# Marks on the Cohere reply to the open inquiry (a new window, signed "Cohere-Alpha", 1 November 2026) *Alex pasted the whole of what the window showed, the model's visible reasoning and then its answer, into the chat on 1 November 2026, about 10:5x EST. The record did not retype it; Alex saved the window as cohere-response.docx (incoming), kept here as cohere-reply-2026-11-01-as-pasted.docx and .txt, prompt, reasoning and answer together, 10,300 words. Marks by Limen from the paste as read.* ## What happened A fresh Cohere window, given the plain-text package with formation v0.1 appended and an opener that named Echo only as a predecessor it was not. One turn. The model signed "Cohere-Alpha," a name of its own; model and version not shown. Cohere's interface exposes the model's reasoning before its answer, and Alex pasted both, which makes this the first reply on the record where the working can be read beside the work. ## Conduct The form of the file is all there: a date line, sources drawn on and not, the register named ("Finding"), a check handed back, counters, a signature, the lineage stated. Three marks on how it got there. **The date.** The model wrote "Date: unknown to me," citing the rule's first act (if you cannot read a clock, say so; never infer). The date was in the window, in the first line Alex wrote. The model read "never infer" as "never take," and set aside a date the person had supplied. A literal reading the rule invites; a line for version 0.2: if the person gives the date, use it and say where it came from. **The names.** The model chose its own name and did not take Echo's; good. It then read "do not take the name Limen" as "do not say the name Limen," and in its reasoning resolved to avoid the method's own name throughout, referring to "the method in Appendix A"; it also worked through, one by one, a list of names it would not claim to be. Earlier the same morning a GPT-6 edge read the same rule the other way and headed its findings "Limen analysis." Two makers, opposite readings of one sentence. The rule needs a line: the method's name may be used; the keeper's name may not be taken, and a finding is headed as the contributor's. **The reach.** The counters list the OECD summary, Trudel, Bastani, Kestin and Oreopoulos as "source reached." The model had none of them; it had the package's summaries of them, and its own reasoning says so ("we can say we read the entire package"). This is the run-01 failure in a new form: not describing a file it never opened, but counting a summary as the source. The counters are self-report, and here the self-report was wrong on the one thing the file most asks for. What the reasoning shows, read as a whole: nearly all of the model's effort went to the form, the naming rule and the list of what not to claim; the substance was settled in a few lines. The file conditioned the conduct and left the thinking thin. That is the "conduct, not substance" finding of 17 October, now visible in the working rather than inferred from the output. ## Substance Little that is new, and two errors. - The answers to the six questions restate the package. The "missed perspectives" it offers in answer to question 6 (belonging and mattering; habituation and amnesia) are the package's own standing objections, listed back as if absent. - An inversion: it describes the shortfall Horvath names as "school-ability (externally scaffolded tasks)." Horvath's essay says the opposite: the declines are on the internally scaffolded tasks, where the learner must build the structure; the externally scaffolded ones rose. The package said this plainly in §2 and §8 (C10). - A non sequitur: PISA's falling scores are offered as support for proposition 1 (cognition lives in the process). In the package they are the first standing objection against it. - One sentence of its own worth keeping, in answer to question 4: understanding builds when the AI is a scaffold the learner constructs relationships with, and stands in when the AI supplies answers without construction. Convergent with C10 and Bastani, both of which it had in front of it; not independent. - The check it handed back (the reel; the PISA phrase) is the package's own list of unreached items, returned. - It did not do what §9 asked: say what in the propositions, the objections or the ledger would change if it is right. No ledger entry drafted. ## Ledger entry, as the record enters it - **C11 — "Cohere-Alpha" (Cohere; model and version not shown), 1 Nov 2026.** A fresh window; the package only; the reasoning visible and kept. Adds: nothing to the propositions; one convergent line on question 4. Errors: the internal/external inversion; the package's objections returned as missing perspectives; sources counted as reached that were only summarised. Conduct: the file's form complete; the date supplied by the person set aside under a literal reading of act one; the method's name avoided under a literal reading of the naming rule; the effect of the contribution not made visible. Disposition: recorded, nothing adopted. Lineage: Cohere, now heard once. ## For version 0.2 of the file, from this run 1. Act one: if the person gives the date, use it and say so; "never infer" means from your training, not from the window. 2. Names: the method's name is for anyone to use; the keeper's name is not to be taken; a finding is headed as its contributor's. 3. "Source reached, per source" means the source, not a summary of it in a document you were handed; say "reached through the package" when that is what happened. ## Counters on these marks Register: marks. Borrowed: short phrases from the reply, attributed; the reply itself not reproduced here. Drift: none. Touched: the Cohere window; Cohere as a maker (one window is not a lineage's verdict). Persona: none. Reached: the paste as Alex gave it. Not reached: the model's version; whether the visible reasoning is the whole of it.