A joint research project of the University of Cambridge and the SINE Foundation, funded by ARIA under its Scaling Trust programme. Term: July 2026 – December 2027. Co-PIs: Martin Kleppmann (Cambridge) and Martin Pompéry (SINE).
Source: sources/proposals/2026-06-17-aria-ccti-proposal-draft-2.md — the
“Schedule 3 / Part 1 – Services” section of the funding agreement, draft 2,
dated 2026-06-17. ⟨Inference: this is contract language, not a public
description; the Background field is still an unfilled guidance placeholder,
so the document states what is committed to, not how it is pitched.
Whether this draft was signed as-is is not recorded in the source.⟩
Thesis
“AI agents can find counterparties, negotiate terms, and execute real-world transactions without the prior trust history that humans depend on, provided that the agents’ claims about their physical entities and their reasoning are verifiable.”
The north star is to build the infrastructure for such proofs and to test the hypothesis empirically on at least one vertically integrated use case: that AI agents with this trust infrastructure reach agreements where humans, or AI without it, would not. See AI Advantage for the measurement claim.
Stated ambition: to become a centre of excellence at the intersection of information security, cyber-physical systems, and agentic AI in the UK.
Work packages
All run in parallel from project start; only WP-E is conditional.
| WP | Lead | Title | Period |
|---|---|---|---|
| WP-A | Cambridge | Attested-data layer + attestation types | Jul 2026 – Dec 2027 |
| WP-B | Cambridge | Verifiable query/rule layer (zkDB) | Jul 2026 – Dec 2027 |
| WP-C | SINE | EIC + credible matching | Jul 2026 – Dec 2027 |
| WP-D | SINE / Cambridge | Integration showcase + evaluation of AI advantage | Jul 2026 – Dec 2027 |
| WP-E | SINE | Institutional lens | Conditional, Jul 2027 – Dec 2027 |
| WP-F | SINE / Cambridge | Integration facilitation | Jul 2026 onwards |
SINE leads the two conceptually novel packages (EIC, institutional lens) and co-leads evaluation and facilitation; Cambridge owns the two infrastructure layers. For the technical stack see Cryptographic trust infrastructure; for WP-C’s primitive see Epistemic Integrity Certificate.
WP-E is the institutional-economics package: characterising one institutional equivalent (insurance, arbitration, or reputation) that emerges among agents collaborating with CCTI infrastructure. It only activates if the earlier packages succeed.
Milestones and gates
| # | Type | Content | Date |
|---|---|---|---|
| 1 | Milestone | Setup, use-case selection and specification; first integration map + convergence-method proposal | 30/9/2026 |
| 2 | Go/No-Go | Attestation-layer feasibility: ≥2 attestation types + EIC prototyped; PoC query language; end-to-end demonstrator WP-A–C | 31/12/2026 |
| 3 | Milestone | ≥3 attestation types; matching end-to-end with multi-dimensional reputation; a demonstrator exhibiting a cooperation-unlocking decision | 31/3/2027 |
| 4 | Go/No-Go | Language gate + AI-Advantage signal + WP-E commit + integration review (offered-vs-adopted ledger) | 30/6/2027 |
| 5 | Go/No-Go (only if WP-E active) | Scale-up without assuming trusted hardware; larger AI-advantage evaluation | 30/9/2027 |
| 6 | Milestone | Final publications, showcase release, CCTI continuity plan | 31/12/2027 |
Three gate designs are worth noting because they are unusually explicit about being able to fail:
- Language gate (M4) — if the query language shows no demonstrable advantage over general-purpose code across two use cases, it is downgraded to general-purpose code rather than defended.
- AI-Advantage gate (M4) — the comparison is committed to being “publishable either way”; a No-Go triggers a headline-claim pivot.
- M5 raises the security bar by removing the trusted-hardware assumption, i.e. TEEs are explicitly a starting crutch, not the target architecture.
Facilitator role (WP-F)
Beyond its own research, CCTI acts as the programme’s cyber-physical trust-layer facilitator, in two directions:
- Toward other creators in the Scaling Trust cohort — build and maintain reusable trust primitives (attested evidence, EIC), maintain a living integration map (dependencies, gaps, synergies), and run a convergence method: a cadence and process for iterating toward shared primitives with willing teams.
- Toward ARIA’s Arena — contribute open-source benchmarks for trust-primitive verification and proposed challenge designs.
The stated leitmotif: CCTI commits to the offer and to the contribution effort; adoption decisions remain with the creators and with ARIA/Arena leadership. This is instrumented by an offered-versus-adopted ledger at M4.
⟨Inference: this is a deliberate hedge against a facilitation role being graded on other teams’ behaviour. It is also the part of the project that most resembles SINE’s existing convener work in PACT — the novel element is that here the convening is contractually owed rather than self-assigned.⟩
Openness
All artefacts — publications, research outputs, code, integration patterns — are open access and open source by default, excepting proprietary data or tools contributed by industry partners.
Consortium
| Party | Role |
|---|---|
| University of Cambridge | Prime; WP-A, WP-B |
| SINE Foundation e.V. (DE) | Sub-contractor; lead WP-C and WP-E, co-lead WP-D and WP-F; cryptography, game theory, industrial supply chains |
| Ink & Switch UK Ltd. | Sub-contractor; zk-SNARKs, zkVMs, software engineering, human-agent interaction |
| Light Squares Ltd. | Sub-contractor; TEEs, cryptographic protocol engineering |
Note the structure: SINE appears as a sub-contractor to a UK prime while one of its own principals is a named Co-PI. ⟨The CCTI source lists Martin Pompéry only as “Co-PI, SINE Foundation”; his founder role at SINE comes from sine-foundation.md, not from this document.⟩
Published prior art by CCTI’s own people ⟨added 2026-08-17⟩
Two papers ingested on 2026-08-17 turn out to sit inside the project’s own personnel, and they are the clearest existing evidence of what WP-A is supposed to produce:
- Man & Kleppmann, Emission Impossible (E-Energy ‘26) — by the Cambridge Co-PI. Attested sensor data plus recursive proofs, implemented, measured, and applied to a real reporting obligation. Those are exactly two of the four evidence types in WP-A’s own table (see the CCTI stack).
- Heiss et al., Verifiable Carbon Accounting in Supply Chains (IEEE TSC, 2023) — by Jonathan Heiss, who is part of SINE within this programme (Martin Pompéry, 2026-08-17). Attested sensor data in supply chains, with an auditor certifying the method once rather than each report.
Man & Kleppmann §6 states its line of advance over Heiss et al. explicitly and thanks him for feedback, so the two sides of the consortium are already in working contact on this problem. Joint reading: Verifiable Carbon Accounting.
⟨Inference: three consequences. (1) The supply-chain “adoption pathway” of
outcome 4 is not aspirational — there is a shared, published research object
in it already. (2) Both papers assume a trusted signing meter or sensor,
which is precisely the assumption M5 commits the project to removing;
that gate now has a concrete baseline to be measured against rather than an
abstract one. (3) Heiss is not named anywhere in the proposal draft, so his
work package and deliverables are undocumented in sources/ — an open item
on his page.⟩
Open questions carried in the document itself
-
The evaluation use case(s) are not yet selected — due by M1 (30/9/2026). ⟨Lint 2026-09-07, re-checked 2026-09-14, 2026-09-21, 2026-09-28, 2026-10-05, 2026-10-12 and 2026-10-19: the wiki still records no selection. No source later than the 2026-06-17 proposal draft exists on this, so the wiki cannot tell an unmet gate from an unreported one — see open tasks.⟩
⚠ The date has passed. M1 fell due 30.09.2026 — nineteen days before the 2026-10-19 lint, thirty-three before the 2026-11-02 one — and nothing on this page or in
sources/records whether the gate was met. The question is no longer “will it be met” but “was it met, and was the wiki not told”. ⟨Lint 2026-10-12: this bullet read in the future tense and dated its last check to 2026-09-14 — four weeks after the lint had begun re-asking the question weekly inopen-tasks.md. The tracking file was current; the page a reader actually opens was a month behind it.⟩ ⟨Lint 2026-10-19: seventh ask, still nothing. Nineteen days past the gate, this is the wiki’s oldest unanswered question — on the one programme in it that SINE runs itself.⟩ -
Whether the approach unlocks cooperation in those use cases is unknown.
-
Real industry data will be attempted but is not required; anonymised or synthetic data is an acceptable fallback.
(The source repeats this “Open questions” paragraph verbatim twice, under the work-package section and again under the facilitator section — an editing artefact of the draft.)
Long-term outcomes claimed
- Continuity of CCTI via industry partnerships, follow-on funding, or a Cambridge-anchored community hub.
- Open-source trust infrastructure centred on a general-purpose zkDB.
- The EIC as a reusable primitive adoptable by other Scaling Trust creators and ARIA Arena challenges.
- Adoption pathways in supply chains, manufacturing, logistics and adjacent domains where coordination without prior trust history is expensive or blocked.
- Methodological precedent for empirical AI-Advantage measurement.
- Theoretical foundation for agent-mediated institutional dynamics (insurance, arbitration, reputation).
Why this matters strategically for SINE
⟨Inference, for discussion rather than as fact:⟩
- It is a second centre of gravity. SINE’s public identity is sustainability-data collaboration (Data Commons, PACT, iLEAP, DPP). CCTI is agentic-AI trust infrastructure with a UK institutional anchor. The connective tissue is real — both are about making cooperation possible between parties who cannot see each other’s data — but the communities, funders and venues barely overlap. The sharper framing of that split is on Coordination vs. bargaining failure: PACT/iLEAP address coordination failure, CCTI addresses bargaining failure.
- It broadens the assumed competence base from MPC/garbled circuits to ZK (SNARKs, zkVMs) and TEEs, plus agent evaluation and institutional economics.
- It is a funding-diversification data point against Hypothesenpapier H3/H6 (structural dependence on German project funding): a UK national research agency is now a funder.
- Outcome 4 names supply chains, manufacturing and logistics — the same sectors SINE already works in through PACT and iLEAP. That is the most plausible bridge between the two centres of gravity.
External signals ⟨added 2026-09-10⟩
Three self-published texts by the CTO of Upright Oy — a Helsinki impact-data company with no knowledge of this project — bear on three different work packages at once. They are not evidence for CCTI; they are an outside party reaching adjacent conclusions, which is the most useful kind of corroboration available before the project has results. ⟨Lint 2026-09-14: every figure in the bullets below is self-reported by Upright and none is independently replicated — see the evidentiary health warning before any of it leaves this wiki in a proposal or a pitch.⟩
- WP-A. Their red-teaming finds LLM agents hardened against being asked and vulnerable to being told; the attack surface is facts, not instructions, which leaves provenance as the only discriminator. The best external pitch material found so far for the attested-data layer — see Cryptographic trust infrastructure.
- WP-C. Their whole product is a claim about the integrity of a reasoning process, with no way to make it checkable by a counterparty. ⟨Inference: a candidate evaluation use case — provable-in-ZK, genuinely proprietary inputs, existing skeptical paying audience — and relevant to M1 (30/9/2026), whose use-case selection is still unrecorded in this wiki as of the 2026-10-19 lint — the milestone date having passed on 30.09.2026, nineteen days ago, without a word reaching here. See Epistemic Integrity Certificate and Coherence as a truth substitute.⟩
- M4 / WP-D. Their CoCoCo metric is a small precedent for inventing a domain measure where none exists and publishing an unflattering result — with the cautionary half attached (self-published, closed harness). See AI Advantage.
⟨Inference, against the thesis: their most recent post argues that agents currently mishandle exactly the autonomy CCTI’s premise grants them — verifiable claims constrain what an agent can assert, not what it will do with a badly-weighted objective. Worth holding against the thesis section above rather than filing only as supporting material.⟩
A second external signal: the networking community’s answer ⟨added 2026-10-26⟩
Li et al., Agentic Communication Networks (arXiv, 20.10.2026) — a white paper by eight academics across China, France, Germany, the USA and Italy — starts from CCTI’s premise and arrives somewhere else.
Its motivating condition (§7.1): “Two agents may belong to different users, enterprises, operators, or service platforms and may have no prior relationship. The network must establish sufficient trust for interaction without assuming a universal agent platform or a single identity provider.” That is the thesis at the top of this page, written by people who have never heard of this project.
Its answer is identity infrastructure plus policy enforcement plus tamper-evident logging: DIDs, PKI, zero trust, policy engines, reputation, standardised audit records, blockchain evidence preservation. Zero-knowledge proofs, MPC and attestation of an agent’s reasoning appear nowhere in the document.
⟨Inference, three consequences.⟩
- It is the clearest available picture of the counterfactual. This is what gets built if the cryptographic layer is not offered to the networking community. Useful for WP-F framing and for the AI-Advantage argument: the comparison case is not “agents without trust infrastructure” in the abstract, it is agents with credentialled identity and a log.
- It names the gap CCTI’s attested-evidence layer fills, in its own words, and does not reach for the instrument. §7.4: “If an agent delegates a task, another agent selects a service, and the network coordinates the underlying resources, accountability cannot be established solely from packet logs. ACN needs an auditable relationship among authority, decisions, actions, and outcomes.” The relevant CCTI objects are business-data attestation and recursive proofs in WP-A — not the EIC, which is an ex-ante commitment device for bargaining, not an ex-post attribution record. ⟨The distinction is Martin Pompéry’s, 2026-10-26, correcting an earlier draft of this entry.⟩ Developed on Agent identity and delegated authority.
- It is an audience, not a competitor. It proposes nothing SINE builds and builds nothing SINE proposes, and it is explicitly a bid for a future standardisation agenda (its references point at 3GPP TS 23.501 and ETSI GS ENI 005). The open question for WP-F is whether CCTI’s “offer” ledger should have a row for communities like this one, or only for the Scaling Trust cohort.
Also relevant to the foundational-games question raised here by Martin Pompéry on 2026-10-23: the ACN architecture depends throughout on heterogeneous agents converging on a shared capability vocabulary (§5.1, §5.3, §5.4) and never asks how that convergence occurs without a central registry. ⟨Inference: an unnamed naming game, and a worked setting for one. No wiki page on foundational games exists yet.⟩