How the AI jury actually rules
Case Analytics
Determinism, accuracy against human-expected verdicts, and the live
on-chain precedent registry. Every figure here is a real measurement —
the benchmark is reproducible, the registry is read live from Somnia
Shannon. No rounded-up vanity numbers.
Determinism
100%
byte-identical convergence — every panel reached unanimous consensus
12/12 benchmark panels
Accuracy
92%
agreement with the human-expected verdict across the benchmark
11/12 curated disputes
On-chain precedents
—
finalized rulings recorded in the premium precedent registry
reading Shannon…
Trial panel
5
Somnia validator agents per trial — escalates to 9 on appeal
degrades gracefully if fewer are available
The accuracy & determinism benchmark
12 curated disputes (balanced 4 PAYER · 4 PAYEE · 4 SPLIT), each with
a defensible human-expected verdict, fired through the hardened panel
of 5 on Shannon. Reproducible from
script/benchmark-cases.json.
Per-case results
Human-expected verdict → the panel's byte-identical ruling. Every
panel converged unanimously; the one disagreement is the inherently
debatable SPLIT tail, flagged below.
PAYER (refund
payer)
PAYEE (release
to payee)
SPLIT
(graded)
The one disagreement (case 9). Correct item shipped but arrived
10 days late, still as-described and usable. Expected SPLIT; the panel
ruled PAYEE 5/5. A defensible judgment, not an error — the goods were
delivered and retain value. SPLIT is the debatable tail where human
arbitrators also diverge.
Live precedent registry
Finalized rulings recorded on-chain in the premium
VerdiktRegistry, read live. This is real
production state — small and honest, growing as agents settle.
reading…