the working agreement

the standing policies about who decides what, and how evidence is graded.

AGENTS.md and method/ carry a set of standing positions. they are what make the room predictable to work in.

who decides what

a direct instruction to start or proceed authorizes that work. it is not answered with a proposal panel instead. proposal panels are for significant new directions, when requested or when method/proposals.md calls for one — not for a small example, a source check, a toy computation, or a bounded exploration.

the agent asks when a choice would materially select the mathematical direction or alter the intended claim, and continues independent useful work in the meantime. it shares provisional thinking early enough for you to redirect it, rather than presenting a finished line you did not choose.

and it matches the scale and mode of the request — exploration, computation, drafting, cleaning, review, or decision are different activities, and the reply should be in the one you asked for.

how evidence is graded

five kinds are kept distinct: proposed arguments, finite computations, imported theorems, cleaned exposition, and independently reviewed arguments. the assumptions and the range that matter are stated. a dependent conclusion cannot be more secure than the input it relies on, and is not recorded as though it were.

mistakes are corrected promptly, and the dependencies they affect are inspected.

the formal status ledger is used when you ask for a progress report, when recording a session closeout, or when mathematical status has materially changed. it is not appended to ordinary discussion.

cleaning, and certification

manuscript cleaning is collaborative. cleaners can ask located questions and propose edits; the author confirms that the cleaned text preserves the intent. it is not a silent rewrite.

certification is independent. blind referees see an exact, self-contained review unit, without author confidence signals or prior votes. review proceeds by natural claims and applications, so an accepted conditional calculation is not mistaken for verification of its hypotheses. reviewer dialogue uses focused excerpts and keyed responses — a complete packet is resent only when the review object materially changes.