Independent Verification for Every Result | Dweve AION

Choose the audience that matches your question

The page contains three selectable readings of the same subject.

Knowledge governance built on AION-style provenance.

Merkle-certifiable parser used by AION for input handling.

Verified runtime that pairs AION certificates with execution transcripts.

Event provenance for systems: hash-chained records of what happened, not reasoning steps inside a decision.

Run an example certificate and check it with aion check, cvc5, or any LRAT-compatible verifier once the repository publishes. Publication is planned for Q4 2026, and the installation instructions arrive with it.

A certificate stays small and stays re-checkable years later.

AION is proof infrastructure for AI decisions, coming in Q4 2026. It emits a compact proof beside an AI result, and a separate checker walks that proof against the original premises and accepts or rejects it without the model, the solver session, or the network.

AION adds a checkable receipt to an AI answer. Your assistant can tell you whether the reasoning holds without asking you to understand the machinery, and the receipt can be checked again without an internet connection. AION is in Q4 2026, and its documentation comes with it.

AION opens its round in Q4 2026: proof infrastructure for AI decisions, with the repository and its documentation published together. Post hoc explanations merely ask you to trust another account from the same system. AION instead emits a proof certificate that any verifier can check offline, in linear time, without the solver.

Your files, your hardware, your numbers, and a checker that reads a stranger's proof

That is a two week evaluation with a written output rather than a demonstration. The corpus is yours, the hardware is yours, the numbers are yours to publish internally, and nothing about the exercise requires a conversation with us. The harness ships with the repository, so the run can start the day access arrives.

Correct is not AION's opinion of itself. Each file in a standard corpus carries the status it is known to have, and the harness compares AION's answer against that annotation instance by instance, so the count at the end is a score against a stated ground truth. The address space per run can be capped, which keeps one pathological file from taking a machine with it.

AION ships a benchmark harness, so the numbers you quote in a business case are the ones your own files produced on your own hardware. Point it at a corpus, and it records parse time, solve time and peak resident memory for every instance, then prints a count of how many answers were correct. The run repeats on the same corpus, so a second evaluation can be set beside the first line by line.

Even with all of that, somebody will dispute an outcome. That needs a route, not a meeting

Somewhere before that last one there was a day when the work was still an afternoon's, and that is the day a manager wants marked, because every option after it costs more. Reading the margin daily is how that day gets found.

The report has an owner rather than a queue, and every line on it implies a different piece of work with a different urgency. One is complete and needs no action from anybody. One has four days left, which makes it this week's plan rather than today's. One has a single day left, which makes it this morning's work. One has been past its date for a fortnight, which needs a name against it and a sentence explaining the gap.

Anything with a deadline in it can be watched while it runs, and AION reports the room left as a signed number rather than a pass. A response window with four days remaining. A retention period a fortnight past the date it should have closed. That is an operational reading somebody can act on this morning. The number moves every day, so the report is read as a margin rather than as an alarm.

Records answer questions after the fact. Some obligations need watching while they run

Attribution is a second mechanism rather than another column of the same record. A signature is applied over the derivation order, so who produced a result survives a change of contractor, and two identical answers from two suppliers stay distinguishable. That holds even when both were produced from the same declared inputs.

Each entry on that list names the component that changed, and the stored certificate behind it is complete, so re-checking is a run rather than a reconstruction. The alternative, which is what most organisations have, is a choice between re-examining everything and re-examining nothing, and in practice that choice is always made the same way.

A result accepted last year was accepted under last year's contract. Seven different things can move underneath it: a checker rule, a certificate schema, a policy, a compiler rule, a runtime contract, a model version, or a simulation contract. When one of them does, the question of which past conclusions rested on the part that changed has an answer, and the answer is a list.

A local record is easy to keep. Keeping it meaningful when the rules change is harder

For a data protection officer that answers a short list of questions about this path, and only this path: no processor to add to a register, no transfer to justify, no model provider in the chain, and nothing of ours that can be unavailable during your incident or change beneath you on somebody else's release schedule. The build you have is the build that runs next year. That answer reads the same whether the site is connected or air-gapped.

Two paths can cross a boundary and both are yours to configure. A cluster runs over addresses you set yourself. An external prover, if you want one in the loop, is a program you install and it runs as a local process on your own machine.

The default build of AION carries no outbound client. Checking a proof, translating a requirement, monitoring a signal and running the search all happen on the machine you started them on, inside your own boundary, on hardware you control. There is no hosted model in the chain and no route to one. The same build runs in an air-gapped room as in a connected one, because nothing on the default path needs a network to finish.

A proof about the rules is one thing. Where the work runs is the next question people ask

That is a statement about the rule set rather than a sample of past cases. Testing a thousand applications shows a thousand applications; a non-interference check covers every pair the rules allow, using self-composition, property directed reachability and k-induction, with the abstraction tightened wherever a pair shows it was too loose. When it does not hold, what comes back is not a statistic. It is two applications identical in every public field, differing in the one you asked about, with two different outcomes printed side by side. That is a finding a compliance officer can act on the same morning.

Take the one regulators and complainants ask most often, in different words: did this field change the outcome? It is answered as a property over two runs. Tie every public input together, leave the field in question free on each side, and require the visible outcome to agree. If it always agrees, the outcome did not turn on that field.

A certificate is only interesting for the sentence it lets you say. This binary corresponds to this source graph. This workload ran under these capabilities. This action had the required approval. This optimum meets this bound. This trajectory respected these invariants. This conclusion follows from these constraints. This output came from this exact model and parameter set.

A case is about one application. Some questions are about two of them at once

Both of those are findings, and both have an owner. Neither is available from a pass or a fail. The counterexample is what tells you which of the two you are holding.

The counterexample also ends arguments that otherwise run for weeks, because it is not anybody's opinion about whether the rule is too strict. It is a case, drawn in the real values the rule works over. Either the case is possible somewhere in your business, in which case the rule is right and the system is wrong, or the case cannot occur, in which case the specification is missing a constraint that everybody assumed and nobody wrote down.

When a rule does not hold, AION returns the specific situation that breaks it, written out as prose: which values, of what kind, and which conditions they violate. An engineer can reproduce that and a manager can price it, and for once the two of them are arguing about the same case.

Which leaves the sentences that can mean two things, and those get asked about

Translation is a pattern library rather than a model, so the same sentence always produces the same reading, and a review does not have to be repeated because the tool changed its mind between one quarter and the next. A second reviewer reading the same sentence arrives at the same condition, which is what makes a review transferable.

Every condition arrives with one sentence of plain English saying what it encodes, and with the quantities it works over, each named with the kind of value it holds and its own line of description. So the review can be done by the person who owns the requirement rather than only by the person who owns the notation.