RCD-1 — Reproducible Coverage Decisions

Why this standard exists

Coverage decisions — prior authorization, utilization review, claims adjudication, downcoding — are consequential, governed by written policy, and subject to appeal, audit, regulation, and litigation. A growing share of these decisions are now produced or materially assisted by software, including probabilistic artificial intelligence.

Courts have begun compelling disclosure of payer decision algorithms. States have begun requiring annual certification of insurer AI. Federal rules now require a specific reason for every denial. The direction of all of it is a single demand:

A decision system that cannot answer cannot be meaningfully appealed, audited, or defended. RCD-1 defines, in testable requirements, what a system must be able to show when examined.

The architectural principle

Reading clinical reality is inherently interpretive. Deciding under written policy is not. RCD-1 requires that systems separate the understanding layer (which reads and structures source material, and may use any technique, including large language models) from the decision layer (which determines outcomes, and must be deterministic, policy-bound, and reproducible).

The standard does not prohibit any technology. It constrains where probabilistic outputs may carry decision authority: nowhere.

Key requirements

R-1

The decision layer must be deterministic: identical inputs and identical policy revisions must produce identical outcomes, on every execution, indefinitely.

T-1 · T-3

Every decision-material field must trace to the specific policy revision and evidence behind it; adverse determinations must carry specific reasons at the criterion level.

P-1 · P-2

Policies must be immutably versioned, and every decision bound to the revision in force when it was made.

D-2

A qualified third party must be able to reconstruct any decision — and re-execute it — from the decision record alone.

U-2

Probabilistic outputs must not directly determine any decision-material field. A system in which a model selects or scores the determination itself is out of conformance regardless of measured accuracy.

H-1

Every decision has an identified, accountable human. Automated processing must not be the sole basis for an adverse determination.

Clause 7

The Re-Run Test. A random, uncurated sample of historical decisions is re-executed. Conformance requires identity on all decision-material fields for 100% of the sample. There is no statistical tolerance, because divergence is an architectural property, not a defect rate.

Conformance levels

Levels create an adoption path: L1 is achievable by well-run platforms today. L2 and L3 are architectural claims that can be verified, not asserted.

Regulatory alignment

RCD-1 is designed so that conformance produces, as a by-product, the evidence demanded by current and emerging law:

• CMS-0057-F (effective January 2026) — specific reasons for every denial; decision timeframes. → T-3, D-1–D-3.
• CMS guidance for Medicare Advantage — AI must reflect the individual's circumstances, not group datasets. → T-1, T-2, U-2.
• Alabama SB 63 (effective October 2026) — annual certification and accuracy monitoring of insurer AI. → Clause 5.7, Annex A.
• Indiana HB 1271 — no AI as sole basis for downcoding without professional review. → H-1–H-3.
• California SB 1120 and successors — clinician accountability in utilization decisions. → H-1–H-3, D-2.
• CMS-0062-P — the Working Group filed public comments (June 2026) urging reproducibility, criterion-level reasons, policy versioning, and human accountability in drug ePA.

Participate

The comment period for Working Draft v0.9 is open through September 2026. The Working Group invites co-conveners and reviewers from provider organizations, payers, professional associations, standards bodies, technology vendors, academic researchers, patient advocates, and regulators.

Version 1.0 will be ratified by the assembled Working Group. No single organization — including the convener — holds veto over its content. The standard is and will remain free to implement: no licensing fee, no certification monopoly, no membership requirement.