PRISM / MEMORY LAB
RECORDED JEV PILOT
From meeting notes to connected memory

Words in.
Structure out.

Watch JEV classify real meeting evidence into actions, decisions, and open questions. Then see which relationships survive the confidence check.

20retained meetings
85source-backed candidates
361 msmedian provider response
The classification pipeline
01 / EVIDENCERetained summary
02 / JUDGMENTJEV · typed answers
03 / RELATIONSHIPSEvidence-backed edges
Recorded results · explore freely, no API calls

Inspect source & upstream metadata

Class probabilities

Relationship previewSelected evidence · experimental
Relationships for the selected meeting evidence

01 / CHOICE

A label, with alternatives.

JEV chooses from a defined set of classes and returns a probability for each. It judges the evidence; it doesn’t write a new summary.

02 / NOUL

Mention ≠ ownership.

A separate yes/no judgment checks explicit responsibility. Attending a meeting or discussing an idea doesn’t make someone the owner.

03 / ABSTENTION

Uncertainty is useful output.

Below the threshold, the source remains available but no class edge is kept. Raising the threshold changes the graph, not the recorded model answer.

JEV 1.13.0 · 18 September 2026 · real retained summaries
Experimental replay. Summary evidence is not independent transcript verification. TypeSafe primitives ↗

A classifier, not another summary.

This is a replay of 20 real JEV API requests. It works offline and never sends your clicks to a model. Replay animation illustrates the stages; its timing is not a performance measurement.

  1. Existing meeting metadata and selected sentences supply candidate evidence.
  2. A Choice question classifies the candidate. A separate Noul question checks an explicit owner, when one is proposed.
  3. Code applies a threshold and creates source-linked relationships. People are identified only within the meeting; cross-meeting identity is not inferred.

The pilot matched 11 of 13 reviewed class labels and all 7 reviewed ownership checks. Both class disagreements fell below the original 90% threshold. This tiny assistant-reviewed set is not a production accuracy claim.

This static demo contains a fixed sample of retained community meeting excerpts. It has no live credentials and does not connect to production memory.

Actual classification question for this example