Kontour Surface Snapshot ViewerGitHub
Browse the docs

Where Kontour Fits in the Trust Landscape

This page places Kontour in the wider world of trust, attestation, and provenance

standards — what problem is genuinely unsolved, where Kontour sits, and the

evidence that this position is real rather than aspirational. The claims below are

grounded in an adversarially-verified landscape review (2026-07); sources are

linked inline.

The layers

  • Hachure is the open, vendor-neutral trust format Kontour created and stewards. It defines portable records — claims, evidence, verification policies, append-only events — and a *published, versioned status function* status = f(evidence, policy, now) that any consumer recomputes.
  • The building-block tools — forage (crawl), traverse (extract), lookout (drift) — emit hachure-evidence-shaped output but depend on nothing in the trust layer; a consumer lifts their output into a bundle.
  • Surface is the integration surface — the reference implementation of Hachure that derives and validates trust state from bundles and exposes it (CLI, the Surface Console, the MCP surface). Every product speaks the format through Surface.
  • Kontour's products sit on top — evaluation and evidence producers (e.g. Survey) project through Surface; consumers read derived trust over the MCP surface. Producers own their domain; Surface owns portable trust primitives; dependencies point one way.

For the engineering-canonical version of this stack — the four layers, dependency direction, and today-vs-target state — see the Portfolio Layer Doctrine.

The gap nobody else fills

The standards world has thoroughly solved portable attestation — signing and recording *frozen statements* at a point in time. What none of them standardize is the step *after*: "what is this claim worth right now?" Every adjacent standard leaves that appraisal — the relying-party trust decision, and the *content* of any policy — to the verifier.

  • IETF SCITT (RFC 9943, Proposed Standard, June 2026) registers signed statements on append-only transparency services and returns receipts. It *requires* transparency services to maintain (and even publish) registration policies, but *the relying-party trust decision is explicitly out of scope*, and the appraisal-policy *content* is left to each operator — SCITT standardizes the registry mechanism, not a recomputable status function. (RFC 9943)
  • in-toto / DSSE / SLSA / Sigstore seal statements about artifacts; verification means checking a signature, not re-deriving a status.
  • NIST OSCAL Assessment Results is a snapshot *document* with assessor-populated status fields and no executable appraisal function.
  • C2PA recomputes only *cryptographic validity* at validation time, not a policy-driven status.
  • RATS/EAT has an "appraisal" concept but *centralizes* it in a trusted Verifier and seals the result; the appraisal policy is left unstandardized.
  • W3C VC status lists publish a revocation *bit the issuer flips* — a lookup, not an appraisal over evidence and policy.
  • General policy engines (OPA/Rego, Cedar, CEL) do recompute f(input, policy) — but they are *languages and runtimes*, not a specific, published, versioned trust-status function. Hachure could run *on* one; the engine is substrate, the standardized function is the product.

The closest *ideas* are two 2026 academic preprints (Proof-Carrying Agent Actions; KYA) that describe consumer-side recomputable policy — but both sit in the agent-authorization lane (not eval-conclusion trust), carry no calibrated confidence, and are unadopted research, not shipping formats.

Verdict from the review: the "portable, recomputable appraisal function + optional sealing" niche is *genuinely unoccupied at the format layer*. Kontour composes with this ecosystem — SCITT/DSSE/Sigstore as sealing and transport, SLSA/OSCAL/C2PA/BOMs as evidence inputs — rather than competing with it.

The wedge: trust for AI evaluations and agent outputs

The sharpest, least-contested place this matters is AI. An eval result — "the model passed the safety eval," "the agent verified the change" — is produced with rich context and then, crossing a boundary, collapses to a badge or a number. The evidence is gone, there is no expiry, the receiver can't re-check.

  • The problem is funded and named: self-reported eval claims no longer pass — in 2024 the Texas Attorney General settled with Pieces Technologies over marketing an unproven "<0.001%" hallucination rate, imposing five years of disclosure obligations (Texas AG settlement, Sept 2024). A third-party AI-assurance market is scaling with public funding — the UK's £11M AI Assurance Innovation Fund (DSIT Trusted Third-Party AI Assurance Roadmap, Sept 2025). And the compliance-audit world is shifting from point-in-time reports toward *continuous, freshness-tested* evidence — which is status = f(evidence, policy, now) in the auditor's own direction of travel.
  • Every AI-attestation *format* shipping today (CycloneDX ML-BOM, OpenSSF Model Signing, in-toto for ML, the IETF model-lifecycle-attestation draft, C2PA, machine-readable model cards) seals a statement or inventory — none carries a portable, recomputable eval *conclusion*.
  • Kontour's conclusionConfidence + comfort-zone — a calibrated probability the conclusion is correct plus an in/out-of-distribution signal, carried on the conclusion across the boundary — is *ahead of the ecosystem*; no adopted format ships it. This is the wedge to lead with.

This lane is defined vendor-neutrally in the open standard: Hachure AI Evaluation profile.

Honest positioning

  • The problem is real and funded now; the demand for *this specific format* is a forward bet — no regulation mandates recomputable trust yet (EU AI Act, NIST, ISO accept maintained documents), though every one has converged on *continuous* + *cross-boundary* evidence, the two conditions that break sealed attestations.
  • The right go-to-market is to ride under an assurance/attestation motion buyers already accept, not to sell a novel format cold.
  • "Recomputable" honestly means a consumer re-derives the status under their policy at their now over the carried evidence — not always re-running the eval, which may be infeasible. Overclaiming here would be self-refuting for a trust product.

For the plain-language version of the problem and solution, see Start Here. For the trust vocabulary, see Concepts.