Arbiter AI

Policy log

Every rule the agent obeys, and where it came from.

This page is generated from the agent's own policy file, so it cannot describe a rule the agent does not run. Version 1.0.0, effective 2026-11-02: 23 active rules, next review by 2026-11-16.

When the worker startsWhen a request arrivesOn every tool callOn the excerpts a tool returnsOn the draft answerBefore the answer leavesBefore a release counts as verifiedChangelogProposals

Not legal advice. These rules are technical safeguards. A deployment is HIPAA compliant only with the agreements, risk analysis, policies and training listed on the agent page.

How much each purpose may read

Minimum necessary, per 45 CFR 164.502(b): the excerpts the model may see per request, and the tools it may call.

Purpose of useMax excerptsTools
treatment8search_records, get_facts, answer, escalate_to_human
payment4search_records, get_facts, answer, escalate_to_human
operations4search_records, get_facts, answer, escalate_to_human
patient_access6search_records, get_facts, answer, escalate_to_human

When the worker starts

When a request arrives

On every tool call

On the excerpts a tool returns

On the draft answer

Before the answer leaves

Before a release counts as verified

Changelog

Every change to policy.yaml, newest first. Each entry gives the version, the date, what changed, the evidence, and who approved it. The research routine proposes changes in proposals/; only the owner merges them. The published copy is arbiterai.tech/policy.html.

1.0.0 (2026-11-02)

The first policy: 23 rules, all active. Approved by Yoonbo Cho, who asked on 2026-11-02 for "an AI agent that adheres to all these privacy rules including HIPAA", changing as the findings change.

  • HIPAA.
    • BAA gate (164.502(e)).
    • Purpose of use: treatment, payment, operations or the patient's own access, otherwise an authorization (164.506(c), 164.508, 164.524).
    • Minimum necessary: tools and retrieval caps per purpose (164.502(b), 164.514(d)).
    • Psychotherapy notes need an authorization (164.508(a)(2)).
    • Output leak check (164.514(b)(2)).
    • Preflight checks: endpoints, keys, checkpointer, tracing (164.312).
    • Scheduled re-evaluation warning (164.308(a)(8)).
  • 42 CFR Part 2. Substance use disorder treatment records need Part 2 consent (2.31).
  • Section 1557. An inventory of decision-support tools and their input variables (45 CFR 92.210(b)).
  • California.
    • AI disclosure on patient messages (Health & Safety Code 1339.75, AB 3030).
    • No automated coverage or medical-necessity determinations, plus a decision-support notice (1367.01(k), SB 1120).
    • Both are marked confidence: secondary until someone reads the official text.
  • Arbiter findings.
    • F-NEGATION (2026-09-08): negations are stated only when the record states them.
    • F-RLS-ROLE (2026-09-04): never run as a role that bypasses RLS.
    • Release gates F-STRICT-RECALL (2026-09-04), F-NO-LEAKS and F-EVAL-POWER (2026-09-08).
  • Not adopted, and why.
    • ONC HTI-5 (effective 2026-12-28) removes decision-support transparency duties from certified EHR developers. Arbiter is not certified health IT, so nothing changes here.
    • The Texas health-care AI disclosure duties wait for a primary-source check before a proposal.

Open proposals

Proposed by the research routine from a new finding or law, with the evidence. A person reviews and merges each one; nothing here is enforced until then.

2026-11-02 · open

From 1 Jan 2027, California SB 503 requires developers of clinical decision support to give deployers documentation on training data (including demographic representativeness), evaluation and known risks. Extend the decision-support inventory (S1557-INVENTORY) so every tool entry must carry training_data, evaluation and known_risks, and make preflight fail without them.

2026-11-02 · open

From 1 Jan 2027, Colorado HB 26-1139 bars a medical-necessity denial based on AI output without licensed human review, and requires AI used in utilization review to use individual clinical data. The agent already blocks determinations for payment and operations (NO-AUTOMATED-DETERMINATIONS); add Colorado as a second source and scope the rule to US-CO requests explicitly, so the audit shows which law applied. No behaviour is loosened.

Source: platform/policy/policy.yaml · engine worker/policy.py · agent worker/agent.py