Status · Working Draft v0.6.2 · Circulated for review by domain experts · Not a published standard

Proof of Insight · An open specification

Evidence for agentic AI analyses,
built for regulators.

Proof of Insight (PoI) is a specification for a content-addressed, signed directed acyclic graph of typed derivation steps — and a single mechanical algorithm that verifies analyses produced by agentic AI systems in regulated contexts.

Version
0.6.2 (working draft)
License
CC BY 4.0 · Apache 2.0
Editor
Arclio LLC (interim)
Canonical URL
proofofinsight.org/spec/v0.6.2/
§1 · The problem

Agentic analyses are reaching regulated decisions before the evidence has caught up.

Drug-development readiness reviews, model-risk assessments, clinical decision support, environmental compliance reporting — analyses produced by agentic AI systems are now contributing to decisions that regulators must scrutinise.

These analyses characteristically combine three classes of operation: ingestion of external data, deterministic computation over that data, and non-deterministic reasoning that interprets and concludes. Existing provenance frameworks — in-toto, SLSA, Sigstore, PROV-W3C — address the first two. None admits non-deterministic reasoning as a first-class operation. None provides a verification algorithm of the precision a regulator requires to act independently of the producer over a graph that includes such operations.

The result is that regulated agentic analyses currently rely on producer-controlled audit logs, narrative descriptions, or post-hoc reconstructions. None of these is mechanically verifiable. A regulator presented with such an analysis has no protocol-level basis on which to confirm even the structural integrity of the work.

§2 · The core construction

A proof is a signed, content-addressed graph of four step types.

PoI's central position is that a single primitive — a content-addressed, signed, typed derivation step — is sufficient to express the evidence required for regulated agentic analyses, and that the properties such evidence must possess emerge as theorems over composed steps rather than as separately-designed features.

A PoI proof is a directed acyclic graph of such steps. Verification is a single algorithm parameterised by step type. The taxonomy is fixed: four step types and three edge relations. Domain specialisation is handled through a profile mechanism without modifying the base.

observe

External data, hashed and signed.

Records the ingestion of an artifact by content hash, attributed to an authorised attestor. The trust-handoff boundary with source-data integrity regimes.

compute

Deterministic functions, replayable.

Records the application of a function to predecessor outputs. Bit-identical replay where the environment admits it; tolerance equivalence where it doesn't.

reason

Non-deterministic models, as first-class evidence.

Records a model invocation with inputs, sampling, visible rationale, and an explicit replay class (R1 recorded · R2 re-executable · R3 reproducible).

attest

Signed claims about predecessor steps.

Records review, validation, qualification, prespecification, supersession, or adequacy — by a named role, about identified predecessors. Never an input.

Note on the novel claim

The promotion of reason to a first-class step type — distinct from compute, with its own predecessor relations, replay regime, visible-rationale evidence field, and verification semantics — is the principal way in which PoI extends prior provenance frameworks. Recording the model's narrative of its own analysis as content-addressed evidence, hash-bound to inputs and downstream consumers, is what makes agentic reasoning legible to mechanical inspection.

§3 · Scope

Verification, not credibility.

PoI establishes that an analysis was produced as recorded. It is silent on whether the recorded analysis was the right analysis. This distinction is constitutive of the protocol's scope, not a footnote.

A conforming proof establishes

What PoI does claim

  • Recorded inputs are bound to recorded operations through identified attestations.
  • The binding is tamper-evident under SHA-256 and the chosen signature scheme.
  • Deterministic operations are replayable under their declared regime.
  • Non-deterministic operations are recorded with stated conditions and explicit re-execution semantics.
  • Predicates over the typed-step structure — required reviews, role authorisations, prespecification ordering — hold mechanically.

A conforming proof does not establish

What PoI does not claim

  • That recorded inputs are accurate, complete, or fit for purpose.
  • That recorded methods are appropriate to the question being asked.
  • That recorded conclusions are correct, calibrated, or defensible.
  • That a reviewer applying domain judgment would reach the same conclusion.
  • That the analysis is sufficient to support any specific regulatory action.

The protocol answers a process-fidelity question and is silent on the analytical-correctness question. Both matter for regulated decisions, but they are addressed by different mechanisms: PoI by mechanical verification, analytical correctness by qualified human review under domain-specific standards. A reproducibly wrong analysis remains wrong; the protocol's value to a regulator is that it makes the analysis legible and mechanically inspectable, so that human judgment can focus on the questions that require it.

§4 · Conformance

Five levels. Each adds constraints, not features.

Conformance levels are defined as constraints over which step types must appear, which replay classes are admissible for outputs, and which roles must sign which claims. The predicate set grows monotonically with level. A producer's claimed level is recorded on the signed manifest and checked against the proof DAG.

L1 Provenance

observe and compute only. All steps signed, replay regime declared. Sufficient for low-risk computational analyses with no reasoning component.

L2 Identity

L1 plus bound organisational or individual identity for every attestor and a recognised timestamp authority. Sufficient for regulated computational analyses without reasoning.

L3 Reasoning

L2 plus reason steps admitted, at replay class R2 (re-executable) or higher for any step ancestor to an output. Model identifiers must resolve.

L4A Independently Attested

L3 plus independent qualified review of every reasoning output, plus prespecification for confirmatory analyses. The level for regulated agentic workflows on closed-weight hosted models with strong governance.

L4R Reproducible

L4A plus replay class R3 — bit-identical reproduction against pinned weights — for any reasoning ancestor of a profile-designated high-stakes output. The level for open-weight, controlled-runtime deployments.

§5 · Relationship to prior work

PoI extends an existing attestation lineage. It does not replace it.

The signature, content-addressing, and timestamping primitives PoI relies on are intended to be instantiated through existing ecosystems via informative profiles. The contribution beyond these frameworks is the typed-step taxonomy admitting non-deterministic reasoning, the typed edge relations, the visible-rationale evidence field, and a single verification algorithm that extends replay semantics to non-deterministic operations.

in-toto · SLSA

The closest cryptographic substrate. PoI instantiates its signing and attestation primitives via the Sigstore/Rekor profile. The observe step's provenance field is the explicit integration point with in-toto attestations.

Sigstore

Deployed signing and transparency-log infrastructure. PoI's §7.1 profile binds signature to keyless signing and timestamp to Rekor inclusion proofs.

PROV-W3C · PROV-AGENT

Provenance data models. PoI differs in being a verification protocol with a single mechanical algorithm, not a data model. PoI proofs can be lifted to PROV representations for interoperability.

Certificate Transparency (RFC 6962)

Closest precedent for a protocol whose verification algorithm is mechanically tractable and whose claims are precisely scoped. PoI borrows CT's discipline of stating exactly what the protocol does and does not assert.

§6 · Editorial process

Open review, with the intent to move beyond a single editor.

Status

v0.6.2 is a working draft circulated for review by domain experts in cryptographic provenance, regulated artificial intelligence, and standards development. No production deployments exist at any version. The protocol is pre-1.0.

Editorial control

Editorial control rests with Arclio LLC pending establishment of an independent editorial body. Arclio is the current steward, not the owner. The CC BY 4.0 licence and the §0.3 patent non-assertion covenant are structured to make this transition possible without re-licensing.

How to engage

Reviewers are invited to submit comments, ambiguities, and proposed changes through the public issue tracker referenced at the canonical spec URL. Threat model and profile-specific feedback is especially welcome — §6 of the spec and the §7.4 / §7.5 profile sketches are explicitly under-developed pending stabilisation of the core.

Versioning

Document and protocol version are intentionally the same. v0.6.2 supersedes v0.6.1 with backward-compatible additions. v0.6.1 L4 claims map cleanly to v0.6.2 L4R; no producer compliant with v0.6.1 is rendered non-compliant.