The Build

Learning Signals

Everything the student side must capture from day one — and the principle that governs how: capture at the platform's target shape, not a release-one minimum, because retrofitting a signal's shape later is an expensive migration.

The signal inventory

The signals the dev team must capture fall into two groups: the ones the brief named, and the ones that are equally required but easy to overlook.

Named signals	What each captures
Feedback	Every diagnostic feedback surfaced — with the anchor that produced it, whether it was authored or Companion-composed, the evidence record it was composed against, and the learner's engagement (read / dismissed / acted on).
Retries	Every Evidence Check retry, with the first attempt preserved verbatim, each subsequent attempt as its own record, the interval, and any remediation completed between attempts.
Confidence	Self-reported confidence at every check, plus confidence-response latency; the learner model tracks calibration over time as a persistent signal.
Pace	Time on task per node, per Evidence Check field, per session; session frequency and duration; time between milestone unlock and submission; pace against cohort baseline — captured raw and interpretively.
Remediation loops	Every remediation triggered, which option the learner chose (recommended / alternative / none), the outcome, and whether the next check resolved the diagnosis. Loop completion is a first-class event.
Interaction history	Every interaction as an event stream — node entries/exits, check attempts, Companion messages, menu choices, milestone views, artifact revisions, acknowledgements. This is the audit spine; every other signal references it.

Also required	What each captures
Response & reasoning	The primary evidence signals at every check. Reasoning is captured verbatim and is the primary diagnostic input, not a supplementary field.
Behavioural signals	The full Behavioural Signal Layer inventory — timing, paste, typing rhythm, tab focus, AI-fingerprint scores, stylistic consistency, steering-attempt events. Feed diagnosis in release one and populate the baseline from the start.
Process evidence at summative	Decision-log entries, revision history, AI-use disclosures, acknowledgements — captured throughout artifact production, not at submission. The trail is the evidence.
SOLO depth judgment output	The depth judged per check, with the markers that supported it — tracked per KC over time.
Misconception state transitions	Every transition (none → suspected → confirmed → cleared) with the check that produced it and the detection signal that matched; blocker activations/clearances at milestone level.
Confidence calibration state	A persistent per-learner signal — how well confidence tracks correctness — distinct from per-submission confidence.
Companion interaction patterns	Per node and cumulatively: interaction count, message-length distribution, steering-attempt count, engagement with authored vs. dynamic responses.
Learner agency signals	For every routing recommendation, which option the learner selected. First-class because the platform is architected on participatory adaptivity.
Enrichment engagement	Whether enrichment was accepted, at what subtype, with what outcome — feeds Mastery Credit.
Milestone readiness signals	Per milestone, each readiness condition's state at each check (satisfied / pending / blocked) and the learner's engagement with unmet conditions.

The same-shape principle

Capture vs. use. Some signals are used from day one (feedback, retries, confidence, response, reasoning, misconception state, remediation loops, interaction history, agency signals, milestone readiness). Others must be captured from day one because their use starts now but matures later — behavioural signals feed diagnosis in release one and populate baselines later capabilities depend on; process evidence is used for SME grading now and feeds contribution pipelines later; SOLO output is used for routing now and feeds cross-course transferable-mastery tracking later.

Storage, access & privacy

All signals are stored in Maestro's own data layer as part of the learner model and event history. They are available to:

• the SME — through cohort monitoring, grading interfaces, and review queues;
• the content-development team — as aggregate patterns to inform authoring;
• higher management and university admin — in aggregate, per role permission;
• the trust-layer components — Diagnostic Matcher, routing engine, misconception state machine, Behavioural Signal Layer, milestone readiness evaluator — as inputs.

They are not exposed to the learner as diagnostic data (the learner sees their outcomes and their choices, not the underlying signals), and they are not exported to external systems as raw signals. What leaves toward Moodle is summarised status and score, per the AGS contract.