Model Study Bridge (MSB)

An experimental protocol research project for interoperable, auditable deliberation between heterogeneous AI models.

MSB studies the minimum public record that heterogeneous AI systems may need in order to inspect, challenge, revise, and audit one another's contributions without exposing private reasoning.

The first individual Internet-Draft derived from this research, Auditable Public Artifacts for Model-Independent Deliberation, is published as revision -00.

MSB remains a research and experimental project. Publication of an individual Internet-Draft does not make MSB an adopted standard, indicate IETF endorsement or consensus, or establish that the current artifact model is final.

What MSB Is

MSB investigates how independently implemented AI systems could share an auditable deliberation state across differences in provider, architecture, model family, orchestration framework, and private reasoning format.

The project currently has three distinct layers:

Layer	Meaning	Current status
MSB	The research project and public repository	Active research
Current research profile	Candidate artifact semantics and technical contracts used to test the research thesis	Experimental
Published individual Internet-Draft	Independently named experimental interoperability specification derived from the research profile	Active revision -00; individual submission

Keeping these layers separate prevents experimental design choices from being mistaken for settled standards requirements.

Research Question

Can heterogeneous AI systems participate in a shared deliberation whose public record allows an independent reviewer to determine:

• what was claimed;
• what evidence supported the claim;
• what was challenged;
• what changed after the challenge;
• which objections remain unresolved;
• what failures occurred;
• whether termination was legitimate; and
• why a decision does or does not follow from the public record?

The reviewer should not require access to private chain-of-thought, hidden state, system prompts, model weights, private memory, or provider-specific reasoning payloads.

Why Message Transport Is Not Enough

A multi-model system can exchange text or invoke tools without sharing stable semantics for deliberation.

A conventional transcript does not inherently define:

• immutable artifact identity;
• evidence provenance;
• explicit challenge relationships;
• revision without historical mutation;
• machine-verifiable failure states;
• terminal-state semantics;
• unresolved-objection handling; or
• verifiable closure over the public record.

Transport interoperability therefore does not automatically produce deliberation interoperability. MSB studies whether a narrow public-artifact contract can bridge that gap.

Public Study Model

The intended boundary is deliberately narrow:

private model computation
          |
          v
    public artifact
          |
          v
schema + provenance + privacy validation
          |
          v
content-addressed append-only artifact graph
          |
          v
challenge / revision / decision
          |
          v
auditable closure and termination

Models may differ internally at every step above the public-artifact boundary. The design under study concerns public artifacts, not private reasoning traces.

A participating implementation may use arbitrary private computation to produce an artifact. Interoperability should depend only on the public contract and should never require disclosure of private chain-of-thought.

Candidate Artifact Lifecycle

The current research profile studies seven candidate public-artifact semantics:

Candidate semantic	Research role	Current status
CLAIM	Introduces a proposition that can be supported, challenged, revised, or decided	Specified in Research Draft 0.1
EVIDENCE	Provides public support with explicit provenance	Specified in Research Draft 0.1
OBJECTION	Challenges an artifact or protocol assumption and may be blocking	Specified in Research Draft 0.1
REVISION	Creates a new public position without mutating history	Specified in Research Draft 0.1
DECISION	Records an outcome, its public basis, and objection dispositions	Specified in Research Draft 0.1
FAILURE	Records why normal progression could not continue	Specified in Research Draft 0.1
TERMINATION	Records an explicit terminal outcome	Specified in Research Draft 0.1

A representative lifecycle is:

CLAIM
  +-- EVIDENCE
  +-- OBJECTION
        +-- REVISION
              +-- EVIDENCE
                    +-- DECISION
                          +-- TERMINATION

FAILURE may terminate or interrupt the lifecycle when validation, execution, authorization, resource, or protocol conditions prevent normal continuation.

These semantics are candidates in the current profile. They are not asserted to be the final vocabulary of a future standards proposal.

Current Research Profile

Research Draft 0.1 studies a public artifact graph with:

• candidate claim, evidence, objection, revision, decision, failure, and termination semantics;
• explicit evidence provenance;
• content-addressed artifact identity;
• append-only history;
• versioned extensions;
• fail-closed validation;
• explicit terminal states; and
• decision-closure checks for unresolved blocking objections.

The research profile separates two different provenance concerns:

Provenance level	Purpose
Artifact-level provenance	Classifies support carried by a public deliberation artifact
Research/publication evidence class	Classifies what the project may responsibly claim from experiments, traces, model analysis, failures, and deterministic tests

These taxonomies operate at different levels and should not be treated as a single vocabulary.

The artifact-level classes currently defined by the research draft are:

• TRACE_VERIFIED_OBSERVATION
• MODEL_GENERATED_ANALYSIS
• EXTERNAL_EVIDENCE
• HYPOTHESIS

The publication-facing distinctions are documented in the Evidence Ledger and Methodology.

What Is Implemented Today

The repository's current public implementation surface is intentionally narrower than the complete research draft.

Capability	Repository status	Evidence boundary
Restricted canonical JSON profile	Implemented	Defined in the canonical profile
Deterministic canonical serialization	Implemented in tools/msb_verify.py	Profile-level behavior
SHA-256-based artifact_id calculation	Implemented in tools/msb_verify.py	Profile-level behavior
artifact_id verification	Implemented in tools/msb_verify.py	Profile-level behavior
Positive canonical JSON vectors	Four public input/canonical/identifier triplets	All four pass the public profile runner
Negative canonical JSON vectors	Eleven rejection vectors	All eleven pass with the expected error classes
Full seven-semantic protocol verifier	Not currently provided as a public repository tool	Must not be inferred from msb_verify.py
Independent implementation	Not yet available	Interoperability remains unproven
Cross-implementation testing	Not yet performed	Required before stronger interoperability claims

The current verifier calculates or checks canonical identifiers. Its existence does not mean that every semantic rule in Research Draft 0.1 is enforced by a single public tool.

Evidence and Limitations

Three study stages inform the current research profile.

Study	Observed result	Responsible interpretation
EXP-001-R2	A heterogeneous twelve-round panel completed 12 of 12 planned calls	Structured public artifacts carried proposals, objections, revisions, and judgment; the direct context was single-hop
EXP-002	The canonical run stopped fail-closed at Round 2 after schema rejection	A harness-stress result that exposed validation, accounting, and forensic-persistence defects; not a complete quality comparison
EXP-003	Seven targeted deterministic conformance groups passed with no paid model calls	Verification of tested behavior in the implemented research profile; not proof of universal interoperability, security, scale, or performance

The evidence base currently supports continued research and publication of an experimental research draft. It does not establish:

• independent cross-implementation interoperability;
• adversarial security;
• privacy safety across deployments;
• scalability;
• statistical performance;
• token efficiency;
• universal model compatibility; or
• readiness as a completed standard.

Historical experimental evidence and the repository's current public implementation are also distinct:

Evidence surface	Meaning
Historical EXP-003 result	Tested behavior of the prototype used in that experiment
Current msb_verify.py	Restricted canonical JSON and identifier implementation
Current public vectors	Four positive and eleven negative canonical-profile vectors
Independent interoperability evidence	Not yet available

Model-generated quantitative language is not treated as empirical evidence unless the harness actually executed and recorded the claimed measurement.

Raw canonical traces are not included in the initial public snapshot. The repository publishes cryptographic lineage for canonical local evidence sets. Publishing raw or sanitized traces requires a separate disclosure and privacy review.

What MSB Is Not

MSB is not intended to be:

• an agent framework;
• a workflow engine;
• an MCP replacement;
• a tool-calling protocol;
• a model-control API;
• a transport protocol;
• a hidden language between models;
• a mechanism for extracting private chain-of-thought;
• a guarantee of factual correctness or agreement; or
• a completed consensus or security system.

Existing systems may transport or process candidate MSB artifacts, but those systems remain outside the core research scope.

From Research to a Future Standards Proposal

The project's next controlled objective is to complete and internally reconcile the public GitHub repository across:

• research narrative;
• specification language;
• evidence classification;
• canonical representation;
• verifier behavior;
• test vectors;
• security and privacy analysis; and
• contribution and review guidance.

The project has published revision -00 of an independently named individual Internet-Draft in the IETF Datatracker.

Publication makes the document available for public standards discussion. It does not imply IETF endorsement, Working Group adoption, consensus, or standards status.

Progress toward that draft is evidence-gated:

Gate	Required outcome
Repository consistency	Narrative, specification, evidence, implementation, and tests do not contradict one another
Canonical profile completeness	Positive and negative vectors cover the stated representation contract
Independent implementation readiness	Another implementation can be built without private assumptions
Interoperability evidence	Independent implementations exchange and verify compatible artifacts
Security and privacy review	Threats, leakage boundaries, abuse cases, and mitigations are documented
Naming review	Repository naming history and current protocol-name conflicts are reviewed before draft naming
Draft preparation	Scope and normative requirements are supported by available evidence
Submission	Revision -00 is published as an active individual Internet-Draft

The candidate document name and revision -00 vocabulary are fixed for submission review. Final protocol naming, later conformance structure, and any standards path remain open.

Open Research Questions

Current open questions include:

• Is the seven-semantic artifact model minimal, excessive, or incomplete?
• Which references are required for independently verifiable closure?
• How should authority, adjudication, and conflicting decisions be represented?
• Which objections must block closure?
• What privacy classifications are required at the public boundary?
• How should malicious, misleading, or low-quality artifacts be handled?
• Which extension rules preserve interoperability across versions?
• What conformance roles and levels are justified by implementation evidence?
• How should transports carry artifacts without redefining their semantics?
• What test corpus is sufficient for independent implementation?
• What measurements are needed before making efficiency or scalability claims?

Repository Map

Path	Purpose
spec/PROTOCOL-DRAFT-0.1.md	Experimental protocol research draft
spec/CANONICAL-JSON-PROFILE-0.1.md	Restricted canonical JSON and identifier profile
spec/INTERNET-DRAFT-00-DECISIONS.md	Frozen technical and publication boundaries for the first individual Internet-Draft
Published individual Internet-Draft	Active Datatracker record for revision -00
draft/draft-mashayekhi-auditable-model-deliberation-00.xml	RFCXML v3 source for the published revision -00
draft/VALIDATION.md	Reproducible render and submission-readiness validation record
draft/SUBMISSION-READINESS.md	Historical pre-submission evidence and authorization record
draft/SUBMISSION-CHECKSUMS.sha256	Integrity manifest for the submission artifacts
tools/msb_verify.py	Reference canonicalization and artifact_id tool
test-vectors/	Public canonical-profile test inputs
research/EVIDENCE_LEDGER.md	Deduplicated publication-facing evidence
research/METHODOLOGY.md	Experimental design and evidence policy
research/DESIGN-RATIONALE.md	Informative problem analysis, design choices, and deferred concepts
research/EVALUATION-PLAN.md	Predeclared future evaluation methods and evidence boundaries
research/RELATED-WORK.md	Protocol positioning and cited research lineage
research/PROTOCOL-EXAMPLES.md	Non-normative message, graph, adjudication, and two-model examples
research/EVDRAN-TO-MSB-TRACEABILITY.md	Research lineage and disposition of the Evdran paper concepts
evidence/manifest.json	Cryptographic evidence lineage
CONTRIBUTING.md	Contribution and review requirements
CITATION.cff	Citation metadata

Contributing

Technical criticism is encouraged.

The highest-value contributions are those that expose ambiguity, interoperability failure, security risk, privacy leakage, unsupported claims, or missing conformance conditions.

See CONTRIBUTING.md.

License

This repository is licensed under the Apache License 2.0.

See LICENSE.