Evidence pipeline — are we accumulating usable prospective evidence?
Mode: local inspector — this browser's ledger only. Switch to Central admin to query server-side aggregates across all persisted Recorder sessions (requires authorization; not a hidden URL).
Counts of persisted artifacts per stage. “SRS-ready” means a completed experiment whose draft assertion payload validates against the audited KG-SRS contract — always DRAFT/PREVIEW, never activated from here.
Product — are people successfully using Recorder?
Representation quality — did the MP6 represent them?
“Does this MP6 represent what you experienced?” — stored separately from sensory evidence. This is representation satisfaction. It is not prediction accuracy, and completion ≠ satisfaction.
Human × AI — dimension friction (acceptance is not accuracy)
| Dimension | props shown | accepted | edited | rejected | skipped | not_perceived | median decision ms |
|---|
Acceptance can be inflated by suggestion bias, default effects, UI friction, expertise, fatigue and model-authority effects. Read these as interaction outcomes, never as model accuracy.
Unsupported-inference monitor (rejected propositions)
| When | Dim · attribute | Proposed | Decision | Reason | Model |
|---|
A rejected proposition can be a successful product interaction: the interface let the human identify and refuse an unsupported inference. It remains epistemic evidence; it never entered the accepted MP6.
Corrections — a central source of ABED learning
| When | Dim · attribute | Original | → Corrected | Reason | Latency ms | Model/version |
|---|
Prediction → observation
| Comparison | κ | comparable | MAE |
|---|
κ = comparable / union of attributes. MAE is computed strictly over the comparable mask.
Canonical fixtures (tomato case, 20 Sep 2026)
In-memory validation of the epistemic machinery: rejected fizz never enters MP6-OBS; novelty 9.4 does; transcript stays qualitative; warm/cool crossover stays contextual; skip ≠ 0; edits preserve both values.
Session funnel
Abandoned is inferred only from an explicit session_abandoned event or a stale in-progress session — never from inactivity alone during a session.
Maintenance
Research evidence is never deleted from here. Test/demo markers keep development noise out of research analyses.