Cross-session analysis (XSA)

Cross-session analysis finds repeated strategies, failures, and behavioral patterns in producer traces. It complements the Experiment Overview: Overview compares outcomes and metrics, while XSA reads session evidence and produces investigator findings.

1. 1
   Finish the execution roster

   Every selected rollout needs a terminal Runtime identity and at least one Attempt before Registry can freeze the analysis roster.

2. 2
   Open Analysis

   Open the Experiment, select Analysis, then choose Analyze experiment. Registry owns the results project, investigator prompt, and execution policy for this path.

3. 3
   Review the child count

   Registry groups sessions by immutable Task digest and plans one investigator Review per Task-local shard. More than ten expected child Reviews requires explicit confirmation before any child model work launches.

4. 4
   Follow Task reports

   The page shows parent progress and one searchable row per immutable Task, with session counts, missing traces, shard status, failures, and report links.

One child Review contains sessions from exactly one immutable Task digest. A child can analyze at most 16 sessions, so a Task with more sessions is split deterministically into separate shards. Open each completed report for its summary and evidence-backed findings.

Registry freezes the exact TaskSet and the complete terminal Run, rollout, Attempt, and seat roster. A trace that is not indexed yet remains explicit as waiting_for_trace; it is not silently omitted. If one trace for a Task is missing, none of that Task's shards launch until every trace for the Task is readable. Other Tasks can continue independently.

• Retry failed or missing tasks re-reads missing traces against the same frozen roster and requeues only failed shards. It never adds new session identities.
• Cancel analysis stops waiting or queued children and requests cancellation of active child Reviews; completed reports remain.
• Start new analysis creates a distinct parent after the current analysis is terminal, freezing the Experiment roster again instead of mutating the earlier analysis.