Where every document lives

One root, one naming scheme. The repo root holds exactly three documents: README.md (status and quick start), AGENTS.md (the agent rulebook — rules, machine facts, and every retraction), and CONTEXT.md (the system map). This directory holds everything else, and docs/glossary.md keeps its path because every document in the repo links it.

directory	what lives there	how to name it
adr/	one record per decision, with a Status line. Retracted decisions stay	NNNN-slug.md
protocols/	the instruments: what a number must satisfy before it may be written down	as the instrument is known
research/	dated findings — what was measured, what was concluded	YYYY-MM-DD-slug.md
reviews/	hostile reads of a lane, written to attack it	YYYY-MM-DD-slug.md
decisions/	the owner’s decision sheets: inputs to a decision, not records of one	YYYY-MM-DD-slug.md
architecture/	the design documents: program plan, model floor, mixture arms, post-training, bf16	as the design is known
guides/	how to run a thing on this box	as the task is known
papers/	transcriptions of other people’s formulas, the PDFs themselves, and provenance.tsv (file → source → sha256)	papers are <arXivID>-<slug>.pdf
glossary.md	the vocabulary, and the live list of where a document and the code disagree	—
archive/	history, not truth — never cite it in an argument	unchanged

Two rules worth stating because they are the ones that get broken:

• A note is a snapshot of a belief on a date. Nothing in research/ is maintained. Where a note and AGENTS.md disagree, the rulebook is later.
• The site is generated from these paths. docs-site/tools/manifest.mjs is the information architecture as data; adding a document means adding a line there, and npm run check fails on a link that resolves to nothing.

If you cannot tell where a new document belongs: it is a measurement → research/; an attack on someone else’s work → reviews/; something a reader needs before trusting a number → protocols/; something superseded → leave it in git, do not add it.