MORNINGSTAR: appendix
Self-reported metrics · schema · limits
Supporting notes for the MORNINGSTAR case study. Everything here is self-reported by the system; none of it is an independent quality measure.
Reported deliberation counts
| Measure | Reported |
|---|---|
| Deliberations | 13 |
| Decisions | 12 |
| Unanimous | 8 |
| Split | 3 |
| Tie-breakers | 0 |
The three outcome categories add up to 11, not 12, so the metrics file is not internally consistent. The discrepancy is recorded rather than smoothed over.
One model plays every seat. Unanimity therefore says little about deliberation quality: it may reflect a consistent persona rather than agreement among independent judges.
Case and proceeding schema
| Field | Shape |
|---|---|
| Case ID | 2026-DEL-005-001: year, category, case number, deliberation number |
| Feasibility tiers | F0 to F5; F3 and above require deliberation, F4 a transcript, F5 a full record |
| Proceedings | Standard, expedited, special-interest hearing, contempt hearing |
| Special-interest hearing | Testimony and findings, no final vote |
The feasibility labels are not fully consistent across the rules files: courtroom/RULES.md calls F0 “Trivial” while core/MFAF.md calls it “Infrastructural”.
What has not been measured
| Missing evidence | Why it matters |
|---|---|
| A single-prompt baseline | Without it, the courtroom structure has no measured benefit |
| A second model or independent judges | Unanimity from one model is not agreement |
| A tamper-evident log in the live path | The hash-chain library exists but the runner does not call it |
| An external evaluation of rulings | The system grades its own output |
Status
The July 2026 docket and the 13-deliberation metrics file are the only outcome records. The overdue F4+ specialist review is documented in state/current.md. Both are reported as-is, with no quality claim attached.