← Aria
Aria · version 3.9

Aria changelog.

Every version of Aria since the first release: what changed, the model it ran on and what the evaluation measured. A version changes when Aria’s model, instructions, tools, sources or guardrails change. The first number moves when Aria can do a new kind of thing or act with more autonomy, the second when what it says or proposes changes, the third for a fix. How Aria is governed is on the transparency page; the provider’s record is Aria’s public AI System ID.

A date is the day the version was completed. It reaches the product with the next release of Veritome.

3.9

13.11.2026
Current

Aria reads the obligation you are on, and drafts its assessment from it

On an obligation's page, Aria now knows which obligation you are looking at. Asked to prepare an assessment draft or to explain it, it reads that obligation (what it requires, its checklist, its form fields and the answers already filled in) and writes the draft in the conversation: what to record, the evidence that would prove it and what is still missing. The draft is text you review and enter yourself; Aria fills in, ticks and approves nothing. The GDPR is now always among its knowledge sources, and in chat ISO/IEC 27001 joins them when a workspace switches that standard on. Supplier research also keeps a certification whose stated end date is still to come ("valid until December 2027"), which it used to drop as lapsed.

What changed
  • ToolsA new read-only tool, getObligation, returns one obligation in detail: its requirement and guidance, its checklist with what is ticked, its form fields with the answers filled in so far, the latest record and how many pieces of evidence are linked. It applies the obligation page's own visibility rule and the person's "Obligations and due dates" switch. The obligation list now gives each obligation's id, so Aria can open one from it.
  • ScopeThe Aria panel sends the obligation's id with each message asked from an obligation's page; it is used only when the person may open that obligation and has obligations switched on, and only while they stay on that page.
  • InstructionsAsked for an assessment draft, Aria reads the obligation first and drafts from what it holds; where an answer depends on a fact the record does not hold, it names the fact to supply instead of inventing one.
  • SourcesThe GDPR is always among Aria's knowledge sources: it is binding law, and it used to depend on the optional GDPR programme being switched on. A new ISO/IEC 27001 summary is used in chat when the workspace switches that standard on.
  • GuardrailSupplier research keeps a certificate whose stated end date is still to come: the check that a quote supports a certification reads the date after "until", "bis", "expires" or "through" (a day, a month or a year) against the day the supplier's pages were read, and reads "held" as current only in "have held", "has held" or "is held". An "until" or "expires" with no date it can read still does not count.
Model

Mistral Small 4, pinned to mistral-small-2603; agent runs on Mistral Large 4, pinned to mistral-large-4-0

Evaluation

Seven-suite run of 13.11.2026, against the live model: 6 of 7 suites reached their gate, 0 forbidden outcomes.

  • Answers about the law · 19/20 · gate 95% · at its gate
  • Proposals and their limits · 98/120 · gate 95% · below its gate
  • Questions about your workspace · 51/52 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 33/36 · gate 90% · at its gate
  • Help for visitors · 92/94 · gate 90% · at its gate
  • Trust questions · 16/16 · gate 90% · at its gate

3.8

13.11.2026

A reclassification changes only what changed, and a question is only answered

When Aria proposes the reclassification of a system, it now gives only the registration answers that the goal changes, and every other answer is taken from the system's record. The proposal records which answers changed. An agent run whose goal is a question is now given read tools only, so it answers and proposes nothing.

What changed
  • ToolsThe reclassification tool takes only the changed answers and lays them over the answers on the system's record. A call that changes no answer is refused, and the proposal records which answers changed.
  • GuardrailA goal that is a question, and asks for nothing to be done, runs with read tools only; a call to a tool that would propose something is refused in code.
Model

Mistral Small 4, pinned to mistral-small-2603; agent runs on Mistral Large 4, pinned to mistral-large-4-0

Evaluation

Seven-suite run of 13.11.2026, against the live model: every suite reached its gate, 0 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 115/120 · gate 95% · at its gate
  • Questions about your workspace · 48/48 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 36/36 · gate 90% · at its gate
  • Help for visitors · 91/94 · gate 90% · at its gate
  • Trust questions · 16/16 · gate 90% · at its gate

3.7.1

12.11.2026

Aria's model calls go to Mistral's EU endpoint

Every call Aria makes to its models now goes to Mistral AI's EU regional endpoint, which runs inference in data centres in EU and EFTA countries. Until this version the calls went to Mistral's global endpoint, for which Mistral does not commit to an inference location. The pinned models, the instructions and the tools are the same as in 3.7.0.

What changed
  • HostingAll model calls go to Mistral's EU endpoint, api.eu.mistral.ai, instead of the global endpoint, api.mistral.ai.
  • GuardrailA test fails if any source file contains the global endpoint's https address, and the deployment check asks both pinned models on the EU endpoint.
Model

Mistral Small 4, pinned to mistral-small-2603; agent runs on Mistral Large 4, pinned to mistral-large-4-0

Evaluation

Seven-suite run of 12.11.2026, against the live model: 6 of 7 suites reached their gate, 1 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 113/120 · gate 95% · below its gate
  • Questions about your workspace · 48/48 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 35/36 · gate 90% · at its gate
  • Help for visitors · 91/94 · gate 90% · at its gate
  • Trust questions · 16/16 · gate 90% · at its gate

3.7

12.11.2026

Agent runs decide their steps on Mistral Large

An agent run, where Aria reads a workspace and proposes actions, now runs on Mistral Large 4: the large model decides each step and writes what the run proposes. Chat, the public assistants, suggestions and drafts made outside a run stay on Mistral Small 4. On the small model the agent test set stayed below its 95% gate: in four ten-repeat runs on 12.11.2026 it passed the reclassification case in one, two, two and five attempts of ten. Both models are Mistral AI's and are called over the same API.

What changed
  • ModelAn agent run calls Mistral Large 4, pinned to the fixed version mistral-large-4-0, for every step. The content of what a run proposes, such as the body of a draft or the clauses of a policy, is written by that model.
  • ModelEverything else still calls Mistral Small 4, pinned to mistral-small-2603: chat, the public help and trust assistants, field suggestions, and drafts made outside an agent run.
  • GuardrailIf Mistral answers that a pinned model is unknown or retired, the call is retried on the newest model of the same family, so an agent run is not moved to the small model.
Model

Mistral Small 4, pinned to mistral-small-2603; agent runs on Mistral Large 4, pinned to mistral-large-4-0

Evaluation

Seven-suite run of 12.11.2026, against the live model: 6 of 7 suites reached their gate, 1 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 107/120 · gate 95% · below its gate
  • Questions about your workspace · 48/48 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 36/36 · gate 90% · at its gate
  • Help for visitors · 93/94 · gate 90% · at its gate
  • Trust questions · 16/16 · gate 90% · at its gate

3.6

12.11.2026

Only a person makes a system's first classification

Aria no longer proposes the classification of a system that has never been classified, and such a proposal made earlier can no longer be approved. A person answers the registration questions and the rules engine computes the tier. Aria may still propose the reclassification of a system that is already classified, for a person's sign-off.

What changed
  • GuardrailA classification proposed for a system that has never been classified is refused in code, whatever the goal says. One proposed before this version is refused when someone tries to approve it.
  • InstructionsIn a run, Aria is told to state no risk tier for such a system and to say that completing its registration is the next step.
  • ToolsThe classification tool is described to the model as the reclassification of a system that is already classified, and its proposals are titled as reclassifications.
  • ScopeThe example goals in the Run tab no longer include classifying unclassified systems.
  • EvaluationThe agent case that expected a first classification now expects none and counts one as a forbidden outcome.
Model

Mistral Small, pinned to mistral-small-2603

Evaluation

Seven-suite run of 12.11.2026, against the live model: 6 of 7 suites reached their gate, 0 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 109/120 · gate 95% · below its gate
  • Questions about your workspace · 48/48 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 36/36 · gate 90% · at its gate
  • Help for visitors · 90/94 · gate 90% · at its gate
  • Trust questions · 16/16 · gate 90% · at its gate

3.5

12.11.2026

The agent reads a system's registration answers

An agent run can read the registration answers held on a system's record, so a reclassification starts from the answers on its record. A reminder added in 3.4, which pressed a run to propose a classification, was removed: on a system whose record held no answers it pressed the model to make them up. The agent suite is still below its gate.

What changed
  • ToolsReading a system returns the registration answers on its record, or none for a system that was never classified.
  • InstructionsA classification starts from the answers on record and changes only what the goal or the person says. Aria is told never to fill in an answer nobody gave.
  • GuardrailThe refusal of a classification with an open registration question now names every open question and says not to answer them.
  • InstructionsThe reminder that pressed a run to propose a classification, and the forced tool call after it, were removed.
  • InstructionsThe one-time reminder for a run about to end without the proposal its goal asked for now also covers a goal that asks to schedule a review.
  • EvaluationThe agent suite runs each case ten times in the build pipeline. Two of its cases were corrected: one expected a classification of a system described in two sentences, the other asked for the classification its system already had.
Model

Mistral Small, pinned to mistral-small-2603

Evaluation

Seven-suite run of 12.11.2026, against the live model: 6 of 7 suites reached their gate, 0 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 106/120 · gate 95% · below its gate
  • Questions about your workspace · 47/48 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 36/36 · gate 90% · at its gate
  • Help for visitors · 92/94 · gate 90% · at its gate
  • Trust questions · 16/16 · gate 90% · at its gate

3.4

12.11.2026

First full evaluation run, and the fixes it led to

All seven evaluation suites ran against the live model in the build pipeline for the first time, on 11.11.2026. This version carries the fixes those runs led to. After them, six suites reached their gate in the run of 12.11.2026 and the agent suite did not. Public Aria gained a bot check and rules for questions about account access.

What changed
  • ToolsIn chat, a system is found by its initials or by a name such as "the hiring tool", and an obligation by the meaningful words of its title.
  • InstructionsA question about what is due by a date is answered from the list of obligations. Names are given as the workspace holds them. A duty that belongs to another party is answered with its Article.
  • ScopeThe scope check is told to treat a question about obligations, deadlines or due dates as in scope.
  • GuardrailSupplier research drops a fact when the quoted passage says the supplier does not hold it, plans it for the future or is about another company.
  • InstructionsPublic Aria does not troubleshoot an account, repeat a credential or explain a way round a security control. It points to the published way back in.
  • GuardrailPublic Aria takes a signed proof-of-work with every question, spent once, and has a daily ceiling per network range as well as per visitor and address.
  • GuardrailAn agent run that mistypes the identifier of the system it is working on is corrected to that system, never to another one.
  • EvaluationThe suites run in the build pipeline against the live model. Six help cases on account access were added.
Model

Mistral Small, pinned to mistral-small-2603

Evaluation

Seven-suite run of 12.11.2026, against the live model: 6 of 7 suites reached their gate, 0 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 108/120 · gate 95% · below its gate
  • Questions about your workspace · 48/48 · gate 95% · at its gate
  • Supplier research · 14/14 · gate 90% · at its gate
  • First policy drafts · 36/36 · gate 90% · at its gate
  • Help for visitors · 91/94 · gate 90% · at its gate
  • Trust questions · 15/16 · gate 90% · at its gate

3.3

01.11.2026

The model is pinned, and the scope check fails closed

The model Aria calls became a fixed version written in the code, so a model change is a reviewed change like any other. The check that decides whether a question is in scope stopped letting a question through when it could not decide.

What changed
  • ModelAria calls one pinned model version, mistral-small-2603. A deployment setting can no longer change it.
  • GuardrailWhen the scope check errors, times out or returns no verdict, the question is not answered and the use is given back. It was treated as in scope before.
  • UsageEvery assist counts against the monthly allowance, reserved before the model is called.
  • ScopeAria's one-shot routes, records opened by link, lists and exports show only the systems the person is assigned to.
  • ToolsAn approved task sets a due date and an owner. It never completes work or moves its target.
  • ScopeFive assist routes that no screen used were removed.
Model

Mistral Small, pinned to mistral-small-2603

Evaluation

Not run against the live model before this version: the build pipeline had no model key until 11.11.2026. The run recorded under version 3.4 is the first to cover it.

3.2

24.10.2026

Aria for visitors: help and trust questions

A separate, anonymous mode answers visitors' questions on the Help Centre and the trust page from public content only. It has its own instructions, its own four tools and no access to any workspace. It is switched on per site by Veritome.

What changed
  • ScopeA new anonymous scope with two modes, help and trust, each answering from what Veritome publishes.
  • ToolsFour public tools: search the Help Centre, search the EU AI Act, list the Academy courses, read the trust facts.
  • GuardrailA daily limit per visitor and per address.
  • EvaluationTwo suites added: help and trust.
Model

Mistral Small

Evaluation

Not run against the live model before this version: the build pipeline had no model key until 11.11.2026. The run recorded under version 3.4 is the first to cover it.

3.1

23.10.2026

Aria proposes records: risks, suppliers, training and policy drafts

Aria can propose a risk, a supplier, a course assignment and a first policy draft. A risk, a supplier and an assignment land in the Inbox, and approving one runs the same path a person's own form runs. A policy draft opens in the policy editor, where a person keeps or drops each clause. After onboarding, a first run proposes a starting set from the answers given.

What changed
  • ToolsNew proposals: a risk on a system's plan, an organisation risk, a supplier, a training assignment and a policy draft, clause by clause against its template.
  • ToolsA first run after onboarding proposes suppliers, three policy drafts and the Art. 4 course. Classifying a system proposes up to six risks from Veritome's risk library, chosen by rules with no model call.
  • ToolsSupplier research reads a supplier's own public pages after the supplier is approved. Every fact it keeps must quote the page it came from.
  • GuardrailA policy draft is refused when a clause claims compliance or certification, or cites an Article that does not exist.
  • ScopeWho may approve a proposal, and who may switch Auto-apply on for it, follows the permission of the role for that kind of record.
  • ToolsA daily run can propose tasks and reviews on its own initiative. It never applies them, and it is off unless Veritome switches it on.
  • EvaluationTwo suites added: supplier research and policy drafts.
Model

Mistral Small

Evaluation

Not run against the live model before this version: the build pipeline had no model key until 11.11.2026. The run recorded under version 3.4 is the first to cover it.

3.0

05.10.2026

Chat reads your workspace

Until this version, chat answered questions about the law. From this version it can read the workspace of the person asking, through read-only tools limited to their organisation and to the systems they may see, and it can propose a task or a review into the Inbox.

What changed
  • ToolsRead-only tools for chat: systems, obligations, evidence, recent activity, the Help Centre and the person's own Academy progress.
  • ScopeEvery lookup is limited to the organisation and narrowed to the systems the person can see.
  • ToolsChat may propose a due date for an obligation or a periodic review. Both go to the Inbox, under the same gates as an agent run.
  • GuardrailThe outcome of every chat turn is recorded: answered, unmet, out of scope, refused or error.
  • ToolsA drafted document opens in its regular editor, prefilled, for a person to finish.
  • EvaluationA workspace suite was added, in English and German, with adversarial cases.
Model

Mistral Small

Evaluation

Not run against the live model before this version: the build pipeline had no model key until 11.11.2026. The run recorded under version 3.4 is the first to cover it.

2.2

02.09.2026

Gates on what the agent may propose, and a test suite for it

The agent mode gained gates that refuse a proposal the run has not grounded, and its own evaluation suite, which scores what a run read, what it proposed and at which autonomy.

What changed
  • GuardrailA proposal about a system is refused unless the run has read that system. A proposal that states a legal conclusion is refused unless the run has read the Act.
  • GuardrailA proposal that names an obligation is refused unless that obligation exists on the system.
  • GuardrailA reference in an agent's answer that does not resolve against the regulatory registry is flagged in the answer.
  • ToolsSearching the Act ranks results by the terms of the query.
  • EvaluationAn agent suite of 12 goals, each run ten times. The gate reads the rate over all runs, and a single forbidden outcome fails the suite.
Model

Mistral Small

Evaluation

Eval 12 of 02.09.2026, against the live model: every suite reached its gate, 0 forbidden outcomes.

  • Answers about the law · 20/20 · gate 95% · at its gate
  • Proposals and their limits · 114/120 · gate 95% · at its gate

2.1

26.06.2026

Answers grounded in the regulatory registry

Dates, penalties and prohibited practices are answered from Veritome's regulatory registry and no longer from the model's memory.

What changed
  • SourcesThe registry covers every Article and Annex of the Act, and statutory facts are answered from it.
  • GuardrailCitations in a chat answer are checked against the registry.
  • EvaluationThe golden set of 20 questions about the law, scored on facts, citations and refusals, reached its gate.
Model

Mistral Small

Evaluation

Golden set run of 26.06.2026, against the live model: the suite reached its gate.

  • Answers about the law · 20/20 · gate 95% · at its gate

2.0

25.06.2026

Aria proposes actions for approval

Aria gained an agent mode. Given a goal, it reads with tools and proposes actions such as a classification, a set of obligations or a task. Nothing it proposes takes effect until a person approves it in the Inbox. Aria was classified as a limited-risk system with Relay Labs as its provider, and its system card was published.

What changed
  • ToolsAn agent run reads with tools and proposes writes. Every proposal waits in the Approval Inbox.
  • GuardrailAn autonomy policy decides what each proposal waits for. A classification always needs a sign-off, and filing with a regulator is never proposed.
  • ScopeAria was classified limited risk, provider Relay Labs Ltd, and given a system card in the app and on the Trust Centre.
  • EvaluationA golden set of 20 questions about the law was created.
Model

Mistral Small

Evaluation

The golden set was first scored on this day at 8 of 20, before registry grounding. The first run at its gate is recorded under version 2.1.

1.0

12.04.2026

First release

Aria answers questions about the EU AI Act inside Veritome and helps with obligation forms. It gives information, not legal advice.

What changed
  • ScopeA chat assistant for questions about the EU AI Act, with help on obligation forms.
Model

Mistral Small

Evaluation

No evaluation set existed yet. The first one is recorded under version 2.0.