Behavioral Provenance Record (BPR) — working draft v0.4

Preprint: https://zenodo.org/records/20710512
Repository: https://github.com/RudsonCarvalho/bpr

A vendor-neutral, machine-checkable traceability contract for AI-assisted software development. Not a matrix format and not tied to any tool, notation, or vendor. The atomic unit is the provenance record; a traceability graph is just a set of records, and a classic RTM is one projection of that graph.

The one rule

Provenance-or-reject. Every derived node MUST cite the upstream node(s) it derives from and MUST carry a justification. Nodes that resolve to nothing upstream are rejected. Human-authored roots (origin: authored) are exempt — this is what keeps rationale and decisions honest instead of forcing fake links.

Conformance levels

Level	Name	Enforced here	Checks
L0	Structural	✅	required fields, unique ids, valid node types, derived↔derived_from consistency, all references resolve
L1	Coverage	✅	every scenario is verified by ≥1 test; every test verifies ≥1 scenario
L2	Bidirectional closure	✅	every derived node reaches an authored root upstream; every approved need reaches a test downstream (gated on status)
L3a	Claim citation	✅	every behavioral/operational claim of a derived node cites a valid upstream claim (no invented behavior)
L3b	Attested support	✅	every citation carries an attestation with verdict supported (unchecked/unsupported fail)
L3c	Policy-bound support	✅	every attesting checker is authorized by the bound profile (--authorized-checkers); requires a profile

L1 is the viable MVP. L2 is deterministic graph reachability, gated on lifecycle status so deferred/rejected/informational needs are not forced to have tests. L3 is the real differentiator and is split into three sub-levels so incremental adoption never silently dilutes its meaning.

Read levels honestly. L1/L2 deliver immediate, deterministic value today; L3 is a longer-term frontier whose strength depends on the (out-of-scope) quality of claim checkers. A conformance level is a statement about structure, not correctness: "L3b conformant" means every behavioral/operational claim is cited and attested supported — it does not mean the attestations are correct, and must not be read as proof that nothing was invented.

How L3 detects invention (without proving a negative)

You cannot prove an agent invented nothing. So L3 inverts the burden:

1. Claims become first-class, in three kinds. Each claim is behavioral (changes observable functional contract — must trace to a functional claim), operational (changes observable non-functional behavior: latency, retries, logging, metrics, security, ordering — must trace to a non-functional/quality claim or be attested policy-compliant), or implementation (strictly internal, no observable footprint). The implementation exemption is default-deny: it holds only if non-observability is attested (kind_attestation: supported); otherwise the claim is treated as behavioral. This closes the laundering hole and gives non-functional-but-observable changes (cache→latency, retry→duplicate events, log→data exposure) a home instead of a semantic dispute.
2. The obligation is graded so incremental adoption never dilutes "L3": L3a every behavioral/operational claim cites a valid upstream claim; L3b every citation is attested supported (unchecked/unsupported fail); L3c every checker is authorized by the bound profile.
3. The checker is substitutable (human, NLI model, LLM-judge, proof). The standard mandates that the verdict exists, is recorded, and at L3c comes from an authorized source — never how it is produced.

Propagation rule (discovered while building): L3 only applies through the sub-chain where claims are declared on both ends. Where an upstream node is claim-less, the node's behavioral claims are origin-level at claim granularity and need no citation — the node-level trace (L0) still holds. This makes L3 incrementally adoptable, claim-bearing subchain by claim-bearing subchain.

What is fixed vs substitutable

• Fixed (the standard): the record schema, the node/edge vocabulary, the invariants, the conformance levels, and the serialization.
• Substitutable (not the standard): the id format (must only be stable and lineage-preserving), the authoring tool, the scenario notation, and the validator engine itself. A tool using a different stack must be able to reach L1 conformance without depending on this reference implementation.

Files

• schema/provenance-record.schema.json — the atomic record (JSON Schema 2020-12)
• examples/profile-retrieval.records.json — neutral end-to-end example (need → example → scenario → contract → code → test), plain text, no proprietary notation
• examples/profile-retrieval.broken.json — same example with 3 planted L0/L1 violations
• examples/profile-retrieval.l3.records.json — same chain with first-class claims, L3-conformant
• examples/profile-retrieval.l3-broken.json — code invents a behavior (uncited) and cites one claim a checker attested as unsupported
• validator/validate.py — L0/L1/L2/L3a/L3b/L3c conformance checker (the substitutable engine)

Run

pip install jsonschema            # optional: enables full schema validation
python validator/validate.py examples/profile-retrieval.records.json --level L2          # -> CONFORMANT
python validator/validate.py examples/profile-retrieval.l3.records.json --level L3b       # -> CONFORMANT
python validator/validate.py examples/profile-retrieval.l3.records.json --level L3c \
       --authorized-checkers "human:reviewer"                                            # -> CONFORMANT
python validator/validate.py examples/profile-retrieval.l3-broken.json --level L3b        # -> NON-CONFORMANT, exit 1

Neutrality check (do this before publishing)

Re-express the same example with a different id scheme and a different source notation, and confirm the schema and validator still hold. Two independent encodings converging on the same conformance verdict is the embryo of "two independent implementations converge" — the line between a standard and a private API.