Scientific Admissibility Evidence Records for Verifiable Research Provenance

Introduction Research governance systems increasingly exchange machine-readable claims, executable analyses, provenance records, and integrity metadata. Interoperability becomes difficult when a single boolean such as "valid" is used to collapse distinct questions: whether a document is structurally well formed, whether its bytes are intact, whether its declared scientific procedure satisfies registered admissibility gates, whether a result has practical value, whether a governance consequence is authorized, and whether a proposition is true. This document specifies a record model that keeps those layers separate. A conforming implementation can validate record structure and integrity and can evaluate explicitly declared cross-field invariants. It MUST NOT infer scientific truth merely from successful serialization, signature verification, or schema validation.

Conventions and Terminology 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 when, and only when, they appear in all capitals, as shown here .

Evidence Record: A JSON object containing a claim, procedure, evaluation, artifacts, actors, lifecycle state, and ordered events.
Admissibility: A decision produced by declared scientific gates. It is distinct from valuation, governance, and truth.
Artifact: A referenced file or byte sequence, such as analysis code, executable notebook, data product, or result object.
Event Core: The event fields covered by canonicalization and digest computation, excluding the integrity object itself.

Design Requirements A conforming profile SHOULD satisfy the following requirements:

Structural validation and scientific consistency checks are separate operations.
Cryptographic integrity proves consistency of recorded bytes and event history, not scientific correctness.
Failed or underpowered designs remain representable without being converted into exclusions.
Automated interpretation MAY assist review but MUST NOT silently become the authoritative research record.
Normative consequences MUST be derived from explicit state transitions and registered predicates.

Record Model An evidence record is encoded as a JSON object as defined by . Profiles of this specification SHOULD define a machine-readable structural schema. The top-level record contains the following conceptual members:

schema_version: record-format version.
record_id: stable identifier for this record.
claim: bounded observable claim and lineage.
registered_model: model family, hypotheses, parameters, assumptions, and prediction fingerprint.
parameter_regions: search, power, and refutation regions.
registered_procedure: preregistered decision procedure and calibration declarations.
admissibility_evaluation: binding gate results and evidence references.
valuation: optional non-authoritative valuation information.
governance: optional consequence predicates and verification requirements.
current_state: state derived from the event sequence.
actors, artifacts, and events: provenance and lifecycle information.

Scientific Consistency Invariants Structural validity alone is insufficient to claim profile conformance. A profile MAY define additional deterministic cross-field checks. Where the corresponding fields are present, an implementation conforming to the core profile defined here MUST enforce the following invariants:

An admissibility evaluation refers to the exact registered procedure being evaluated.
A record MUST NOT report admissibility as passing while a binding calibration or power gate is failing.
An underpowered null result MUST NOT activate a scientific exclusion.
Diagnostics that are declared non-authoritative MUST NOT become autonomous admissibility gates.
A predictive revision MUST create a new claim version; a declared prediction-equivalent reparameterization SHOULD preserve claim identity when the profile's equivalence test passes.
A terminal retirement state requiring independent verification MUST NOT be reached before the corresponding verification event.

Canonicalization and Integrity When an event digest is used, the event core MUST be serialized using the JSON Canonicalization Scheme (JCS) defined by before hashing. The event core consists of the sequence number, event identifier, event type, occurrence time, actor identifier, previous-event digest, and event payload. The integrity object is excluded from its own digest to avoid circularity. Profiles using SHA-256 SHOULD represent digests using an algorithm identifier together with the lowercase hexadecimal digest. Profiles using Ed25519 signatures SHOULD follow . Implementations MUST reject a record when a required digest or signature fails verification.

Append-Only Event Semantics Events are ordered by a monotonically increasing sequence number. Except for the initial event, each event MUST reference the digest of the immediately preceding event. A conforming validator MUST reject a discontinuity, reordered event, or digest mismatch when the selected profile requires an append-only chain. Cryptographic linkage does not establish that the recorded event was scientifically justified or factually honest. It establishes that the protected event sequence has not changed without detection under the assumed cryptographic model.

Example The example is intentionally incomplete and is not itself a complete profile-conforming registry record. It illustrates the separation of claim identity, procedure identity, evaluation, and lifecycle state.

Interoperability Considerations Producers SHOULD expose the exact profile identifier and version used to validate a record. Consumers MUST NOT silently reinterpret an unknown profile as a known profile. Extension fields SHOULD be namespace-qualified to reduce collisions. External standards references SHOULD identify the exact specification and revision on which an implementation depends. A standards reference is evidence of the applicable interoperability contract; it is not evidence that a scientific claim is true.

Security Considerations Implementations need to protect private signing keys, validate signatures before relying on integrity metadata, reject malformed or ambiguous JSON inputs, and avoid treating untrusted paths or artifact identifiers as local filesystem authorization. Signature verification does not establish correctness of the underlying experiment, honesty of the signer, validity of a statistical model, or authorization of a governance consequence. Consumers MUST maintain those trust decisions separately. Canonicalization and hashing are sensitive to exact byte and character semantics. Implementations using JCS MUST enforce its constraints, including rejection of input that cannot be represented by the JCS data model.

Privacy Considerations Evidence records can contain author identifiers, timestamps, artifact locations, experimental metadata, or other information that becomes sensitive when correlated. Profiles SHOULD support data minimization and SHOULD avoid embedding secrets, credentials, private keys, or unnecessary personal information.

IANA Considerations This document has no IANA actions.

Implementation Status An experimental implementation corresponding to version 0.3 of the record model includes a JSON Schema structural contract, deterministic semantic conformance checks, RFC 8785 canonicalization tests, SHA-256 artifact verification, Ed25519 event-signature verification, passing and intentionally failing scientific fixtures, and append-only lifecycle checks. These implementation artifacts are informative and are not incorporated by reference as normative IETF material.

Open Issues

Whether this work belongs in the IETF, an IRTF research group, or another standards/research venue.
Whether a future revision should define a media type after interoperability experience exists.
Whether cryptographic agility should be mandatory in a future profile.
How profile registries should represent scientific-domain extensions without implying universal scientific semantics.