06-reference/research

ed fi vendor neutral ai learning evaluation layer

2026-10-02·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
edtechai-learningdata-standardsed-fiinteroperability

Ed-Fi covers the outcomes half, nobody covers the join: the vendor-neutral evaluation layer is still open

The question

"Does the Ed-Fi Alliance (or another vendor-neutral data standard) already close the 'vendor-neutral evaluation layer' gap for joining usage + assessment data across AI-learning tools?"

Context: this is the kill-or-confirm check on candidate problem 1 from the AI-for-learning network map, where the hypothesis was "the open gap is a vendor-neutral layer that joins usage and assessment data across tools."

Verdict: PARTIAL, and the half that is closed is the half that was never the hard part

So the gap the founder hypothesized is not closed. But it is narrower and differently shaped than "no standard exists" — the standards exist and are mutually unaligned, which is a different business than inventing a standard.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The check comes back confirming the problem, and the confirmation is more useful than a clean kill would have been. The reason is that it reframes the work from "build a standard" — a consortium job the founder has no business attempting — to "build the join," which is ordinary, hard, well-paid data engineering. Ed-Fi already holds the outcomes side with statewide deployments in a reported 14 states (unverified, as above) and the major student-information-system vendors integrated. Caliper already has the vocabulary for the usage side. What is missing is a governed bridge: identity resolution across a Caliper or xAPI event stream and an Ed-Fi student record, a conformed dimensional model over the pair, and a permission model that answers which human may see which learner. Ed-Fi's own paper names two of those three as open (the aggregation-unfriendly read shape and the application-level permission model) and schedules a summer 2026 interest group for them. The identity-resolution step across a usage-event stream is our inference; the paper never mentions usage data, Caliper or xAPI. Only the first two are safe to attribute to the standards body in front of a buyer.

The uncomfortable part is the emitter side, and it is the part that decides whether this is a business. An evaluation layer with nothing to evaluate is a schema. The vendors with the usage data (MagicSchool, SchoolAI, Khanmigo; district counts in [[2026-09-23-ai-learning-network-map-v0]]) have every commercial reason to export rosters in and no reason to export comparable per-learner engagement data out to a neutral judge that could rank them against a rival. Rostering standards spread because they reduced vendor onboarding cost; an evaluation standard increases vendor exposure. That asymmetry, not the schema, is the real moat problem, and the only forces that beat it are procurement mandates. Louisiana tying 50% of tutoring contract value to outcomes and the Gates Foundation gating scale on independent evaluation are exactly those forces, which is why those two facts in the network map are the load-bearing ones for this candidate rather than color.

Against the founder's current stated direction, this lands awkwardly and should be said out loud rather than smoothed over. The declared moonshot space is healthcare and life sciences ([[01-projects/hcls/index|Healthcare / Life Sciences: the moonshot space]]), RDCO is dormant, and phData is the main bet with credibility aimed at phData sales. This problem is in education, not HCLS, and the backlog item mislabeled it as HCLS. Its honest value is therefore not "candidate venture" but two smaller things: it is a genuine phData-shaped data-integration problem in a vertical with public money and a mandate forming, which makes it usable as credibility material; and it closes one of the three candidate problems on the network map with a documented answer instead of an open question. If the founder wants to spend more on it, the next move is not building anything. It is a half-day check on whether the summer 2026 Ed-Fi interest group produced output, and a direct pull of the 1EdTech certified-product directory to convert my soft adoption read into a hard one.

Worth noting for completeness: xAPI and a Learning Record Store are the obvious third candidate for the usage half, and I did not reach a primary xAPI source within this round's caps. My prior is that xAPI's near-total vocabulary flexibility is why it does not solve the join — any two vendors can emit conformant statements that will not reconcile — but that is an untested prior and is flagged as such below rather than asserted here.

Why this is in the vault

This closes the explicit "Open questions / next pass" item in [[2026-09-23-ai-learning-network-map-v0]] — "Check existing vendor-neutral data and evaluation layers (for example the Ed-Fi Alliance standard) before treating problem 1's gap as open" — and changes candidate problem 1 from an unverified hypothesis to a PARTIAL verdict with a named missing piece (the usage-to-outcomes join and its emitter incentive problem). Read it before any networking meeting where the "outcomes evidence for school AI" thesis would be pitched, and before re-running the network map's quarterly refresh.

Open follow-ups

Related

Sources

Vault:

Web (primary unless noted):

Not reached within this round's research caps, flagged as gaps: the 1EdTech Certified Product Directory itself, and any primary xAPI or Advanced Distributed Learning (ADL) specification source.