[PierreMardon]

The other pages a person reads showed the pattern #119 treated in ARCHITECTURE.md: paragraphs that gained sentences until nobody reads them through, history told in a page that says how things are, and no bound (#120). The developer asked on 2026-10-12 for the same treatment.

Closes #120

What changes

• evals/README.md, cut by measure. Each measure is an item of its own, where three list items of 954 to 1,706 bytes defined four or five each. What is read the same way for several has a short section: a question handed back, the pull request stop by stop, what a page announces, the faults put there on purpose. The run stopped at the hand over has its section.
• evals/reports/README.md, new: the notes on the campaigns kept. The guide ended on 4.3 KB that told, campaign by campaign, what was judged before which line of the rubric, counted before which change of the method, played in which image. They are needed to read a comparison against one of those campaigns, and they are not how the evaluations work today: they move next to the campaigns, and the guide points to them, with the scale of the counts made by hand.
• docs/guide.md: the paragraph of 1,756 bytes on the blueprint is a list of what the page opens on, holds and closes on, then the assumptions and the critical zones under a title each. The paragraph on what counts as a pass is a list of two.
• tests/e2e/README.md: its seven paragraphs over 600 bytes are cut where their subject changes. One clause of history leaves, on the path "the runs took before the stand-in", and the scenario expected to fail names its issue, A session that dies between the commit of a conformant plan and its push leaves the push to nobody #114.
• Five decision records: six paragraphs cut at a sentence end, with no word changed.
• The rule, in AGENTS.md, under "Documents": what each page is for, that it says what is true today, where the rest goes, and that the prompts are the product and no such page.
• The bounds, for every page a person reads: 600 bytes for a paragraph of the guides, of the README, of AGENTS.md, of CONTRIBUTING.md and of each record; each page at its size plus a tenth; a record at 4,700. The index of the records and the notes on the campaigns grow with what they list: their paragraphs are bounded, not their size.

The critical zone

ARCHITECTURE.md changes in two sentences, for the developer to read: under "Tests", the form check now bounds "each paragraph and each page a person reads, this one, the guides and the decision records"; in the codemap, reports/ holds "the campaigns kept and what each compares on".

Measured

Page	Bytes before	Bytes after	Over 600 bytes before	The longest after
evals/README.md	27,145	24,061	12	528
evals/reports/README.md		4,949		445
tests/e2e/README.md	25,750	25,787	7	565
docs/guide.md	12,861	12,950	2	562
The records of docs/adr/		no word changed	6	597

These pages are no shorter: the guide of the evaluations and its notes hold together 1.9 KB more than the guide did, for the titles and the lists. What changes is that a reader finds a measure, a campaign or a rule without reading a wall through.

Checked

• scripts/gate.sh, green on 3.11 and on 3.14, 1,897 tests, before the corrections below; the prompt tests after them.
• An independent reading, by a fresh agent, of the guide of the evaluations against its two new pages: 166 sentences compared. Every measure named exists in evals/surface_evals/, and what the notes say of each campaign was checked against the folders and the summaries of evals/reports/.
• It found one caveat lost, on the figures counted by hand, two reasons dropped, and five places where the new text said more than the old: a pointer that covered five measures where its section speaks of two, the pages of the old campaigns said to predate a form where the old page said most of them fall outside it. All are corrected before the first commit.
• The guide, the end to end guide and the records were read word by word against their diff: reorderings and cuts, and the changes named above.

Left alone, on purpose

• The prompts and their templates, agents/ and skills/, 11 files and 25 paragraphs over 600 bytes. They are the product: a change to them changes what the chain does, which only a campaign of the evaluations measures.
• What a run or a campaign wrote: docs/plans/, evals/reports/*/, docs/e2e/. They are records.
• The corpus of the evaluations: evals/cases/, evals/host/, evals/judge/. A change there changes the measures.

To weigh

• evals/reports/README.md has no bound on its size, by design, and is where history now goes: it can grow as the guide did. Its paragraphs are bounded.
• Two statements the old guide already made sit closer together in the notes: the campaign of 2026-10-07 is "the first counted by the method as it stands", and was "counted before the sentence on how far the plan goes into a zone was set aside". The notes also name two campaigns by rank, "the first" and "the second", where three folders share a date. Both are kept as written: settling them takes someone who was there.