Corpus Log
Append-only, reverse-chronological. Newest first.
[2026-09-10] ingest | April 23–24, 2026 TAC meeting
Action: ingest
Initial construction of the corpus from the April 2026 meeting materials, built for demonstration at the September 10, 2026 TAC meeting.
Pages created: 76.
1 meeting, 5 sessions, 7 artifacts, 12 topics, 38 recommendations (including the register), 8 questions, 12 people, 11 organizations, plus schema.md, glossary.md, and index.md.
Sources ingested
Everything in raw/2026-04-23/ — the agenda, the 9-page summary minutes, and five presentation artifacts. Text was extracted with pdftotext -layout, PowerPoint content including hidden slides and speaker notes via direct OOXML extraction, and the 52-page CT-PO simulation deck was rendered to page images and read visually.
Key additions the source minutes do not contain
The simulation results. Roughly thirty of the deck’s fifty-two pages are chart images carrying no text. Reading them recovered: the empirical non-normality that justifies abandoning the multivariate normal assumption; the eigenvalue evidence that the data is effectively one dimension (first eigenvalue ~12.7 against second and third ~1–2.5); the finding that CT-PO beats SBAC on accuracy and consistency in every simulated condition; and — most importantly — the Batch 2 gradient showing that CT-PO’s advantage collapses as the latent dimensions become less separable, which is the realistic operating region for a test engineered toward unidimensionality. Recorded in CT-PO vs SBAC Simulation Results Deck.
The hidden slides and speaker notes. Eight hidden slides in the growth deck carry the canonical statement of Michigan’s current growth model. Two speaker notes carry material found nowhere in the minutes: the repeal of educator evaluations and school grades from state law, and the post-COVID data limit that explains the WIDA projection horizon. Recorded in TAC Growth Update Deck, April 2026 and Michigan Growth Model.
The participation trend. The handout’s own totals show interim testing down about 2% in a year when six new schools joined — a retention problem the presentation does not name. Recorded in Smarter Balanced Interim Participation Trend Data.
The design decision that makes it work
recommendations/ as a first-class page type with status: and raised:. A TAC’s value is longitudinal and the current record cannot support the question “what did we advise, and what happened?” Open Recommendations from April 2026 is the register that answers it: 37 discrete recommendations, 2 accepted, 1 declined, 34 open.
Fidelity decisions
Body-level attribution preserved throughout — the minutes attribute to “TAC,” never to individuals, and inferring otherwise would fabricate a record the source cannot support. The loss is documented rather than repaired.
Transcription errors corrected in wiki prose and catalogued in TAC Meeting Notes, April 2026, including a correlation reported as “between 2.7 and 2.8” (certainly 0.7–0.8) and “Cronbach’s coefficient theta” where the source deck says coefficient α. The error rate is evidence, so it is recorded rather than quietly fixed.
Where the agenda and minutes disagree — session order, session naming, an off-agenda WIDA item, an unrecorded minute-approval item — both are recorded and the minutes are followed for content.
Open — the meeting this was built for
Nothing in the register has been closed, because no meeting has occurred between April and September to close it. The September 10, 2026 meeting is the first test of whether the register does the job it was built for.