[LTSCommerce]

What this changes

The acceptance process, so that the cohort test proves what it exists to prove, the clarity of the document to a cold reader, and nothing else. ACCEPTANCE.md is rewritten around five fixes, with CONTRIBUTING.md, acceptance/README.md, acceptance/runs/README.md, the three question sets and the three keys brought into line, and two new files, acceptance/CHANGES.md (the record of process changes) and acceptance/ROUTED.md (findings routed out of a run to a change of their own).

1. Keys ask what the question asks. Each key lists, per load-bearing item, the numbered elements an answer must state, one for each part of the question, and the assertions that make an answer wrong; everything else in the key is explanation and never marks an answer down. The questions are reworded to ask for exactly those elements (question 2's instruction asks for "the sentence or sentences"; detector 2b asks whether one of the two things is enough; detector 2c asks at what version; the question 3 verdict items ask "and why"). Detector 2b no longer requires "need not remove or forbid", which the question never asked and 28 of 30 readers were marked down for.
2. A confusion finding is settled run by run. A passage named by two or more readers of one run is a finding; the next run is judged on its own readers, and a passage fewer than two of them name is clear. Nothing is carried from run to run except the record.
3. A change is judged on what it changed. A finding in a paragraph the pull request added, reworded, moved, deleted from or changed a term or link in (read against its merge base), or that a reader sets against such a paragraph, blocks the change. A finding anywhere else is routed to acceptance/ROUTED.md and a change of its own, and does not block. The keyed questions still cover the whole document on every run, so the document is still proved; a keyed miss blocks wherever its text lies.
4. Marking judgements are written into the rules. The Marking section defines a keyed item, quoting (a six-word verbatim run), reading for sense, question 1 counting, a correct question 2 answer, a question 3 grade with its condition, a wrong reading the text invites (stated in the key as a grade plus what the reasoning says or omits, blocking on one reader), an unanswered item, naming a passage, the same passage, where a finding lies, and the direction every open point takes. A run record states no judgement the rules do not; an open point is settled in the rules or the key before that pull request's next run.
5. The "traceable miss" row goes. Any miss whose reasoning quoted the document was a finding whatever the count, which made the all-but-one tolerance meaningless. Now a single miss under questions 1 and 3 counts against the document only where the key names that reading; everything else is the reader's, and a second miss on the same item fails the threshold without a judgement.

Kept from pull request #13, which this supersedes: the odd cohort of at least five, the isolated reader process and its example command, the packet form, set-aside readers and void runs, the sheets posted on the pull request, the fuller run record, the route for a process change and acceptance/CHANGES.md.

The observed case

Thirty-three runs on two changes (19 to 30 on pull request #14; 31 to 33 and 35 to 52 on the defined-terms change) in which every keyed answer reached five of five and the runs still failed: on key elements the questions never asked for, on passages the change never touched, on flags that one later reader kept open, and on judgements J1 to J12 that the markers invented and later runs cited as precedent. Runs 050 to 052 got worse after a sixth rewording, not better. The run records on those branches are the evidence.

Kind of change

• Clarity edit
• Obligation

Neither: a change to the acceptance process, recorded in acceptance/CHANGES.md and not tested by a cohort, as ACCEPTANCE.md says under "Changes to this process". tools/changes.py --base origin/next and spec-qa.py are clean; the unit tests pass.

Where it is looser than the rules on next, and why

Two rows are narrower than next's, and the pull request says so rather than hiding it. The traceable-miss row (point 5) measured whether a reader had quoted the document, not whether the text was unclear. The confusion row (point 3) measured the whole document on every change, so that a one-line rename was reviewed sixty paragraphs at a time by five readers and could not converge. Neither change alters a mark already given; every run under the new rules is a fresh run. ACCEPTANCE.md under "Changing the test" now states the test for touching a rule: a rule that measured something other than whether a cold reader could follow the text is corrected, with what it measured recorded; a rule that measured that is left alone and the text is fixed.

Review

Three independent reviews by a reviewer that did not draft the change, each applied in its own commit (review 1: 19 findings; review 2: 20; review 3: 9, none blocking). acceptance/CHANGES.md carries the line for this change; its pull request number is filled in by the next commit.

Not done here, by design

No cohort run: a process change is not tested by one. Merging is the editor's. The two held changes and the two changes routed out of their runs were rerun under these rules at a commit named in each record, as "Changes to this process" allows, and each passed: pull request #14 (runs 70 and 79), #24 (runs 64, 80 and 87), #23 (run 78) and #25 (run 86).

Settled during those runs

Open points the run records raised were settled here before the next run of the pull request that raised them, as "Changing the test" requires, each in its own commit: a remark that another document was not supplied, or that a term is defined only there, is a note however much of the document it quotes to locate the reference; a paragraph the pull request parts from what it refers to, by inserting a paragraph between them or moving either, is the change's text; the toolchain key's 4.3 row and the detector key's 4.2 row say which case the fixture presents, so that a grade given on a stated condition is the grade for that case; the Lintomatic fixture says check --only prints its findings as check does; and acceptance/ROUTED.md carries every finding the runs made in text no change had touched.

🤖 Generated with Claude Code