Most AI events end with a panel agreeing that trust matters. Ours ended with a room watching a year-old decision replay, byte for byte, on a laptop nobody had prepared. That was the brief for the evening we hosted in London with Google DeepMind and Microsoft: not a conversation about trustworthy AI, but a working definition of what an AI system must prove before an institution lets it execute a decision, not merely suggest one.
The distinction carried the night. A system that advises can be wrong in interesting ways and survive it. A system that executes, that records the grade, clears the payment, closes the case, has to be right in a documented way, and it has to stay that way on every release, under every audit, for years.
What "trusted to execute" must prove
The evening's working definition came out of our research line with Google DeepMind, which asks an unfashionable question: how far can a judgement pipeline be verified with no probabilistic step in the loop? The answer, formalised over the past year, is a set of properties a system must demonstrate before a decision can be delegated to it. Identity of output: the same input returns the same record, every run, on demand. Separability: the reasoning that finds evidence is fenced from the rules that judge it. Accountability: every verdict carries the rule, the quote and the model version that produced it. And gateability: no release ships unless the entire recorded past replays identically beneath it.
None of these is a model property. All of them are system properties, which is the evening's real thesis: trust in execution is engineered around the model, not extracted from it.
Advice versus execution
A system you can advise with
- Fluent most of the time, wrong in interesting ways
- Answers drift with the batch, the prompt, the release
- Trust is a vibe: demos well, audits poorly
- A person re-checks everything it touches
A system you can execute with
- Same input, same record, proven on replay
- Reasoning fenced from the rules that judge
- Every decision carries rule, quote and model version
- A release ships only if the recorded past replays clean
The demonstration, not the deck
Microsoft's part of the evening answered the obvious follow-up: where does a system like this run when the institution is a ministry or an awarding body? On sovereign cloud regions or the institution's own floor, with the boundary drawn by the customer and the pipeline identical inside it. The live demonstration ran the full loop on stage: a decision recorded last year, replayed in front of the room, landing identically while the release gate re-ran the calibration corpus behind it.
The verification record, as shown on stage
Recorded judgements in the replay corpus, each re-runnable on demand
1.9m
Divergences tolerated per release: one flip and the release is blocked
0
Production releases gated this way since 3.0
14
every one byte-clean
Of decisions carrying rule, quote and model version
100%
Anatomy of a decision you can execute on
- 01 · Evidencequoted, pinned to source
- 02 · Rulewritten, versioned, applied as a check
- 03 · Verdictrecorded with its reasoning
- 04 · Replayidentical, whoever asks, whenever
Why these two rooms, together
The pairing was deliberate. Google DeepMind brings the verification research: the formal account of what a judgement pipeline must satisfy before its output can be treated as fact rather than suggestion. Microsoft brings the sovereign substrate: the in-region and on-premise environments where institutions may lawfully run such a system over their most sensitive records. Sokros brings the layer where the two meet in production, marking real qualifications for real learners today.
The field has spent three years asking whether models can reason. The institutions that matter ask a harder question: can the system prove what it did?
What we publish next
The joint programme continues with the verification methods paper, co-authored with the DeepMind team, and a reference architecture for deterministic execution on sovereign infrastructure with Microsoft. Both land this year. The thesis they document is already running: determinism is not the ceiling on what AI may do in regulated work. It is the floor under what AI may decide.