Request
The NIST AI Agent Standards Initiative is developing an interoperable and secure foundation for agents that act on a user's behalf. This open request asks NIST to evaluate a complementary technical object: a portable, versioned decision record.
DRCS-EC is a candidate for a bounded public pilot. It is a working specification whose field performance, privacy properties, interoperability, and production cost require testing.
Why a decision record belongs in the standards discussion
Authentication can show which agent acted. Authorization can show that the user delegated a task. A transaction receipt can show the selected outcome. These objects usually do not show the serious option set, rejected alternatives, scoring trace, feed provenance, or the agent's prediction before it acted.
A common record shape lets independent authorized evaluators apply different tests to the same evidence. The standard defines the evidence interface and conformance rules. Institutions retain authority over their evaluation rules and consequences.
This separation maps to the NIST AI Risk Management Framework's Govern, Map, Measure, and Manage functions: governance sets authority; mapping defines context and affected parties; measurement tests record and outcome; management supplies response, review, and correction paths.
Proposed technical work program
- Core schema: stable identities; schema, policy, model, scoring-rule, and feed versions; delegation; authority; evidence; uncertainty; prediction; alternatives; stakeholders; system welfare; review and reversal conditions.
- Consumer e-commerce profile: seller and product identity; total landed cost; shipping, returns, payment, reliability, affiliation and placement signals; selected and rejected options; scoring trace; and outcome follow-up.
- Conformance: normative JSON schema, deterministic fixtures, pre-decision creation tests, scoring-trace reproduction, import/export, version migration, integrity profiles, and negative fixtures.
- Privacy and security: purpose-based minimization, selective disclosure, reader authorization, retention, correction, appeal, and threat models for fabricated alternatives, feed substitution, omission, and post-decision rewriting.
- Testing and evaluation: instruction adherence, steering, omitted serious alternatives, score reproducibility, feed provenance, outcome follow-up, and repeated cross-provider patterns, with false-positive and false-negative boundaries.
Evidence and limits
The findings show why conformance and consequence must be tested separately. A NIST pilot could supply the missing deployment evidence.
- The evidence path alone has no pricing effect under common seeds.
- Forecast scoring has no pricing effect in this implementation.
- Welfare-priced reward changes the private objective and conduct.
- Outcome-contingent restrictions change conduct at higher tested audit intensities; targeting matters there relative to dose-matched random interruption.
- The rejected-alternatives rule and several-reader score add no effect because their implemented triggers do not activate.
- The experiment does not establish field performance, prevalence, legal sufficiency, durable filing, reader diversity, privacy adequacy, or production cost.
Requested outputs
- A public problem statement and use-case boundary.
- A draft DRCS Core and DRCS-EC crosswalk to the AI RMF and current agent standards work.
- A normative schema, conformance corpus, producer, validator, and human-readable renderer.
- A privacy and security threat model.
- An interoperability exercise across independent implementations.
- A field-pilot protocol with registered success and stop criteria.
- A report of negative results and inactive tests.
- An open issue, correction, and revision ledger.
Sources and downloads
Corrections and supersession
This is version 1.0. Corrections will be listed on this page and in the site corrections log. Any successor will identify the exact version it replaces, preserve this page and its downloads, explain material changes, and link to the superseding version.
Current corrections: none.