Product · OutFigure

§ 1 Built

What exists today

Evidence record Built: Every event is stored immutably with its receipt. Each response is annotated with whether it counted and why not, under versioned eligibility rules (eligibility-v1).
Learner state Built: A deterministic summary per learner and concept (observed-v0): counted responses, correct responses, sessions and recency over the latest 200 counted responses. It stays insufficient evidence below 5 counted responses or 2 sessions, and becomes stale after 30 days without new evidence. Those are disclosure thresholds, not validated measurement thresholds. It has no forecast.
Readiness report Built: An offline command-line tool that reads a CSV or JSON Lines export and reports validation problems, missing context, comparable evidence and which analyses the data can support.
Partner API and TypeScript SDK Local preview: Idempotent event ingestion, processing jobs, state and evidence reads, with database-enforced separation between projects. It runs locally today; hosted sandbox access is in preparation.
Evaluation against outcomes In development: Predictions frozen before outcomes, simple baselines first, calibration, coverage and learner-clustered uncertainty. A candidate that loses to a baseline is reported as losing.

§ 2 Outcome audit

An audit before an integration

For AI-enabled, assessment-heavy professional, certification and technical learning products with returning learners. The audit asks what your existing data can support before anyone builds an integration.

You provide

An authorized, pseudonymized export of responses: event ID, learner ID, item and version, attempt, time with timezone, correctness.
Where available: sessions, assistance, hints, feedback visibility, delivery context.
Your item-to-concept mapping, if one exists. It is not required for the first findings.
Any later outcomes you hold, and the metric your team uses today.

You receive

A readiness report you can run yourself, with no learner identifiers in its output. Available
A baseline evaluation with calibration and uncertainty. In development
A reproducible report in which each conclusion names its evidence design. In development
A clear view of whether continuous measurement is worth integrating.

What an audit will not claim

Improved learning, when the data cannot show it.
That an intervention caused an outcome, from observational data.
That missing outcomes are failures. They are recorded as missing.
Anything based on synthetic data.

Discuss an audit

§ 3 Sample

What a readiness report says

This table is produced by the readiness report itself, run on a synthetic export of 577 events from 10 invented learners. It shows the format, not a result about anyone.

577 of 577 records valid 396 of 541 responses comparable 49 responses without reported assistance

Analyses a synthetic export can support, with status and reason
Analysis	Conclusion type	Status	Why
Validate records against the event contract and measure missing fields	descriptive	Supported	Valid events: 577 of 577 records.
Describe item difficulty and item-rest correlation, without concept mappings	descriptive	Limited	No item version has eligible responses from at least 10 learners, so statistics are withheld. Item versions with eligible responses: 49.
Summarize observed evidence per learner and concept (observed-v0)	descriptive	Supported	Learner-concept pairs with enough recent evidence to summarize: 17 of 71. With enough evidence that is more than 30 days old: 7.
Evaluate predictions of the next eligible response chronologically	predictive	Supported	Learners with a later eligible response: 9. Whether that is enough for a comparison is decided by a prespecified protocol, not by this report.
Find later unassisted observations of the same item or concept	descriptive	Supported	Unassisted observations 7 or more days later exist. They are re-observations, not controlled retention measurements.
Separate assisted performance from unaided performance	descriptive	Limited	Responses without reported assistance: 49 of 541. Only the others can be separated.
Compare later performance of learners who did and did not receive an intervention	observational	Not supported	The export contains no intervention events.
Estimate the causal effect of an intervention	experimental	Not supported	Causal conclusions need randomized or quasi-experimental assignment records, which a schema 1.0 export cannot carry (planned: EXP-01).
Evaluate predictions or interventions against delayed, transfer or external outcomes	predictive	Not supported	Schema 1.0 cannot mark a measurement as a delayed check, transfer task or external assessment (planned: OUT-01). Later responses are reported only as re-observations.

§ 4 Boundaries

Not on the roadmap

A tutor, chatbot or content generator
A mastery score or universal learner score
Rankings of people for employment
Biometric or emotion inference
Pooling one customer's learner data with another's
Automated high-stakes decisions about learners