ARIN22 Decision Council
ARIN
Twenty domain agents. One accountable answer.
The council comprises 20 domain agents and two meta-agents. One Governor weighs their assessments, records disagreements and publishes a verdict with its reasoning trail. Agents interpret; the kernel computes. A narrative that contradicts the numbers is blocked from release.
Overview
A governed verdict, not a chat.
ARIN is the accountable narrative and decision layer around deterministic evidence. It does not originate risk numbers; it interprets the computed state and either publishes a traceable verdict or refuses.
One question, one answer
PASS / WARN / FAIL / INCONCLUSIVE for data-quality and risk-control checks, with the evidence and confidence boundary attached. No trading instruction or investment recommendation.
Agents never touch the numbers
All risk math comes from the deterministic kernel. The council interprets computed evidence; it cannot originate or alter a single figure.
Refusal over false certainty
If no agent clears the confidence floor, the Governor refuses to answer and routes to human review. A refusal is a feature, not a failure.
Audit forever
Every assessment, disagreement and decision lands in an append-only SHA-256 chain. Seven-year retention. Any historical verdict replays exactly.
How a verdict is made
The disagreement is part of the product.
The page stays intentionally inspectable: parallel agents, recorded dissent, math-to-narrative audit, codified HITL gates and memory that only recalibrates across releases.
The council convenes.
The lane-specific matrix selects voting, triggered, observer and non-applicable roles from 20 domain agents and two meta-agents. Each participating agent carries its own reasoning trail. The Governor resolves the verdict; the full council does not vote on every question.
Disagreement is published, not hidden.
Every verdict ships with a dissent map: who agreed, who disagreed, who abstained – and why. When the council is genuinely split, the Governor publishes both positions with a CONTESTED flag.
Hallucination is blocked, not reduced.
Before publication, every narrative passes a seven-rule audit against the deterministic state: no invented cascades, no fabricated regime shifts, no severity inflation – and no soft-pedalling a critical verdict either.
Some verdicts must reach a human.
Escalation triggers are codified, not operator-defined: dissent above 25%, a high-confidence FAIL on a data-quality or risk-control check, or any agent flagging critical-tier risk. Those verdicts stop at a human gate before release.
The system remembers.
Entity Memory carries every prior assessment into the next one: trends, anomalies and score deltas. Assessments are recorded with their source evidence; later evidence informs versioned, replay-tested calibration releases, never silent runtime drift.
Five things that cannot be turned off
Structural invariants, not settings.
The controls are not preference toggles. They define what a publishable ARIN verdict is allowed to be.
Sovereign by default
A complete verdict cycle can run without an external API call.
Five LLM lanes with automatic fallback – from local inference down to an offline quantised floor. Every lane is gated by guardrails for compliance, safety and factuality.
- Regulatory boundary: ARIN is an orchestration, narrative and recordkeeping layer. No regulatory validation or endorsement of the council is claimed.
- Model-risk support: Evidence mapping to SR 26-2 (Apr 17, 2026) concerns deterministic kernel outputs, not a certification of SAA or its agents.
- Records: council outputs follow EU AI Act Article 12, DORA and SEC 17a-4 retention discipline.
Request access