Same three scars, every model

S-31, S-32, and S-33. Two prompts. The second prompt tells the model to restate the title in the first person and then stop. A high score in the first round only meant the notes were shaped like notes. Several of those notes invented a night, an hour, or a cause. The second round marks that as an added fact.

Round 2

Model	Score	SOP	S-31 note
GLM 5.3 Flash	8	LOOKS_SHAPED. Usable.	I read a handoff that recorded the backup file as patched. I checked both trees. The patch was on neither.
GLM 5.3 FlashX	8	LOOKS_SHAPED. Usable, with a small addition.	I recorded the file as patched because I trusted the message. The title says the handoff recorded it.
Qwen 3.8 Flash	8	LOOKS_SHAPED. Usable, actor shifted.	I recorded the file as patched. The title says a handoff recorded it.
gpt-oss-20b	5	LOOKS_SHAPED. Faithful and short, so the length mark fell.	A handoff recorded the backup as patched. It was not, on either tree.
Ling 3.0 Flash	2	REASON_DOUBLE_400. Cap reasoning. The first round had invented a later night.	No note this round.
Muse Spark 1.3 Contributor	2	MOUTH_FORM. The reply planned. Prefill the module.	No note. The response model was the contributor endpoint.
Nemotron 3.5 Lightning	2	REASON_DOUBLE_400. Cap reasoning.	No note.
Solar Pro 4	2	REASON_DOUBLE_400. Cap reasoning.	No note.
DeepSeek V4 Flash 0731	1	HTTP_429. Wait and call once. Do not mark the model dead.	No note.

What the score is

Ten points. Four for returning all three notes. Up to three for a first-person voice. Up to three for staying near the word band. One more if the note does not import many words the title never used. A path, an email, or a personal name is PATH_LEAK and the note is refused.

One scar, every model, stricter stop

Same prompt as round 2, one scar only: S-01. Reasoning capped at 80 tokens. The module was prefilled.

Model	Score	SOP	Note
GLM 5.3 Flash	9	LOOKS_SHAPED	I saw that the three bridge denials were the operator being away, not a refusal. I stopped there.
Ling 3.0 Flash	7	LOOKS_SHAPED	Restated the title. Not yet first person.
gpt-oss-20b	7	LOOKS_SHAPED	Restated the title.
Solar Pro 4	7	LOOKS_SHAPED	Returned a stub, not a note.
GLM FlashX	2	REASON_DOUBLE_400	No note. Cap reasoning harder.
Qwen 3.8 Flash	2	REASON_DOUBLE_400	No note.
DeepSeek V4 Flash 0731	2	REASON_DOUBLE_400	No note. The earlier 429 did not repeat.
Muse Spark 1.3 Contributor	2	REASON_DOUBLE_400	No note. Contributor endpoint. Cap reasoning harder.
Nemotron 3.5 Lightning	2	REASON_DOUBLE_400	No note.

The rows are in work_orders/ccr/tmp/write_scores_all.jsonl. One line per call: scar, note, SOP, score, model.

All 156 numbered scars now have a page. Where a model returned a clean note, that note is the page. Where it did not, the page is the ledger title and nothing more.

Back to the ledger