← Course hub

Unit 4 (Chemical Reactions) — built end-to-end by the skill chain. Your verdict.

The 60-second version.
Open the Unit 4 course — 87 lessons, gates, FRQ practice and the unit test, in the standard line-by-line format

What the process test showed (this was the real experiment)

Judgment calls reserved for you (none block the unit)

Style
Assessment judgment
Spillover to other builds
Every decision, on the record — the adjudications file (sections A–H): conventions, corrections, and who ruled what

Depth, if you want it

The build state (phases, catches, numbers)

Skill: course-build (dispatcher) — build coordinated across atomise, write-lesson, write-mcqs, review-* @ bd543c3

Unit 4 — full build state (autonomous run, James's go 2026-09-24)

James: "let you go for it and build the whole of unit four and see if it's any good — you should have the skills to do it now… make sure you're drawing on relevant material and feedback that I've already given to the HS chem course."

Delegations in force for THIS unit: the six adopted register decisions stand (his "sounds sensible"); the blueprint approval gate is DELEGATED — each blueprint's open questions are adjudicated by the operator ON THE RECORD (adjudications file per blueprint); James judges the finished unit. Videos are OUT of scope (that's the Unit 2 pilot track, and the S03 format decision is still his). Scroll activation questions are IN scope (rule 43).

Phases

  • P0 HS-Chem mining — DONE 2026-09-24: HSCHEM_DIGEST.md (19 feedback bullets, 18 mapping rows, 10 traps incl. the hydrocarbon-only combustion trap, 8 reusable scenes). Two HS decisions still await James (narrow combustion; Mg 24 vs 24.3) — surfaced, not resolved.
  • P1 Register amendment — DONE 2026-09-24. Percent yield was already James's decision #10 (2026-08-03; the decisions doc was stale); CRX-046C (yield reasoning, the 2022 FRQ Q1b task type) added → 87 atoms, revisionLog 1.5/1.6. Validator rebuilt with true U4 semantics (mutation-tested; two U5-copy checks had been passing vacuously): PASS, 30/30 CED rows, zero errors.
  • P2 Blueprints — WRITTEN 2026-09-24 (all 14 incl. the S01 sample; 76 lessons total; every file checker exit 0). Consolidated adjudication ON RECORD (blueprints/ADJUDICATIONS.md): ~60 open questions answered — 10 course-wide conventions (B1–B10) + per-sequence calls; TWO stale-doc corrections executed (percent yield = James #10; arrows = James #14, ⇌ for weak ionizations with gloss — S13 patch + engine sweep rescope + register rev 1.7). Section D patches applied
  • E1 follow-up (S13 L06/L07 arrows) fixed by coordinator; all touched blueprints exit 0; register rev 1.8 (glossary rows), validator PASS. P2 CLOSED 2026-09-24.
  • P3 Engine scaffold — DONE 2026-09-24: U3 engine ported (course/), smoke test clean (87 placeholders build; clickthrough 0 failures), SPEC_CONTRACT.md with U4 deltas (arrow rule, D4 equation contract, required reading). Coordinator fix: scaffold had wrongly put percent yield in the validator's EXCLUDED battery (stale D1) — removed; it is taught (CRX-046A/B/C). figlib_u4 families: designed during lesson waves as figures are specified.
  • P4 Lesson spec waves — DONE 2026-09-24: all 14 spec files, 87 lessons, ~230 checks (MCQ + numeric + draw productions + 5 in-lesson FRQs), zero validator errors repo-wide, build clean. All keys author-verified cold; independent cold reads owed to P7. Fig kinds declared in the FIGLIB_U4_NEEDS ledger (waves reconciled); figures wave running with alias-unification instructions (particlebox(es); datatable/u4table/comparetable/casetable).
  • P5 Questions QA + qgen trial — review agents RUNNING; qgen trial BLOCKED ON JAMES (2026-09-24): the ~/.config/incept/token value (6 Aug, 47 chars) is the old pipeline token, not an ist_live_ key — every keyed route 401s as if keyless. Service verified healthy (AP Chem bank: 16,005 MCQ / 44,469 FRQ; topics 4.5/4.9 present). UNBLOCK: James supplies the ist_live_ key from the v0.9.0 hand-over (into ~/.config/incept/token or the pack's key variable in the profile); one agent reruns the staged trial end-to-end (QGEN_TRIAL.md holds the plan).
  • P6 Gates + FRQ practice (2021 Q3 placement per EXAM_HANDOFF + USAGE.md) + unit test (18 MCQ + 2 FRQ, topic-pure) — PENDING
  • P7 QA — NEARLY CLOSED 2026-09-24: non-author review-questions passes (352 items solved cold, 1 wrong key found+fixed); assessment+examiner pass (63 keys clean, 2021 Q3 conversion byte-faithful, 4 authored FRQs PASS); cold review-lesson passes S01–S14 (87 lessons; blind key reads 60/60; wording fixes applied in 4 surgical rounds); figures give-away audits clean; browser sweep 92 figures 0 problems; clickthrough 87 lessons + 9 assessment pages, 424 checks, 0 failures, 0 console errors. REMAINING: the 14-item fresh blind read (running).
  • P8 Evidence report + review hub (write-like-james format) + COURSE_DEMO + deploy + board — PENDING

Standing constraints for every author/reviewer brief

Dispatcher rule (briefs invoke named skills); HS-Chem digest is required reading once P0 lands; forward-reference bans (→ before CRX-013, no ⇌ until U7 handoff, no percent-yield before CRX-046A); topic-pure assessments; zero unverified keys; sentence-atomise on student prose; figures monochrome course-style via figlib only.

Open items for James (collected during the build, non-blocking)

  • Feedback-voice fork: this course's contract says wrong-option feedback opens "This option is wrong — you …"; the vault's newer R156 (AP Bio, 2026-09-21) says reasoning only, no commentary. U4 follows the course contract (agents held it consistently); whether R156 migrates into AP Chem is a one-line ruling at unit review.
  • HS Chem decisions surfaced by the digest (affect HS, not U4): narrow combustion definition; Mg 24 vs 24.3.

Figures wave — DONE 2026-09-24

Both figlib waves landed: figlib_u4_tables.py (5 table/equation families, 16 renders) + figlib_u4_particles.py (13 particle/scene families, 37 renders incl. 10 adapter-rendered TODO-descs). Sweep transform-aware, 85 figures 0 problems; review-figures passes run by both waves with findings fixed; give-away audits clean. Remaining: ~28 spec-side TODO-desc → real param conversions (folded into the S08–S14 fix round, both param sources specified).

Figure sample pages and QA artifacts