The company behind the architecture.

02 · Guiding Principles

Four ethical principles govern every requirement — beneficence, non-maleficence, autonomy, and justice — and emotional data is treated as sensitive by default. On that foundation the Standard adopts the NIST trustworthiness characteristics and adds one specific to this domain: eight characteristics every conforming system is held to.

Human-centered design completes the principles: people can know they are interacting with a machine, can pause or end emotional interaction at any time, and hold rights over their data. Continued engagement is earned through accumulated usefulness.

03 · Lifecycle

The Standard follows a system through its whole life — governance, context and risk mapping, data stewardship for emotional inputs, and security and privacy by design. At its center stand seven design requirements for the persistent-state cognitive core, stated at capability level.

1. Inspectable State

   The internal condition that influences behavior is computed from defined inputs, bounded, and observable to the operator at an appropriate level of disclosure.

2. Regulation by Design

   The system maintains its condition within a designed operating band and recovers toward it after disturbance. Regulation applies continuously and by degree.

3. State-Conditioned Processing

   Where internal state influences generation, that influence passes through defined, testable pathways — never through undocumented side channels.

4. Memory with Structure

   Retention is selective and organized. The design states what is retained, what softens, and what never fades — and why.

5. Bounded Autonomous Activity

   When the system initiates internal activity without a user request, the conditions of initiation are defined and documented.

6. Restraint Pathways

   The system can limit or decline its own behavior as a function of its condition and governance requirements.

7. Governed Decisions as Evidence

   Control decisions are documented when made and retained in durable, examinable form.

Around the core: operating context is documented before deployment, including the risks persistence introduces — accumulated state, long-horizon behavior change, memory as an attack surface; emotional inputs are collected only as the purpose requires, under a written retention policy; and memory derived from interaction with a person belongs to the relationship with that person.

04 · Evaluation

Evaluation is a continuing obligation, not a launch gate. Claims about the system are established by records — reproducible tests, matched comparisons, and retained evidence — and capability claims and evaluation status are stated separately. Six dimensions are evaluated.

Metric targets are set per deployment context and documented at design time — the Standard deliberately sets its targets where the system will actually run, in place of universal numeric benchmarks. Methods include controlled intervention, matched comparison, component ablation, longitudinal observation, and adversarial testing — deliberate attempts to bypass restraint, poison memory, and destabilize regulation. Safety evaluation includes conduct toward users in distress, evaluated across demographic groups with representative data.

05 · Risk

Risk management runs on a continuing cycle — identification, analysis, evaluation, treatment, monitoring, and review, with records — following NIST AI RMF and ISO/IEC 23894 practice. Nine risks specific to emotionally intelligent AI anchor the register.

Incident response covers graduated containment — regulation by degree, up to and including full stop — preservation of decision evidence records for reconstruction, user notification where warranted, and post-incident correction feeding the Standard's revision process.

06 · Transparency

A trustworthy system leaves a trail. Each system maintains its documentation, its operator can explain it, and every governed decision leaves evidence.

Documentation

A capability-level system specification; data statements for emotional inputs covering provenance, characteristics, and limits; a system card summarizing use, evaluation status, and known limitations; and the records of testing and risk management.

Explainability

The operator can obtain an account of how internal state influenced a given behavior, at a level of disclosure appropriate to the audience. End users receive honest, non-technical disclosure.

Decision Evidence Records

Every governed decision produces a durable, examinable record — created when the decision is made, sufficient to establish later what was decided, under what conditions, and on what basis.

Designated Accountability

A named role answers for the system's conduct in operation.

07 · Societal Impact

Emotionally intelligent AI enters people's lives, so the Standard weighs deployments against their effects on those lives.

08 · Compliance

The Standard is written alongside the NIST AI Risk Management Framework, so organizations can adopt it inside governance language they already use. Each area maps to one of the framework's four functions, and the Standard carries the regulatory analysis with it.

Crosswalks published by NIST extend the mapping from the AI RMF to ISO/IEC 23894, the OECD AI Principles, and further U.S. instruments.

09 · Evolution

The Standard is a living document: reviewed formally at least every two years, and immediately upon events that matter.

Revision 1 (2025)

The original standard: founded on early exploratory theory, with fixed capability benchmarks and a proposed certification mark.

Revision 2 (August 2026)

Re-founded on the implemented EM-NOUS Framework — persistent computational-affective state and regulation, governed generation, memory continuity, restraint, and decision evidence. Withdraws Revision 1's physical-substrate concepts and fixed benchmarks, replaces certification with evidence-based conformity assessment, adds EU AI Act alignment, and updates every normative reference. Supersedes Revision 1 in full.