Select a Case

Loading cases...
Select a case to review the setting, participants, and any optional reference dialogue.
▸ Customize (optional)
Temperament layered on the case: starting state, escalation speed, and direction. Random is revealed in the debrief.

Past Sessions

Loading sessions...
Agent pipeline & skills map
Hover nodes for details · Click ★ agents to inspect skills
1
Collect Training Sessions
Run completed simulations first. Skills Studio defaults to completed dialogue and can filter to unused data only.
2
Review Proposed Changes
Residents can inspect source sessions, approve or reject proposed edits, and refine prompts before any system change is made.
3
Admin Promotion Only
Approved changes stay in review until an admin account merges them into the active skills on disk.
Orchestration Actors Evaluators Self-improvement Infrastructure ★ = has SKILL.md · click to inspect PER-TURN LOOP reads state state feedback guidance guidance K cands. K cands. passing next turn SKILL.md → agents LOOP 2 / 3 — SKILL.md SELF-IMPROVEMENT Case JSON scenario · case data Conversation State emotion · pattern · trust Advantage · Select Best GRPO score → best candidate → history Storage JSONL · SQLite · GCS Batch Profile failure patterns Orchestrator turn loop · state management Expert Coach coaching-rules · learner-pitfalls ★ Parent Actor parent-acting · emotions · escalation ★ Clinician Learner sic-framework · learner-pitfalls ★ Hard Filter word count · clichés · role violation · jargon Judge: Realism parent voice · naturalness judging-realism · weight 0.35 ★ Judge: Grounding topic consistency · coherence judging-grounding · weight 0.40 ★ Judge: Educational SIC skills · pitfall detection judging-educational · weight 0.25 ★ Skill Editor batch profile → SKILL.md delta self-improvement core ↑ ★ Validation Gate keep if improved · else rollback
Change requests
Loading…
1 · Select training sessions ▼
Loading sessions…
2 · Target agents

Dialogue Review

Pre-generated serious illness conversations between the AI clinician and the AI parent actor. Read the dialogue and answer the Revised Rubric 2.0 questions in the panel beside it. Both speakers are simulated: the clinician follows (or deliberately breaks) the SIC Guide; the questions are about the parent actor's believability.

Supervisor Review

The supervisor agent watched each dialogue and wrote turn-by-turn feedback plus an overall debrief for the clinician. Its comments appear under each clinician turn. Read what the supervisor said about the good and the bad clinician, then answer the short questions beside the dialogue about whether you agree with it.