From the current synthetic debrief
One observation. Not a complete assessment.
Teacher notices → AI proposes → teacher decides
The claim-proposal step in this demo is a deterministic heuristic, not a separate model call.
Beyond a result. Toward a next step.
A grade tells us
how they did.
Where do we
go next?
An Educational Navigation System helps teachers connect what they observe, build a revisable picture of a learner’s understanding, and choose what to explore next.
See one small partSupporting the teacher’s judgment. Not replacing it.
The broader ENS vision · AI-generated conceptual artwork, not an implemented recommendation engine.
The larger vision. One bounded demonstration.
Start with
one moment.
Today, we’re showing one small part of ENS: turning a teacher’s classroom observation into a claim they can inspect, question, and decide whether to keep.
Follow this observation →A small slice. A larger direction.
See more clearly.
Choose what
comes next.
This slice keeps a classroom observation connected to the claims made from it—and leaves admission with the teacher.
The larger ENS vision is to build that picture over time, so teachers have more to draw on when choosing the next learning opportunity.
Proposed next steps · beyond this slice
From a demonstration to something useful.
With the help of Meta · beyond this slice
What provisioned support would unlock.
Three Meta models would grow this slice — each teacher-chosen, each feeding the same gated pipeline. All advisory, all inspectable, the teacher still decides.
The hosted route already works end to end on synthetic data; provisioned, isolated infrastructure would make it the everyday route.
The next step stays human.
The Vision →Educational Navigation System / the long-term ambition
The Vision
A clearer picture of learning.
For everyone who helps it happen.
Empowered Teachers
Connect classroom observations to a richer understanding of each learner—with evidence they can question and judgment that remains theirs.
Involved Parents
Understand the story beyond a grade and take part in meaningful conversations about their child’s learning, with appropriate privacy and sharing boundaries.
Guided Learners
Recognize progress, ask questions, and explore possible next steps—with support, agency, and room to grow.
Learning Institutions
Learn from patterns in teaching and learning to improve support—without reducing people to labels or rankings.
Learning moves forward. Together.
AI-generated concept art · long-term vision, not capabilities demonstrated in this sliceGuided walkthrough / the live slice
What does this observation support?
Use the actual debrief, channel admission, and human decision below. AI outputs propose; people decide.
1 · The teacher's debrief
Synthetic class: . Edit the debrief, or use the sample.
2 · The Nine Gates — channel admission
The channel is post-class narrative debrief → per-learner sightings. Gates 1–8 pass in minutes. Gate 9 asks where extraction runs.
3 · Manual sighting entry — model off
Switch the model off and record what you observed, by hand. Pick the learner and type the observation in your own words. It flows through the same verification, sign-off, and warrant trail as the automated path, but under a teacher-authored posture: the text is yours, not a quote a model pulled from the debrief. No model runs.
4 · Sightings → claims → the Twin
Evidence and interpretation stay separate. A sighting quotes the debrief — and code, not the extractor, verifies the quote is verbatim before it counts. A claim interprets sightings, waits for teacher sign-off before it enters the Twin, and can be contested, never deleted. The heuristic prior decays at read time; drag the slider to see it.
Advisory only (DecisionJudge, D-016): a System-One-style harness over the same local Glimmer. It never admits, signs, or blocks. Slow on this hardware — about 30 s per claim.
5 · The warrant trail
Every consequential act leaves a crumb (Warrants v0 shape, ens.* namespace). Walk it: each claim traces to its sighting, its extraction, and the gate decision that allowed it.
Agent Console / same session, same evidence
Now, show your work.
The mechanism behind the teacher walkthrough. Changing views never starts a second run.
Agent Console — the technical layer
The same run the teacher sees, as turns. Step through turns pauses after each phase so a presenter (or judge) can inspect what the model saw before the run continues.
Verification — code, not model
No verification has run yet.