Unit 3 course-wide blind review — results and adjudication (2026-08-06)

Process: 446 scoreable items (390 lesson: 333 MCQ + 57 numeric; 56 gate MCQ) in 5 bundles (A–D lessons, G gates); 5 independent reviewers, no keys, no spec access; answers cross-checked programmatically against sequestered keys (cross_check_blind.py, report in crosscheck_report.json). Reviewer reports preserved verbatim in report_A..G.md.

Headline

446/446 reviewer answers matched the keys — zero keying errors across the entire unit, lessons and gates. No unanswerable stems, no undecidable items, no factual chemistry errors found by any reviewer (several real literature values re-verified independently by reviewers).

Defects accepted and fixed

1. Maxwell-Boltzmann peak-shift flaw items were label swaps, not single-feature flaws (B089, B091, G023 — found independently by two reviewers). Root cause: the mbcurve renderer applied its equal-area height scaling to the DRAWN (swapped) peak positions, so the heights read wrong along with the positions while stems claimed exactly one error. Fixed at the renderer (heights now always follow the labels — one wrong feature by construction) plus surgical item repairs (stems → "Which statement identifies an error in the sketch?"; area options replaced with clean non-errors; alts updated). Operator-repaired, coupled to the renderer change.
2. A055 Cl⁻/water hydrogen-bond defensibility (IUPAC/lit. allows halide H-bond acceptors; the figure draws canonical H-bond geometry) — option replaced with a different named error path. Same class as Unit 2's CHCl₃-azeotrope catch in gate authoring.
3. C022 "number of gas particles" defensible for n — distractor rewritten cleanly wrong.
4. D095 boundary item's extrapolation was only ~12% out of range (defensible short extrapolation) — unknown moved decisively beyond the standards; key unchanged.
5. Same-form cueing: D030's keyed option printed the full spectral ordering (contains D029's answer) — option reworded to the relevant comparison only.
6. Absurd/throwaway distractors (A001, A010 — contradicted by its own stem, A013, A024, D001, D002, D016, G011, G012, G013, G039, G051) — replaced with named-error-path options.
7. Length-cue repairs: gate explanation-item pattern (G011/G032/G040/G043/G045/G050/G051), C011/C020/C042, A081 — keys trimmed / one distractor lengthened, none weakened.
8. Small wording repairs: C002 stem giveaway ("doubles to") removed; C012 co-claiming options split; C019 self-eliminating axis options reduced; D028 self-undermining option replaced; D090 "cannot be right" overstatement reworded; B063 stem tightened to "arrangement"; B028/B029 and A025/A059 and D034/D071 near-identical pairs differentiated.

Flags adjudicated as acceptable (with reasons)

• Template drill families across lessons (noble-gas ranking runs A011/A019/A025/A045/A059; partial-pressure runs C043–C045; Kelvin-scaling C002/C003; alloy what/why runs B026–B033; graphite pair B015/B016; microwave/IR pairs D034–D043): items live in different lessons taken in order, not one exam form — the Unit 2 "template drill family / teach-then-use continuity" precedent. The worst same-form pairs were differentiated (fixes above); the rest stand.
• Figure-optional stems on named molecules (A050–A057, A088, D007/D008, G021, G052): the figures standardize the intended structure; removing them punishes students who need them — Unit 2's "answerable without the figure on famous molecules" precedent. Kept.
• B086/B087 "answer printed on the guide label" — that IS the design: the pilot (twice blind-reviewed, James-approved) deliberately made every numeric read come from a labeled dashed guide so no cross-axis numeric error exists (pilot notes, CSP-049 pattern). Kept; seeds are locked. D094 follows the same approved pattern. Kept.
• D022/D045/D056 same two-wave figure reused — different concepts (λ, ν, E) in different lessons; the recap-shows-the-representation rule requires re-showing. Kept.
• Gate near-duplicates across different atoms (G001/G002 = PSM-005 vs PSM-008; G006/G044, G014/G045, G032/G048, G035/G049 = G5-vs-G6 PSM-100 parallel forms): the register's gate table itself samples sibling atoms and repeats PSM-100 in two gates — register-driven, kept, noted for James.
• D029-style cumulative-cueing within the S17 block (many items collectively teach the ordering): course-context teach-then-use; only the direct option-leak (D030) was repaired.
• Verdict+reason gate items ×3 (G2/G4/G6) — sanctioned sparing AP-mimicry use per the 2026-08-06 ruling; all flagged by their authors; the duplicate PSM-053 format question is parked for James in review_notes.md.

Re-review round (bundle R — the 38 repaired items, fresh blind reviewer)

38/38 answers matched the keys. The reviewer confirmed the Maxwell-Boltzmann repairs hold (B089/B091: "the earlier-round confusions appear fixed; no cueing detected"). Four residual findings, all fixed same-day (operator, surgical): 1. D095 — the figure's unknown read-arrow actively invited the extrapolated read the key forbids. Figure reworked: arrow removed, dashed guides mark the TOP STANDARD instead, the unknown's A = 1.80 moved to the stem, alt updated ("the line is not extended beyond the highest point"). Extrapolation distractor arithmetic preserved. 2. G023 — the first repair round introduced logically-equivalent mirrored options (Ne lower / Kr taller); one replaced with an independent clean non-error (gradual high-speed fall). 3. C019 — stem's "fixed amount … which other variables" contradicted the key re-listing amount; stem now asks plainly which variables must be held constant. 4. A017 — the first repair's replacement option contradicted the stem's own "identical electron counts" (dead option); replaced with the bonds-vs-IMFs misconception. Post-fix: validator 0/0, rebuild clean, geometry sweep 0 problems, full click-through 0 failures. The reviewer's borderline note on A055 ("expect some name-the-dominant-force pushback") is recorded as acceptable — the CED's LDFs-between-all-particles convention governs.

Bundles (current + pre_repair/), sequestered keys, reviewer reports, the cross-check report, and repaired_item_ids.json are preserved in this directory. All repairs re-validated (0 errors), re-swept, and re-click-tested.

Exam-artifacts round (bundle X — module + Unit 3 test + practice, 2026-09-07)

30/30 MCQ answers matched the keys (module 11, test 18, practice 1); all six FRQs attempted blind — reviewer's independent answers would earn every rubric point; 5 of 6 rubrics judged fair.

Arbitrated AGAINST the reviewer (primary sources checked): the F01 finding claimed the module's "correct mass with no visible setup earns nothing" clause was an unfaithful hardening of the 2023 Q4(a) rubric. The 2023 scoring COMMENTARY (Sample 4C) explicitly records a real student who "provided the correct mass" earning zero because "insufficient supportive work is shown to justify the calculation" — the clause is faithful to the applied scoring. Rejected; the related X004 stem was still tightened to "certain to earn the point" (removes any scoring-variance defensibility argument at zero cost).

Accepted and fixed same-day: X002 option-format asymmetry (rationale added to the key); X010 keyed feedback now acknowledges the exam's ±1 sig-fig tolerance; X030's cross-reference made self-contained (the "leak" was a bundle-reordering artifact — in the course the MCQ follows the FRQ by design); F02's long FRQ now supplies the BaSO₄ molar mass in-stem (the real exam's periodic table isn't in our surface); set-level longest-key pattern broken by elaborating one distractor each on the three pure-content items (4, 7, 16), none weakened.

Noted, no action: F03's piston render omits the 2021 Q7 valve (the stem supplies it); candidate-response items keep complete answers longest inherently — James has the exemption question in review_notes.

Post-fix: both courses rebuilt; validator 0/0; click-throughs clean (module 12 checks; Unit 3 479 checks incl. test + practice).

Authored long-FRQ round (bundle Y, 2026-09-07)

The topic-purity replacement (authored distillation/vapor-density long FRQ) was blind-reviewed by a former-AP-reader-lens agent: worked cold, all 10 points earned by the reviewer's own answers, zero BROKEN findings, verdict "would put this on a topic test as-is." Accepted refinements applied same-day: (c) rubric accommodation for answers engaging the dispersion data (adjudication kept the electron counts in-stem — the course's taught ranking hierarchy makes "more electrons wins" cleanly wrong, and defeating lazy polarizability reasoning is the design); (f) bundled point condition made concrete with consistency protection; (b) full-structure requirement stated and lone-pair dashes accepted. Adjudicated keep-as-is: all-or-nothing (a) (standard on real rubrics), (g) packing (real-practice weight), the Dumas idealization (2019 real-exam precedent), and guessable verdict points (justification scored separately, per AP practice). Re-validated 0/0, rebuilt clean.