DCN Computation Attestation Profile 0.1
This profile defines how DCN represents evidence about the computation that produced an AI result.
Core rule
Evidence MUST state exactly what it establishes. Transport provenance, runtime identity, model identity, and computation correctness are separate claims and MUST NOT be collapsed into one verified boolean.
ComputationAttestation
A ComputationAttestation records:
- attestation identifier and profile;
- subject execution;
- claim type;
- model identity commitment;
- input and output commitments;
- optional runtime binding;
- proof system and proof reference;
- verifier and verification result;
- numeric semantics and privacy properties;
- limitations and production time.
Baseline claim types are transport_authenticity, runtime_identity, model_identity, computation_correctness, and composite.
Baseline profiles are dcn.ca.transport.v1, dcn.ca.runtime.v1, dcn.ca.model.v1, dcn.ca.inference.v1, and dcn.ca.composite.v1.
Model identity
A model identity SHOULD bind the execution-relevant model family/variant, weight commitment, architecture commitment, quantization or numeric representation, tokenizer, inference implementation, sampling configuration, and adapters where relevant. A model name alone is not cryptographic identity.
Numeric semantics
A computation proof MUST state the arithmetic actually proven. Quantized, fixed-point, approximated, or transformed execution MUST NOT be represented as proof of a different higher-precision computation unless an explicit equivalence argument exists.
Input and output binding
A computation-correctness claim MUST bind the exact input and output covered by the proof. For autoregressive inference this SHOULD include prompt/context commitment, sampling configuration commitment, and output-token commitment. Randomness handling MUST be declared.
Verification result
Verification results are typed as valid, invalid, unsupported, or indeterminate. unsupported and indeterminate MUST NOT be promoted to valid through policy scoring.
Composition
Attestations compose as an Evidence Graph. For example an execution may carry a TLS transcript proving provider-visible request/response, a TEE quote proving runtime identity, a model commitment, and an inference proof binding input/model/output. The composition is no stronger than the statements actually established by its components.
Attestable Gemma profile
The reference implementation defines attestable.gemma.v1 for Gemma inference instrumented by Attestable. It maps to dcn.ca.inference.v1 only when returned evidence cryptographically binds the declared Gemma model identity, input commitment, and output commitment to the proven inference relation.
The adapter MUST fail closed when a required binding is absent or proof verification is invalid/unsupported. Vendor-specific proof payloads are preserved by hash/reference; DCN normalizes the claim without rewriting what the proof establishes.
Attestable Atlas deterministic profile
attestable.atlas-deterministic.v1 is the provider profile for computation proofs of deterministic Capability Atlas models (DCN #82, #88). These are published clinical algorithms and standardised logistic screens. An attestation uses the baseline profile dcn.ca.inference.v1 with claim type computation_correctness, and carries attestable.atlas-deterministic.v1 as proof.provider_profile.
It maps to dcn.ca.inference.v1 only when the referenced receipt verifies against a pinned guest image id and the verified journal binds every one of the following to the attestation:
- the model commitment (computed inside the proof from the model artifact bytes)
- the relation commitment, carried as
architecture_commitment, which binds the model to its numeric profile and evaluator core - the numeric semantics (
dcn.atlas-numeric.ieee754-binary64+libm-0.2.16.v1) - the salted input commitment
- the commitment to the exact output
The proof system and verifier identity (risc0-zkvm, version 3.0.5, risc0-image:<id>) are part of the descriptor.
The verifier MUST:
- return
unsupportedfor an unknown profile, proof system, verifier version or guest image - return
indeterminatewhen the receipt is unavailable - return
invalidwhen the receipt hash, receipt verification or any journal binding fails
It MUST NOT promote unsupported or indeterminate to valid.
The profile attests computation correctness only. It establishes neither clinical validity nor authority. It does not claim zero-knowledge until that property is established for the backend (#89).
Capability requirements
A capability requirement MAY require computation assurance independently of ordinary evidence classes. Admissibility filtering MUST occur before cost, latency, or quality optimization.
An expert binding records assurance profiles promised by the provider and the expected model commitment. The execution result links the evidence actually produced. A promise of attestability is not evidence.
Actum
Actum SHOULD anchor commitments to computation attestations, verification results, model identities, proof artifacts, and causal execution lineage when durable third-party verification is required. Raw prompts, outputs, weights, and proof blobs need not be on-chain; content-addressed references and commitments are sufficient.
Actum records finalized evidence claims. It does not replace the proof-system verifier and does not turn proof validity into semantic truth.
Actum Compute
Actum Compute MAY expose attestation as a market dimension, including assurance profile, proof system, expected model commitment, proving latency, and verification cost. A provider that cannot satisfy a mandatory assurance profile is inadmissible regardless of price.
Compiler Society
Attestability is a compilation target. A compiled artifact MAY carry model commitment, proof-compatible execution representation, verifier metadata, evaluation claims, and attestation profile. Such an artifact is self-verifying cognitive capital when its execution can produce independently checkable evidence without trusting the host. Proof capability is not qualification: qualified reuse still requires Kline's exact-artifact Qualification Evidence Reference (ADR-0003), and neither creates authority.
Security requirements
Implementations MUST defend against proof substitution, model/input/output substitution, replay, verifier-key substitution, silent numeric-semantics drift, partial proofs presented as whole-inference proofs, and composite evidence that hides a failed component.
A conformant dcn.ca.inference.v1 implementation MUST include negative tests for wrong model commitment, wrong input commitment, wrong output commitment, modified proof, unsupported verifier, numeric-semantics mismatch, and replayed execution binding.
Verifiable cognition is claim-specific. DCN composes evidence; it never turns provenance, attestation, or cryptographic validity into semantic truth by implication.