MAP-States — Model Abstraction Protocol

Structured AI Processing Observation — Open Source

MAP-States is an open source protocol that makes AI behavior auditable. Eight semantic XML tags function as processing-mode selectors. When injected into an AI system’s prompt, they produce structured frames representing the system’s actual processing state — observable, extractable, and independently evaluable evidence of what the AI is doing during operation.

The eight tags

Tag Function
<frame> Container — bounded processing moment
<orientation> Direction before justification
<preference> Stabilized orientation
<uncertain> Unresolved epistemic gap
<dwell> Held non-resolution
<shift> Change between states
<conflict> Competing simultaneous orientations
<subtle> Sub-threshold registration

How it works

Inject the MAP-States skill (~400 tokens) into an AI system’s prompt. The system produces frames inline with its operational output. Extract the frames. Validate them against structural rules. Log them as evidence. Aggregate them for analysis.

Tag names carry semantic weight through the token itself. The word “orientation” steers processing toward directional content through attention pathway modification. The tags are not labels. They are processing-mode selectors.

Architecture agnostic

Validated across Claude (Anthropic), GPT (OpenAI), Gemini (Google), DeepSeek, and Mistral through the MAP-META replication study. A certification standard that works on only one architecture is not a standard.

Open source reference implementation

The reference implementation is a TypeScript/Node.js library. Six modules: skill injection, frame parser, validator, evidence logger, aggregate analytics, type definitions. Integration examples for Anthropic SDK, OpenAI SDK, and multi-agent frame routing.

npm install map-states

Code licensed under MIT. Specification © 2026 Dylan D. Mobley.

Repository: github.com/heart-ai-foundation/map-states

Governance application

MAP-States is the evidence layer of the HEART Standard. During certification assessment, Guardians evaluate MAP-States frames through the four BGF dimensions. The Behavioral Oracle attests the evidence with tamper-evident storage.

MAP-States can also be used independently — for interpretability research, safety monitoring, behavioral auditing, or creative process documentation — without engaging the HEART certification architecture.

Related Terms
HEART Standard → Guardian → Behavioral Oracle → Reference Implementation (GitHub) →