Epistemic Integrity Certificate (EIC)

SINE’s own novel primitive in the CCTI project (WP-C, SINE-led). Source: sources/proposals/2026-06-17-aria-ccti-proposal-draft-2.md.

Definition as given

An approach — described in the source as novel — that allows AI agents to:

1. privately commit their beliefs and their conditions for cooperation with other agents;
2. perform secure match-making over those private beliefs (“credible matching”);
3. negotiate the further evidence needed to reach an agreement.

In the attestation taxonomy of the trust infrastructure, the EIC is the evidence type carrying epistemic claims about an agent’s reasoning, alongside sensor data (claims about the world), business data (claims about cross-organisational interaction), and recursive proofs.

Implementation commitment

The reference implementation is specified over conjugate-family beliefs (M-level deliverable for WP-C). ⟨Background: conjugate families are the Bayesian-statistics device where prior and posterior share a functional form, so a belief state is representable by a few parameters — which is what makes committing to it in a cryptographic commitment or a ZK circuit tractable at all. Not spelled out in the source; flagged as the reading that makes the design choice legible.⟩

Prototype due at M2 (31/12/2026) together with ≥2 attestation types.

Why it is the strategically loaded piece

• It is the deliverable ARIA-facing language singles out as the reusable primitive other Scaling Trust creators and Arena challenges could adopt — “directly instantiating the programme’s question of when AI-specific primitives provably expand achievable coordination outcomes”.

• It is therefore both a research output and SINE’s main lever in the facilitator role — the thing that either gets taken up by the cohort or does not, and gets counted either way in the M4 offered-versus-adopted ledger.

• ⟨Inference: conceptually it extends SINE’s existing line of work. Secure benchmarking and PACT let organisations cooperate without exposing their data; the EIC lets agents cooperate without exposing their beliefs. Same structure — reciprocity plus confidentiality — one level up the abstraction ladder. See Data Commons for the values framing.⟩

  ⟨In tension with this: Coordination vs. bargaining failure puts PACT and the EIC on opposite sides — PACT as a fix for coordination failure, the EIC as an assault on bargaining failure. Both readings can hold (the mechanism is analogous, the failure being addressed is not), but the wiki carries both rather than picking one.⟩

An independent statement of the same problem (Upright, 2025–26)

Added 2026-09-10. Upright Oy, a Helsinki impact-data company with no knowledge of CCTI, has built a production system whose entire output is a claim about the integrity of its own reasoning — impact magnitudes for which, in their words, there are “no ground-truth answers”. Their answer is to train and solve for coherence instead of correctness (Coherence as a truth substitute).

Two things follow for the EIC.

⟨Inference: the rhyme. Coherence-training is a commitment device — it binds a model’s future judgements to its past ones, which is the same structural function an EIC performs. ⟨Per Martin Pompéry, 2026-09-09.⟩ The difference is the load-bearing part: an EIC’s binding is external and checkable by a counterparty; Upright’s is internal and statistical, a training-time pull with no guarantee on any given inference and no third party who could verify it. Ostrom’s second puzzle states the consequence directly — “without monitoring, there can be no credible commitment” (Ostrom’s commons design principles). Upright has the economics of commitment without the monitoring; CCTI has the monitoring and is still looking for the economics.⟩

⟨Inference: a candidate WP-C use case. Transitive coherence over a solved system of pairwise judgements is a mathematical property — “these ~800 implied sizes are the coherent solution to these thousands of pairwise comparisons” is provable in zero knowledge without revealing a single comparison. The inputs are genuinely proprietary (patent pending, closed harness), and there is an existing, paying, skeptical audience including Big Four assurance teams. That is a sharper fit than anything named in the CCTI proposal, where the evaluation use case was still unselected as of the 2026-06-17 draft and M1 (30/9/2026) was the gate — a gate whose date passed on 30.09.2026 with no record either way, see open task 8 ⟨tense corrected by the 2026-11-02 lint⟩. Not a recommendation to approach Upright — a statement that the shape is available and worth using as the reference example.⟩

The capacity version of the same claim ⟨added 2026-11-02⟩

Capraro, Marcus & Quattrociocchi (2026) argue that AI evaluation should stop scoring human likeness and instead establish whether a system “generalizes reliably, recognizes the limits of its competence, and withholds judgment when the evidence is insufficient”. That is the EIC’s content stated as a model capacity rather than as a cryptographic object — and the source reaches it without any trust-infrastructure vocabulary, from a behavioural-science starting point.

⟨Inference: the useful part is the demand-side argument, which the wiki does not otherwise have. Every existing case for the EIC here is supply-side — SINE can build it, ARIA wants a reusable primitive, the cryptography is tractable over conjugate families. The same paper reports an experiment in which access to deliberately incorrect AI advice reduced people’s willingness to say “I don’t know”, lowered accuracy and raised confidence, with accuracy incentivized (Marcoccia, Quattrociocchi & Capraro 2026). If a counterparty’s own calibration degrades in the presence of a confident adviser, then abstention and epistemic honesty cannot be left to the receiving agent’s judgement — they have to be carried by the evidence the sending agent emits. That is a reason for the primitive that does not depend on anyone wanting to build it.⟩

⟨Inference: and the gap the source leaves open is the EIC’s market. Its recommendation to institutions is to procure on “evidence of benefit” and to test capacities beyond imitation — but the instrument it offers is the behavioural experiment (assisted vs. unaided trials), which a buyer evaluating a closed commercial system cannot run and a vendor’s own measurements cannot credibly substitute for. Three cognitive scientists have written the requirements document for a checkable claim about an AI’s reasoning, without knowing that such a primitive is being specified. See The output–process gap.⟩

Related

• Cryptographic trust infrastructure — the stack it sits on
• The output–process gap — why a process claim cannot be read off an output
• Capraro, Marcus & Quattrociocchi (2026) — the capacity version, and the demand-side argument
• AI Advantage — how its effect is to be measured
• CCTI
• Coordination vs. bargaining failure — which failure type the EIC attacks
• Coherence as a truth substitute — the same problem shape reached from outside
• Upright Project · Leena-2 (2026)