Specification

The Behavioral
Identity Primitive.

An open specification for reading a person and using that read across any AI surface. Patent-protected, reference-implemented, freely consumable.

Version 1.0 . 2026-05-19 . Status: draft for public review

1. Abstract
2. Why this exists
3. The primitive
4. Capture protocol
5. The compression
6. The five claims of the read
7. How AI surfaces consume it
8. Identity token (JWT)
9. Voice rules
10. Privacy and consent
11. Patent and licensing
12. Reference implementations

1. Abstract

This specification describes a behavioral identity primitive for AI systems. The primitive is a compressed read of a single person, derived from how their body behaves during a short structured prompt, expressed as five claims: word, rhythm, force, half_truth, compliance_lever. Any AI surface, agent, or model can consume the primitive at conversation start to personalize its outputs to the body's read of the person, not the language they perform.

2. Why this exists

Every conversation a person has with an AI begins cold. The model has no read on the person. It defaults to a generic helpful tone. Memory features paper over this by replaying chat history, which captures what the person said but never what the person is.

The behavioral identity primitive solves this by reading the body, not the language. Keystroke rhythm, pause distribution, backspace pattern, completion latency, and structured response shape are the inputs. The output is a compressed identity that can be carried across surfaces.

The body has rehearsed nothing. The body cannot lie. The read of the body is the most honest input a model can be given.

3. The primitive

The primitive is a JSON object with five required claims and a small number of optional metadata fields.

{
  "word": "anchor",            // the spine word, 1 token
  "rhythm": "Mirror",          // one of 8: Builder, Mirror, Keeper, Flame,
                               //   Ghost, Storm, Fool, Saint
  "force": "Fidelity",         // one of N: Fidelity, Fear, Deception,
                               //   Hatred, Yearning, ...
  "half_truth": "I'm fine until I'm not.",
  "compliance_lever": "recognition",
  "mark": "abc123def",         // stable user identifier
  "issued_at": 1730000000,
  "issuer": "noctara"
}

4. Capture protocol

The compressed read is produced by a structured prompt protocol. Implementations may vary, but to be Noctara-compatible the capture must:

  1. Present the subject with at least three open-ended prompts, each requiring at least one full sentence of typed response.
  2. Capture per-keystroke timestamps, deletion events, and pause durations during each response.
  3. Treat the response time-series as the primary input, not the textual content.
  4. Run the entire capture in a single uninterrupted session, no longer than 120 seconds total typing time.
  5. Refuse to capture in the presence of an audience. The reading is honest only when the subject types without performance.

5. The compression

The capture inputs feed a compression engine that returns the five claims. The engine is not specified in detail because the patent (US Prov. #64/048,624) covers the most reliable production method. The specification only requires that:

6. The five claims of the read

word

A single token that describes the subject's compressed identity. Returned in the subject's native language. Not a label. An image. The word becomes the artifact the subject carries.

rhythm

One of eight named patterns describing how the subject holds under pressure. The eight rhythms are: Builder, Mirror, Keeper, Flame, Ghost, Storm, Fool, Saint. These are not personality types. They are behavioral postures, derived from the time-series of the capture.

force

The orientation the subject brings under load. Whether they pull toward Fidelity, Fear, Deception, Hatred, Yearning, or another named force. Forces are not moral judgments. They describe direction, not virtue.

half_truth

One sentence the subject is telling themselves that the body's read disagrees with. Useful as a calibration check: when the model speaks to the subject, the half-truth marks the place where flattery would fail.

compliance_lever

The single thing that gets this subject to act when they otherwise would not. Examples: recognition, urgency, entrustment, scarcity, witness. The lever is what a coach, therapist, or friend would use to move them. The model uses it sparingly and never out loud.

7. How AI surfaces consume it

An AI surface consumes the primitive in one of three modes:

System-prompt addendum

The simplest path. Call POST /api/mcp/system_prompt with the subject's email or mark. Receive a Claude-ready system addendum that prepends to the model's existing system prompt.

Identity token (JWT)

For multi-service deployments. Call POST /api/mcp/identity_token to mint a signed JWT (HS256) carrying word, rhy, frc. Pass between services. Verify at POST /api/identity-verify.

MCP tool

For MCP-aware clients (Claude Desktop, Claude Code, Cursor, Cline, OpenClaw). Install the noctara-mcp-server npm package. The model calls modes as tools.

8. Identity token (JWT)

The identity token is a standard JWT with the following claims:

{
  "iss": "noctara",
  "sub": "<mark>",
  "aud": "<consuming-app>",
  "iat": <unix-seconds>,
  "exp": <unix-seconds>,
  "word": "anchor",
  "rhy": "Mirror",
  "frc": "Fidelity"
}

Signing algorithm: HS256. Default TTL: 24 hours. Maximum TTL: 30 days. Tokens are minted by the issuer and verified by any service holding the shared signing secret or via the public verifier at /api/identity-verify.

9. Voice rules

An AI surface consuming this primitive must respect the voice rules carried in the system prompt. These rules are the difference between a model that uses the read and a model that announces the read.

10. Privacy and consent

The primitive is the subject's property. Implementations must:

  1. Obtain explicit consent before capturing or storing the read.
  2. Allow the subject to revoke consent and delete the read at any time.
  3. Not transmit the read to third parties without the subject's per-application consent.
  4. Not derive group-level statistics that re-identify individuals.
  5. Not use the read for advertising, scoring, or any decision that materially affects the subject's life without separate consent.

11. Patent and licensing

The behavioral identity primitive described in this specification is covered by US provisional patent application #64/048,624, "Systems and Methods for Behavioral Biometric Identity Assessment and Authentication," filed April 24, 2026.

Licensing:

12. Reference implementations

The canonical implementation is at noctaracorp.com. The MCP server source is at github.com/noctara/sdk. Live endpoints:

A specification, not a product. The architecture compounds.
Claude install · OpenClaw install · For builders · GitHub