Guide v2.8: the opening states what the recommended build needs (fixed lines) - #91

Corrects v2.7's item A. The opening is the part of a spec learners are expected to read.

Problem. v2.7's 10.1 clause kept "the build is finished when the offline tests pass, which needs no key". Every v2.7 spec (#88, #89, #90 and all 24 test runs) opened with "They need no key", although each course's recommended build needs an OpenAI key. Owner: the opening must state the truth for the course's recommended build.

Change (10.1 item 1, in place). The opening carries two fixed lines, filled in per course and otherwise copied word for word, like the section 8 gate:

**What the recommended build needs to run:** <each key, account, network access and download its parts use; or "no key, account or download">.

**Fastest path:** answer "recommended build" at the first question. The build is finished when the offline tests pass. The `live` tests run once <the key> is set and are skipped until then; that is not a failure.

Plus: never write that the build, the app or its tests need no key unless the recommended build uses none. The 2026-09-09 completion rule (finished offline; skipped live tests are not failures) stays, as the owner chose.

Checklist

• Reason: owner's reading; build evidence (a keyless run of a keyed build gave canned stand-in answers). P2.
• Word count: guide 4,354 → 4,403 (+49).
• Register row 39, naming row 36 as corrected.
• Runs: a described clause first (v2.8 draft, 24 runs): requirement first 24/24, but "need no key" still in 20/24. Fixed lines (6 runs, 2 per course, owner-approved size): text check 6/6, the fastest-path line byte-identical in all 6, requirement true for the course 6/6, every other check 6/6. The blind scorer's "implies keyless" check passed 3/6 on the identical sentence, so it reacts to the kept completion rule, not to the copying; see the follow-up issue.
• Builds: none needed (wording of the opening only; what the build does is unchanged).
• Generator notes reported to the owner.
• Who can read: generators, build agents, sessions; all course-agnostic, no course names in the diff.

Not tested: a course whose recommended build uses no key (none of the benchmark courses has one).

🤖 Generated with Claude Code