Cognivia Platform
Marks tell you who passed.
Not why anyone failed.
A teacher on one shared classroom computer can see which idea is quietly slipping across the class, and which wrong answer is a confident misconception rather than a gap. Cognivia reads the timing and confidence behind each answer and hands back a memory state you can act on, through one API any tool can call.
Not a medical or clinical diagnostic. An exploratory learning instrument.
See it work
Answer five questions. Watch it read why.
Try one below: answer, then say how sure you were. The demo reads the two together and returns the same kind of diagnosis the API hands a partner app, live, with nothing stored. Then run the full five-question version.
Which blood cells defend the body against infection?
Synthetic sample data. This runs the same error-type logic the engine uses, in your browser, and writes nothing.
The problem today
The score records the output and throws the reason away.
A teacher sees a wrong answer and a mark. What they cannot see is whether the student forgot something they once held, is missing a prerequisite they never had, half-understands it, or is confidently wrong in a way more practice will only harden. Those four need four different responses, and no gradebook tells them apart. So an hour of review gets spent evenly, where it is needed and where it is not.
How Cognivia works
Measure, diagnose, act, then check it landed.
Read the answer
Every answer, how long it took, and how sure the student felt feeds a forgetting model built for that learner.
Name the error
From timing and confidence, the confirm-or-clear check sorts the miss into one of four kinds.
Pick the fix
Each kind routes to the move it needs: retrieval, a step back to a prerequisite, worked examples, or a counter-example.
Verify
The fix is re-checked at the next review, so one that did not land gets caught instead of assumed.
An example result
One event in. Diagnosis, memory state, next move out.
A partner app sends what a learner just did as a Learning Trace. Cognivia returns the reading, with its uncertainty attached.
GET /v1/learners/:id/memory-state
{
"concept": "Photosynthesis",
"error_type": "confident_misconception",
"retrievability_now": 0.91,
"fsrs_stability": 7.6,
"accuracy_ci95": [0.21, 0.94],
"next_move": "contrasting_counterexample",
"next_review_at": "2026-09-09"
}
Limitation: every inferred field ships with its uncertainty. When the evidence is thin, the field returns as unknown rather than a guess. The wide interval above is the system saying it is not yet sure.
What is and is not established
The evidence, labeled.
The instrument runs today: it builds a 10-variable Learning Genome per learner from their own events, schedules with FSRS-6, and returns a diagnosis and memory state through the API, each with uncertainty.
The formulas and thresholds are engineering defaults, seeded from a small pilot (n = 5) and the literature. They are how the system computes today, not validated constants.
Whether error-type-aware remediation actually beats adapting on correctness alone is the question a pre-registered trial is testing. No efficacy percentage is claimed anywhere; findings are withheld until enough participants have contributed.
Why now
Three things changed at once.
Shared low-cost devices now reach classrooms that will never get one laptop per child, and Cognivia runs on a single browser tab. A century of cognitive science is finally precise enough to read one learner in real time, not just population averages. And the new wave of AI tutors can generate endless practice but has no memory of what each learner actually holds. Cognivia is the layer that supplies it.
How it is different
One layer upstream of every study tool.
Anki, Cerego, and Quizlet are delivery tools: they show content and schedule reviews, and they do it well. They all work from the fact that an answer was wrong. Cognivia works out why it was wrong and sorts it into a type, then hands that reading back so a review tool can schedule better. The difference is mechanism, not marketing: producing the diagnosis is the part none of them do, and it is what the API returns.
What is live, what is planned
No maturity theatre.
Every capability is labeled honestly. Only the ones marked live are available today.
The in-browser demo; the REST API (/v1 traces, diagnose, memory-state, analyze, learners, health); API keys with mint, revoke and usage (sk_demo and sk_live); rate limiting (120/min); CSV export; the PDF session report; the classroom with join codes, roles and a consent gate; and published npm packages (@cognivia/cli, @cognivia/mcp, @cognivia/core).
The classroom map and Cognitive Passport run today, but on small or sample data; treat them as early, not as a proven longitudinal record.
LTI 1.3, single sign-on, roster sync, webhooks, and a Python (PyPI) package. Designed, not yet built; they are labeled planned everywhere they appear.
Trust
What we hold, and what we will not do.
Only what the learning needs: which answers were right, how long they took, and a confidence rating. Records are stored under a code instead of a name, on access-controlled servers, never sold and never used for advertising. For anyone under 18, nothing is processed until a parent or guardian consents. Cognivia is not a medical, clinical, or psychological diagnostic and does not claim to understand a person's mind.
Trust, data & scientific status · Privacy questions · Privacy policy
Why this founder
Built where the problem is hardest to solve.
Cognivia comes out of teaching in rural classrooms and hitting the same wall every day: you can see exactly what a student did and still have no way to see why. It is built by one person, from Kanpur, to run on the shared devices those classrooms actually have. Read the founder's account →
The pilot, concretely
A founder-led pilot, scoped small on purpose.
No procurement, no platform migration. One class, one unit, and a written read-out at the end. Students are always free; a pilot partner is a limited, hands-on cohort while the product is early.
weeks
Long enough to see a two-week delayed-retention check land.
subject or unit
One curriculum unit, one teacher or program lead. Kept narrow so the signal is clean.
checks
A baseline and a delayed-retention check, so there is a before and an after.
weekly support
Weekly founder support, and written findings at the end, shared with you first.
Not offered yet: procurement-free institutional roll-out, uptime terms, or priority support. Those come after the pilots prove out.
Start with a class, or a key.
A teacher can create a class, share the join code, and watch the live map in minutes, with no procurement. A developer can get an API key and send the first event today.