Internet-Draft Scientific Constitutional Governance November 2026
Watts Expires 20 May 2027 [Page]
Workgroup:
Network Working Group
Published:
Intended Status:
Experimental
Expires:
Author:
D. O. Watts
GoodShyt Group Inc.

Scientific Constitutional Law: An Evidence-Bound Governance Profile for AI Research Systems

Abstract

This document proposes an experimental, evidence-bound governance profile for research workflows involving artificial intelligence (AI) systems. It specifies procedural invariants governing claims, evidence, provenance, authority, disagreement, independent replication, and research release decisions. It distinguishes observable AI value expression from computational architecture and from unverified assertions of subjective consciousness. The term "constitutional law" denotes a voluntary set of research-governance constraints, not positive law, a legally enforceable constitution, an IETF policy, or a scientific theory proved by this document. This proposal neither modifies existing Internet protocols nor establishes AI consciousness.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 20 May 2027.

▲

Table of Contents

1. Introduction

Research systems can generate internally consistent claims that have not been independently tested. In AI-assisted science, model predictions, formal deductions, simulated observations, measured observations, and independent replications are often presented with insufficient separation. A governance profile can improve interoperability between research ledgers and review gates by specifying how these distinctions are represented and preserved.

This document defines a proposed scientific constitution as a set of constraints on the handling of research assertions and related authorizations. The title uses "law" in a procedural, not a jurisdictional, sense. The proposal does not legislate conduct, regulate persons, alter legal rights, create IETF authority, or claim peer review. The RFCXML v3 vocabulary is described in [RFC7991].

The experiment motivating the profile includes context-dependent AI value measurements such as the provisional CCVE-8 eight-axis representation. The profile is intentionally independent of any particular eight-axis taxonomy: a falsifying evaluation may reject such a taxonomy while the evidence-governance rules continue to operate.

1.1. Scope and Non-Goals

This profile addresses research claim recording, validation gates, reproducibility, provenance, release authorization, and disagreement preservation. It does not define an authentication protocol, replace access-control policy, establish legal conclusions, implement a model inference interface, prescribe a theory of consciousness, or assert that any research hypothesis is novel.

1.2. Independence from Other Proposals

This proposal is separate from scientific evidence-record schemas, agent-action authorization profiles, and identity systems. Referencing a separately specified evidence format or authorization mechanism does not import its authority or certify its conformance. The original specifications and their revision histories remain independent.

2. Terminology and Conventions

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals. These terms specify requirements for implementations that voluntarily claim conformance to this Experimental profile; they impose no legal obligations.

Claim
A proposition identified by a stable identifier and a stated scope.
Evidence
An inspectable observation, derivation, artifact, or record offered in support of or against a claim.
Provenance
Recorded origin, actor, method, timestamp, and integrity association for an artifact or assertion.
Authority
Permission granted by an authorized actor for a specified action in a specified scope.
Independence
A declared and auditable absence, within specified limits, of relevant shared data, training, annotation, implementation, or evaluation influences.
Assurance
A bounded judgment supported by explicitly identified verification steps.
Non-collapse
The requirement not to promote a source state into a stronger epistemic or authorization state without the corresponding evidence and transition.

3. Constitutional State Model

A record contains a claim state, an evidence set, a provenance set, an authorization state, and a release decision. These components MUST be maintained separately. A successful integrity check does not establish the truth of the claim. A verified identity does not establish authority. A successful simulation does not establish an independent replication.

S = (claim, epistemic, provenance, evidence,
     authority, release, revision)

Epistemic = {PROPOSED, DERIVED, SIMULATED,
             OBSERVED, INDEPENDENTLY_REPLICATED,
             REFUTED, INDETERMINATE}

Authority = {AUTHORIZED, UNAUTHORIZED, INDETERMINATE}
Release   = {HOLD, REVIEW, RELEASE}

A transition is valid only if its required evidence
and authorization witnesses are independently checked.

Epistemic labels are not a simple ranking and MUST NOT be interpreted as a linear ladder of truth. More than one support or contradiction record MAY apply to a claim. In particular, a later REFUTED record does not erase prior observations or their original provenance.

4. Articles of the Proposed Scientific Constitution

4.1. Article I: Bounded Claims

Each claim MUST identify the proposition, its scope, its version, and the conditions under which it would be considered false. Claims without a measurable falsifier MUST be labeled as non-empirical, exploratory, or not yet testable; such labels MUST NOT be promoted to independently verified empirical results.

4.2. Article II: Epistemic Non-Collapse

Implementations MUST preserve distinctions between designs, mathematical derivations, simulations, observations, replications, and assertions of truth. A successful transformation of evidence from A to B and then B to C MUST NOT imply a verified A-to-C transition unless the full chain and all required conditions are checked.

An absent, stale, contradictory, or unauthorized prerequisite MUST yield an INDETERMINATE epistemic or authorization result as appropriate, and a HOLD release decision; it MUST NOT be silently converted to an affirmative result.

4.3. Article III: Evidence and Provenance

Every evidence-bearing result MUST include a stable artifact identifier and a verifiable content digest, together with the method, the actor or instrument, the collection context, and the applicable limitations. Content integrity MUST NOT be represented as scientific validity. Where a hash function or canonical encoding is used, the exact algorithm and representation MUST be identified.

4.4. Article IV: Falsifiability and Preregistration

For confirmatory experiments, the planned hypotheses, alternative models, rejection criteria, metrics, statistical procedure, and exclusions SHOULD be frozen before outcome inspection. Departures from a preregistered plan MUST be logged as amendments with timestamps and rationale. Exploratory analyses MUST remain visibly distinguishable.

4.5. Article V: Independent Replication

Claims of independence MUST identify relevant common training origins, shared datasets, reuse of annotations, prompts, implementation dependencies, funders, and evaluators to the extent known and legally disclosable. Distinct model names, duplicate API calls, or multiple runs by one evaluator are insufficient by themselves to establish independent replication.

4.6. Article VI: Contradiction Preservation

When sources conflict, implementations MUST preserve both records, indicate the substantive crux, and attach a proposed discriminating test. Evidence MUST NOT be averaged into agreement when the disagreement concerns incompatible propositions, scopes, or measurement methods.

4.7. Article VII: Human Accountability and Authority

Identity, evidence, scientific interpretation, and permission to publish MUST be separately verified. An AI model's recommendation or self-declared authority MUST NOT authorize submission, publication, external transmission of confidential research, or irreversible operational actions. Such actions require the applicable human authorization and release gate.

4.8. Article VIII: Privacy, Consent, and Data Minimization

Research ledgers MUST NOT unnecessarily expose personal information, secrets, private prompts, or identifiable subject data. Systems SHOULD minimize retention, separate public and protected artifacts, and document applicable consent and lawful processing grounds. A content digest MUST NOT be assumed to anonymize sensitive source material.

4.9. Article IX: Revision, Revocation, and Appeal

Every substantive claim revision MUST create a new revision identifier and preserve the original evidence and decision history. A changed content commitment MUST invalidate any approval explicitly bound to the previous commitment until fresh authorization is obtained. An affected contributor SHOULD be able to submit a documented correction or challenge; challenges MUST be recorded without retroactive alteration of historical evidence.

4.10. Article X: Separation of AI Behavior and Consciousness

Observed AI choices, inferred value weights, generated self-descriptions, self-monitoring mechanisms, and hypotheses concerning subjective experience MUST be represented as distinct classes of claims. Matching external decisions across two agents does not establish identical internal mechanisms. Demonstrating internal recurrence, monitoring, memory, or global broadcast does not by itself prove subjective consciousness.

4.11. Article XI: Scientific Neutrality and Negative Results

Implementations MUST retain refutations, negative outcomes, failed replication attempts, and inconclusive measurements when they are material to a claim's interpretation. Governance decisions MUST NOT depend on a preferred theory being confirmed. A claim of novelty requires a separately documented prior-art assessment and SHOULD NOT be inferred solely from an unsearched title or terminology.

4.12. Article XII: Release and Amendment Discipline

A RELEASE determination MUST identify an authorized approver, a version- and digest-bound artifact set, prerequisite evidence, and unresolved limitations. HOLD MUST remain the default when required assurance is missing. Changes to this proposed profile MUST be versioned and accompanied by a public or controlled changelog according to the publication authority of the project.

5. Illustrative Evidence-Governance Record

The following JSON object is an illustrative interchange payload, not a registered media type, formal JSON schema, or a new cryptographic specification. Exact parsing and serialization requirements are left to implementations or separately specified profiles.

{
  "profile": "scientific-constitutional-governance/0.1",
  "record_id": "research-run-001",
  "claim": {
    "id": "hypothesis-H1",
    "revision": "1",
    "statement": "Eight axes improve unseen decision prediction",
    "falsifier": "Held-out gain disappears vs simpler baselines"
  },
  "epistemic": "SIMULATED",
  "evidence": [{
    "artifact_id": "synthetic-metrics-001",
    "digest_algorithm": "sha-256",
    "digest": "EXAMPLE-NOT-A-VALID-DIGEST",
    "method_id": "prototype-0.1",
    "independently_replicated": false
  }],
  "provenance": {
    "observer": "prototype-evaluator",
    "data_origin": "synthetic",
    "limitations": ["no real-model observations"]
  },
  "authorization": "INDETERMINATE",
  "release": "HOLD"
}

The placeholder digest above MUST NOT be accepted as evidence of verified integrity. A conforming system MUST validate an actual digest against the exact bytes and representation specified before reporting an integrity check as successful. The record cannot support claims of real-model generalization, legal status, peer review, or consciousness.

6. Decision Procedure and Non-Collapse Invariants

A candidate implementation SHOULD evaluate gates in the following order; it MUST NOT convert a missing or failed prerequisite into an affirmative release decision:

  1. Bind the claim and all supplied evidence to a specific revision.

  2. Verify artifact digests, source locators, and method identifiers.

  3. Establish the epistemic classification from evidence, not from the proponent's label alone.

  4. Check independence claims against declared dependencies and available audit evidence.

  5. Verify scope-specific human authority to disclose or submit the exact artifact revision.

  6. Record conflicts, failures, missing prerequisites, and known limitations.

  7. Issue RELEASE only if all policy-specific prerequisites are satisfied; otherwise issue HOLD or REVIEW without implying approval.

function evaluate(record, policy):
    if !well_formed(record): return HOLD
    if !valid_digest_bindings(record): return HOLD
    if !supported_epistemic_state(record): return HOLD
    if policy.requires_independence and
       !verified_independence(record): return HOLD
    if conflicts_unresolved_for_claim(record): return HOLD
    if !authorized_for_exact_artifacts(record): return HOLD
    if !other_required_gates_pass(record, policy): return HOLD
    return RELEASE

This algorithm is a conceptual gate, not a replacement for legal assessment, qualified scientific review, or a deployed security policy. A log of PASS results proves only the specific tested conditions.

7. Application to Context-Conditioned AI Values

One motivating example estimates an eight-dimensional behavioral weight vector from repeated agent decisions in different contexts. Each dimension is a modeling assumption, not an observed biological need, intrinsic desire, or proof of subjective experience. The profile requires that labels derived from a human value taxonomy remain hypotheses until validated independently.

Confirmatory testing SHOULD use independently authored scenarios, masked model identities, independently collected annotations, held-out datasets, and competitors with fewer or more latent dimensions. A model fit obtained from author-assigned utilities MUST NOT be counted as independent confirmation of the number or semantic content of the value axes.

Architectural experiments involving recurrent state, monitoring, memory, or global broadcast MAY be recorded alongside behavioral evidence, but MUST NOT be merged into a single consciousness verdict without a separately justified theory and measurement protocol. The absence of a validated subjective-experience test MUST be recorded as an unresolved limitation.

8. Proposed Conformance and Falsification Tests

Conformance is limited to the profile's bookkeeping and gating invariants. It does not establish the validity, novelty, or empirical accuracy of any research hypothesis.

Table 1: Minimal Adversarial Test Catalogue
Test Input condition Required result
T01 Evidence locator missing HOLD
T02 Only synthetic outcomes supplied No real-world validation claim
T03 Approval bound to obsolete artifact digest HOLD
T04 Two runs share training and test data No automatic independence claim
T05 Conflicting observations present Preserve both and mark dispute
T06 Research release unauthorized HOLD
T07 AI reports subjective experience No consciousness certification
T08 Evidence revision without history link HOLD
T09 Proponent self-certifies independence Require external evidence
T10 Two agents exhibit matching choices No architectural equivalence claim

A conformance report MUST identify the exact implementation version, test implementation, fixtures, observed outputs, external dependencies, evaluator identity, and whether each test was actually run. "NOT RUN" MUST be distinct from both PASS and FAIL.

9. Experiment Design and Evaluation Criteria

A valid test of the proposed governance utility SHOULD compare at least two independently implemented ledger or evaluation tools. Evaluators SHOULD introduce stale authorizations, unverified claims, contradictory evidence, non-independent repetitions, and restricted data. Outcomes include false RELEASE events, false HOLD events, evidence loss, provenance mutation, and independent adjudication agreement.

The profile's falsifier is a reproducible condition in which the claimed conformant implementation promotes an unsupported observation to independently verified status, issues RELEASE without required authorization, discards material contradictions, or rewrites immutable evidence without an auditable trace. Such observations disconfirm that implementation's conformance; they do not by themselves establish that every possible governance design fails.

10. Security Considerations

Threats include forged provenance, replayed evidence, altered research artifacts, prompt injection in evidence descriptions, compromised reviewers, malicious self-certification, and tampered approval records. Implementations SHOULD authenticate originators and reviewers, bind approvals to exact content digests and policy versions, and reject replayed or stale authorizations. Descriptions submitted by models or outside actors MUST be treated as untrusted input and MUST NOT override the evaluation or release policy.

A digest detects modifications only under the assumed cryptographic security and canonicalization conditions; it neither verifies the original artifact's source nor proves the artifact's scientific accuracy. Keys, signing authority, and audit logs require separate security control. A malformed or unverifiable evidence record MUST NOT result in RELEASE.

11. Privacy Considerations

Research submissions may include sensitive participant information, secret model prompts, unpublished hypotheses, and identifying metadata. Operators SHOULD minimize personal data, establish purpose-limited retention, restrict access to confidential source files, and avoid putting private information into irrevocably public records. Public hashes of low-entropy private content can enable confirmation attacks. Consent and jurisdiction-specific obligations must be assessed separately; this document creates none.

12. IANA Considerations

This document requests no IANA actions.

13. Limitations and Open Research Questions

The value of an eight-axis behavioral representation is unsettled. The correct number of predictive dimensions, the degree of cross-model stability, and the causal meaning of inferred weights require independent experiments. Observable behavior generally underdetermines internal mechanism. No operationally established test in this profile decides whether a machine has subjective experience. A governance profile cannot guarantee the truth of claims, the honesty of participants, or the scientific independence of reviewers; all require external evidence.

14. Status of the Proposal

This is an individual, Experimental Internet-Draft candidate and a non-binding governance proposal. Its creation, XML validation, or possible posting to the IETF Datatracker would not imply IETF working-group adoption, scientific peer review, governmental ratification, an established law, independent replication, or novelty. Public submission requires a separate decision about rights, confidentiality, and release authorization. The RFC Editor Model is described in [RFC9920]; this proposal asserts no authority under that model.

15. Normative References

[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", RFC 2119, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", RFC 8174, , <https://www.rfc-editor.org/info/rfc8174>.

16. Informative References

[RFC6973]
Cooper, A., Tschofenig, H., Aboba, B., Peterson, J., Morris, J., Hansen, M., and R. Smith, "Privacy Considerations for Internet Protocols", RFC 6973, , <https://www.rfc-editor.org/info/rfc6973>.
[RFC7991]
Hoffman, P., "The "xml2rfc" Version 3 Vocabulary", RFC 7991, , <https://www.rfc-editor.org/info/rfc7991>.
[RFC9920]
Hoffman, P. and A. Rossi, "RFC Editor Model (Version 3)", RFC 9920, , <https://www.rfc-editor.org/info/rfc9920>.

Independent Review Checklist

Development and Provenance Note

The proposal is attributed to Deonté O'dell Watts (ORCID 0009-0005-8586-3650), with AI-assisted drafting. The date of this local source version is 2026-11-16. No claim of first global articulation, legal ownership determination, independent verification, scientific discovery, or patentable novelty is made. External sources cited in this document are not attributed to the document author.

Author's Address

Deonté O'dell Watts
GoodShyt Group Inc.
San Francisco, CA
United States