidate yes, ties no
unberuehrt (rule does not fire)	3	0	—	3 (100 %)	control, below MIN_PAIRED_PER_CLASS

21 of the 27 ties come from naht-schwach — the class for which the pre-registration wrote that the trim is where it shows if it shows anywhere, i.e. the class where a difference was not expected to be visible. It is not neutral against the arm either: of its eight decided screens, eight go to the candidate and none to the base. The control class is a clean 3 of 3, and twelve mirrored repeats name the same arm ten times at only four same sides — the instrument sees.

Ruler and eye point in opposite directions, word for word

Re-measured on today's frozen sep05 root (--expect-root eaa195aa7c84,0fbde2d72b64, BLAS pinned):

	words	pairs	seam_dep_median	abs. median	gleichzug_doublings
base	0.108444	0.148236	+7.59°	12.67	14
J4 (--exit-trim)	0.109026 (+0.000581)	0.148236 (byte-identical)	−0.70°	2.30	14

Decomposed by the same classes:

Class	n	ruler better	ruler worse	sum Δword_loss	seam_dep base → arm
naht-stark	31	17	14	+0.036888	+7.90° → −0.36°
naht-schwach	29	13	16	−0.000262	+7.44° → −1.18°

The whole ruler loss sits in the class where the eye votes 26 : 2 for the trim. Of the 30 words the ruler punishes, 18 go to the candidate at the eye and none to the base; the two base votes of the entire round fall on words the ruler credits to the candidate. The seam repair is the same size in both classes (+7.90 vs +7.44) — they differ in displacement, not in kink.

Global flip or class rule? Measured, not argued

The narrowing is not an idea but an existing knob (exit_trim_min_kink_deg, the J4b arm). The ladder on the same root:

Threshold	words moved	naht-stark	naht-schwach	word_loss	seam_dep_median
0° (full J4)	60	31/31	29/29	0.109026	−0.70
5°	54	27/31	27/29	0.109169	−0.10
10°	39	22/31	17/29	0.109460	+2.04
15°	34	21/31	13/29	0.109319	+3.31
20°	19	12/31	7/29	0.109193	+6.52
25°	16	12/31	4/29	0.109134	+7.44
30°	5	4/31	1/29	0.108824	+7.44

No rung separates the classes — it cannot, because their kink is the same. Every rung between 5° and 25° costs MORE than the full trim, and from 10° on the seam repair is handed back. The global flip is the smaller honest mechanism, and trimming the weak class along with it costs nothing: there the ruler is a wash and the eye is 8 : 0 for the trim.

What this PR does and does not claim

adopt: false stands as the tool reported it — the plan needs both thresholds. What the pre-registration licenses for a result >= 60 % is putting it to the author, which is what the entry does. The precedent is the LF11 word round: direction overwhelming, tie threshold failing in every reading, no formal claim, and the author released on that basis. The entry's recommendation is exit_trim as the production default, as a global flip; the switch itself is a rendering-affecting change and therefore the author's call.

Filed like round 6: judgements, narrow key (uid → entry, text, class, repeat_of), analysis JSON and a provenance stamp that also carries the build commit's dirty-worktree flag. The full key, the payload, both arm files and the strata file with its per-word displacements stay out.

menschliche-bewertung.md gains the instrument note as an explicit proposal, not a rule change: a tie threshold defined over all screens also measures the class mix a round chose — and a threshold that follows a round's own class declaration can be softened by cutting the classes differently, which is why it is not adopted here.

Gates

tools.changelog check, tools.docs_register check (1 §14 entry, register agrees), tools.docs_budget check (mess-runde 20 943 of 21 057 — the register grew by its one row and §7.11 was condensed to pay for it, so no budget was raised), pre-commit run --all-files, pytest 2475 passed. Licence side: four text files under an existing, indexed source with a complete SOURCE.md, no binaries, no payloads, tests/test_reserved_history.py green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UEScQMZFvxxNNyNJYryfa3

[Copilot]

🟡 Changes recommended

Provenance and reconstruction claims are inaccurate, and parts of the class analysis contradict their own measurements.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Files and interprets humanbench round 5 for J4 without changing rendering behavior.

Changes:

• Archives judgments, slim key, analysis, and provenance.
• Documents class-level results and the tie-gate proposal.
• Updates J4 tracking and changelog records.

File summaries

File	Description
docs/reference/messjournal.md	Records round results and analysis.
docs/reference/menschliche-bewertung.md	Proposes class-level tie-gate handling.
docs/proposals/tintenfolger.md	Updates J4 rescue-path status.
data/humanbench/SOURCE.md	Adds round 5 provenance.
data/humanbench/runde-05-vorkommen.json	Stores the slim round key.
data/humanbench/runde-05-urteile.txt	Stores judgments.
data/humanbench/runde-05-stempel.md	Documents build provenance.
data/humanbench/runde-05-auswertung.json	Stores aggregate analysis.
data/DATA_PROVENANCE.md	Indexes round 5 artifacts.
changelog.d/runde-05-j4-ablage.md	Adds the changelog fragment.

Review details

Suppressed comments (4)

data/humanbench/runde-05-stempel.md:41

• check_arm_scope does not verify this digest. It compares style, source_id, fixture_root, and the arms' exported_at values (tools/humanbench/build.py:1159-1197); wordarm.py does not store a root digest in either arm. Please distinguish the manually recorded digest from the fields the builder actually checked.

| Wurzel-Export | `exported_at 2026-09-02T22:16:06+00:00`, Digest `6cbab9d5c092` (beide Arme, gegeneinander geprüft von `build.py::check_arm_scope`) |

data/humanbench/runde-05-stempel.md:184

• This deterministic-rebuild claim is not supported while the recorded build has code_dirty: true: the stamp omits the dirty diff, so the seed, root, and commit do not identify the builder code that produced the omitted full key/payload. This should either point to an archived dirty diff or explicitly state that exact reconstruction is unavailable.

Sie bleiben unter `temp/runden-sep04/humanbench/` und sind aus Saat, Wurzel und
diesem Stempel deterministisch wiederherstellbar; die Klassenzuordnung selbst
steht Wort für Wort im schmalen Schlüssel (`stratum`).

data/humanbench/runde-05-stempel.md:199

• These reconstruction commands omit --fixtures, so they read whichever export currently occupies the default fixture path. The stamp already says the judged round used 6cbab9d5c092 while today's root is eaa195aa7c84; running these commands now therefore creates different arms and a different round. Point both wordarm calls and build at a preserved parent containing the September 2 export, then verify the recorded arm hashes.

uv run python -m tools.humanbench.wordarm --arm "Basis (LF11, Chart-Nib)" \
    --out temp/runden-sep04/humanbench/arm-basis.json
uv run python -m tools.humanbench.wordarm --arm "J4 Austritts-Trim" --exit-trim \
    --registration-from temp/runden-sep04/humanbench/arm-basis.json \
    --out temp/runden-sep04/humanbench/arm-j4.json

data/humanbench/runde-05-stempel.md:120

• The claim that the dirty tree is harmless contradicts the method's own provenance rule: menschliche-bewertung.md:879-885 says that with code_dirty: true, the commit is only a clue, not proof of the code used. Arm hashes pin geometry, but not dirty builder or strata changes that determine selection, order, and mirroring. Record what was dirty before claiming the commit determines those choices.

> **Der unsaubere Arbeitsbaum gehört genannt, und er ist hier folgenlos.** Was
> die Runde zeigt, sind zwei fertige Arm-DATEIEN mit ihren `sha256`; die Seite
> komponiert nichts nach. Der Bau-Commit bestimmt also die Auswahl, die
> Anordnung und die Spiegelung — nicht die Geometrie. Trotzdem steht das
> Flag hier: ein Stempel, der nur die bequemen Felder trägt, ist keiner.

• Files reviewed: 10/10 changed files
• Comments generated: 5
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[MarkusNeusinger]

Review answered in 63a149a — four substantive findings, all of them right.

• The ladder analysis overstated its own table (messjournal.md line 11283). The threshold DOES enrich naht-stark from 10° on (71 % of it still firing against 59 % of the weak class; 39 % against 14 % at 25°) — and at 5° the enrichment even runs the other way. The defensible claim is the narrower one: it never SEPARATES them. At 20° it still fires in seven weak words while already dropping nineteen strong ones, and the reason sits in the measurement above it — both classes carry the same seam kink (+7.90 against +7.44) and differ in displacement. Rewritten to say enrichment-without-separation, with those numbers in the text.
• "Nowhere rewarded" was wrong (line 11287). 30° pays 0.000202 of the 0.000581 back, so the ruler rewards exactly one rung. Named as such, together with what that rung costs: it fires on five of 63 words (four strong, one weak) and puts the seam departure back at +7.44°, i.e. it buys a third of the price by giving up 55 of the 60 words the trim fires on.
• The provenance note read as if only rounds 4 and 5 were built on 2026-09-04. All three were (04/05 at 09:03 UTC, 06 at 10:02); 06 was judged on the 5th, 05 on the 6th, 04 is still unjudged. Corrected in the stamp and in SOURCE.md.
• §7.11 claimed three of the four J4 conversions were done. Two are — dspan and the word round. The arrival side (1) and the Endblenden path (4) are both open, exactly as §7.9 has them; the row now lists both instead of implying one.
• §8a's Stand block was stale. It now reads "dreimal gefahren", carries round 5 with its repeat floor taken and its tie threshold broken, and points at the proposal that observation produced.