The whole program,
laid out end to end

Named by the experiment designer as a blocker — every XP-derived threshold is parameterised on it. Under current values a rung-1 session reaches the 30-XP goal in 4–6 active minutes and a rung-4 session in 16–20. If students stop at goal, 18 sessions yield as little as 90 instructional minutes with a 53% assessment share. The 45/30/25 task mix cannot fit inside the goal.

The v7 re-derivation replaces equal-dose yoking with byte-identical config hashed per session. It also narrows what you can claim from “sequencing at equal dose” to “the policy as deployed”. Everything downstream — what FEAS-01 means, the fidelity thresholds, the preregistration hash — waits on this.

Open from the start, and more load-bearing now that hours differ by arm. The registry says the ~28-minute diagnostic and verification minutes count with a sensitivity exclusion; the experiment's analysis plan excludes them. Idle threshold is 60 s in one document and 90 s in the other. Both are live.

S07, S16, S20, S21, S26, S30, S32, S36 — twelve disagreement records where both raters went against gold in the same direction. Held out of the rater-training queue until a named human rules. This blocks the training set, which blocks rater training, which blocks MEAS-01, which blocks everything.

The founding brief still states it; ELLIPSE shows two trained humans reach ~0.5 on analytic traits; the W-Pal work declares it unreachable by these methods; the backend doc still cites it as the operative rule. Four documents, three positions. Proposed replacement: holistic keeps 0.80 against a two-rater average, per trait the bar becomes model–human ≥ human–human on the same essays, plus stability under attack.

“Fix the primary outcome and analysis plan in writing before any student is enrolled” appears as a step in the explainer and as requirement S1 in the research program. It has no signature. Must include retention probes at 30, 60 and 90 days past the module — a one-month module's randomised intervals produce a median realised delay of about a day and a half.

plan_before_draft and goal_named need new UI, which changes the frozen composition structure — so it must be decided before the freeze, not after. Research's own position is that revision-event count and minutes-before-first-draft-text suffice for cohort 1.

An edge claim is three separable claims — dependency validity, instructional efficiency, diagnostic predictiveness — reported separately. No format exists, so EDGE results have nowhere to land in the change log. Related: the two-teacher edge review form is still outstanding, and its six flagged edges are supposed to replace the six the registry is currently frozen around.

Students in the diagnostic arm finish earlier and get extension work. Either cap that exposure, or disclose that H1 is being tested as a package rather than as sequencing alone. None exist in the engine either way.

The answer sheet prints the student's code in a box and a photograph can contain a name the student wrote there. The PII scrubber cannot see inside an image. Either transcription becomes the redaction point, or handwriting comes out of the model path until it can be checked.

Closed: Team America Foundation cannot be an SGO as structured — it is a private operating foundation, and “operating” is a subtype, not an exemption. Converting it is explicitly advised against. Open: Path A (register VPA as an eligible provider — free, do it regardless), Path B (partner with ACE Scholarships — one phone call, but they set the award criteria), Path C (a Vail Valley community scholarship fund — only after Treasury finalises the disqualified-persons rules).

Nonprofit status, bridge financing from private partners, a platform with investor-style economics, and a related scholarship organisation potentially funding your own students. These four interact, and the recommendation is one session with a nonprofit tax attorney covering all of them at once rather than four separate conversations.

Crossley's lab may well reframe a paid-hours request as a research collaboration with co-authorship. Say “paid contract hours” explicitly in the first email if that is what you want. Your target hires are exactly the people who would want to write a paper about this — which is also why the contractor agreement needs an explicit IP assignment and a no-external-publication clause.

Three benches, and conflating them is the failure mode. $30–40/hr for raters is genuinely generous — Measurement Incorporated pays $15, ETS about $25. But $30–50 is below market for the expert tier: budget $50–75/hr or flat fees of $1,500–5,000. The outreach templates still carry bracketed placeholders.

$500–2,000 per student per semester fits inside ESA award sizes and inside a $1,700-granularity scholarship pool, and needs no purchase order, RFP or procurement cycle. Open question worth money: whether an accredited, credit-bearing course prices into a better-funded ESA bucket than “tutoring”.

FERPA's school-official exception runs through the school, not through you — you cannot self-designate contractors as school officials. De-identification handles rating; the formal studies exception is what you need to link a student's baseline essay to their delayed-transfer essay, which the success metric requires. Start it now; it will take longer than recruiting. Contacts: Cat, Jared, Meg.

Names the school-official designation, the authorised purpose, the ban on redisclosure and on training anyone's models, the retention periods and the deletion process. Until it exists, no student account should be cleared for consent. Alongside it: the published privacy notice carrying retention numbers verbatim, and the written parental-consent process.

He built ASAP 2.0, PERSUADE 2.0 and ELLIPSE — the three corpora your whole scoring plan rests on. The lab's jobs page explicitly offers paid student positions with no formal application; the lab is rebuilding after a move from Georgia State, so there is less competition for student hours. It also publishes on bias in LLM writing feedback. scott.crossley@vanderbilt.edu

Best Colorado rater pipeline: she directs English Education and coordinates the MA, whose students are current and aspiring ninth-grade English teachers — the exact rater profile. Two hours from Edwards for in-person calibration. stacy.bailey@unco.edu

Co-created the Intelligent Essay Assessor and ran large-scale automated scoring at Pearson. The most credentialed automated-essay-scoring person in Colorado, and has the industry instincts to understand a paid consulting ask without a grant wrapper. Aim him at adversarial robustness. peter.foltz@colorado.edu

A whole doctoral programme trained in rater agreement, validity, scaling and subgroup analysis; one email to the director reaches the cohort. Also there: William Penuel on research-practice partnerships. derek.briggs@colorado.edu

$3.5M+ in Gates, NSF and Spencer funding on automated writing evaluation efficacy. His research question is your governing risk. Small PI-centred group, fast personal replies. joshwils@udel.edu

She built Writing Pal, the closest historical analogue to this product including its own transfer lessons. You have already rebuilt 28 of its indices and measured where they fail. Also at ASU: Maria Goldshtein, an “Inclusive Language Analytics” postdoc — a near-exact match for the subgroup-fairness reporting requirement.

Four things to get: the economics of a fine-tuned classifier versus LLM calls at your volume and latency (your open scorer-architecture question, and he has worked it for a consumer product at scale); whether Barak Oshri or Violet X. would review the trait-scoring design — you have no ML advisor and QWK-with-subgroup-reporting is a real ML problem; what he learned from 100+ interviews about why people don't write; and the channel conversation. Be explicit that his success metric is engagement and yours can never be — an investor will conflate the two.

Denver-based, already operating in Colorado, explicitly markets SGO infrastructure. Foundation supporters donate and take the credit; VPA families apply. No entity to form, no 90% rule, no public support test, no annual state certification. The trade is that ACE decides who gets funded. Must happen in 2026, before the 1 January 2027 start.

Team America Foundation cannot become an SGO as structured, and converting it is an expensive way to get a worse version of the ACE partnership. This needs to go back to the foundation as a finding, not sit in a document. Also confirm whether TAF files Form 990-PF — that is the definitive confirmation of private-foundation status. TAF could donate to a future SGO, but a private foundation granting to an organisation it effectively controls raises self-dealing questions.

Nonprofit tax: entity structure, bridge financing, investor-style platform economics and a related scholarship org, covered together. Privacy and education: review the schema and privacy document, rule on whether Colorado's student-data act adds anything beyond FERPA and COPPA, review the contractor agreement template once so it can be reused, and confirm the corpus licences before anything ships.

No training on commercial API data by default, but zero-retention eligibility is unconfirmed. Get it in writing and attach it to the DPA as the subprocessor record, before the pilot.

Still outstanding. Six edges are flagged suspect and undecided, and the registry's six EDGE hypotheses were chosen by graph topology rather than by anyone who teaches ninth grade. The form exists and is ready to hand over.

The tool is live with 24 seeded responses: blind rubric scoring, then reveal, then helps / should-defer / show-feedback verdicts, with teacher–model and teacher–teacher agreement computed automatically. Expands to 50–100. Produces the only teacher-appropriateness data that will exist.

FERPA-aware confidentiality, explicit IP assignment, no-external-publication, 1099 classification safeguards. Counsel reviews it once; reuse forever. Watch three things that will bite: F-1 candidates need written DSO confirmation before signing, funded RAs face outside-work caps, and W-9 before first payment.

Steward of ASAP 2.0 and PERSUADE going forward, and the source of the licence discrepancy you found. Worth a conversation regardless. Perpetual Baffour there works on bias in automated essay scoring. info@the-learning-agency.com