Triodian · provable AI governance
From the board’s decision to proof an outsider can check.
Can you prove what your AI did? And what your providers’ AI did?
Supervisors, boards, insurers and funders are asking for evidence of AI behaviour, not statements about it. Triodian turns your board’s direction on AI into a tested appetite and a register of every AI arrangement measured against it, in open-source software you own, with a regulated worked example running end to end today. Next, the same policy checks each AI action and leaves a sealed record a supervisor, auditor or insurer can verify without trusting you or your vendor.
One policy, three steps
Decide it. Enforce it. Put it beyond reach.
One policy, declared by the board, carried through every AI action to a record someone outside can check. Each step carries its horizon and its claim mark.
01
Decide it
Available nowBoard direction, risk appetite and AI-use assessments, run through your counsel in the open-source app suite. The appetite becomes limits and triggers a record can be tested against, including for the AI inside your providers.
See the engagements →
02
Enforce it
In build · from December 2026Each AI action checked against that same policy: allow, deny or refer to a person, with a sealed record bound to the policy version, for a third party to verify without trusting the operator or the provider.
See the platform →
03
Put it beyond reach
Next · prototype mid-2027Patent pendingA hardware check the AI and its operators cannot switch off, for when tamper-evidence is not enough. Same policy, same record.
See the appliance →
Each step makes the next one cheaper: the board’s direction bounds the appetite, the appetite becomes the policy, and the policy’s verdicts become the next review.
Proof
One regulated worked example, run end to end on every push.
The first step of the arc is not a slide. A fictional superannuation trustee, Coastline Super, is run through AI Board Direction, AI Risk Appetite and AI Use Assessment with their real service bindings, in a browser, on GitHub Actions: on every push to the installer and once a day. Every file that crosses a link is checked against its published schema. No customer data and no Cloudflare account are involved.
01
Board Direction
The chair records the board’s directions.
02
Risk Appetite
Loads the directions over the service binding, shows them on decision 1 and holds back any ceiling above them. The CRO agrees the working appetite; the board approves it.
03
Use Assessment
Imports the trustee’s own provider register, refusing two malformed rows; assesses the six arrangements behind critical operations; loads the approved calibration set; issues the in-use feed.
04
Risk Appetite
Decision 8 loads the in-use feed without reopening the approved appetite.
05
Board Direction
The chair takes the board view, which must match the approved calibration set, then gathers the APRA-letter evidence index from all three apps, by letter bullet and by ISO/IEC 42001 control.
Check it yourself.
The test, its fixtures and its workflow are public. The badge below is GitHub’s own report of the latest run.
A passing test shows the apps carry one policy from the board minute to the evidence index. It does not show that any appetite was right: that is the board’s, through counsel.
Two doors
Start from your situation.
Most of the record that would answer a supervisor, an insurer or a funder does not exist yet. Where it is missing, and who can produce it, depends on who is asking.
Door 1 · Regulated entities
Regulated by APRA, MAS, OSFI or EU financial supervisors
Your supervisor expects you to oversee AI inside your providers, within appetite and with triggers. Find out what each provider can actually evidence.
Door 2 · General companies
Outside prudential regulation, and using AI anyway
Private companies, not-for-profits and the mid-market: a board that needs a position on AI, a deal that must stay out of it, a funder or insurer asking questions, or plant that acts on its own.
Available now
Five engagements, advisory, and the open-source app suite that runs them.
Each engagement is diagnostic and names no product. It runs through your counsel, takes no customer data and needs no system access. The app suite that runs the assessments is open source under Apache 2.0.
AI Obligations Atlas
Which AI obligations are in scope, by jurisdiction and source tier, mapped to ten use cases. The matter brief every engagement starts from.
See the Atlas →AI Board Direction
Where the board wants AI, what it rules out, what must come back to it, and who is accountable. Standalone, or feeding the appetite.
See the steps →AI Risk Appetite
Eight board decisions that turn "within appetite" into limits and triggers a record can be tested against.
See the sessions →AI Use Assessment
Where AI acts for you, inside your providers and the platforms behind them, and what evidence each boundary can produce.
See the assessment →AI Confidentiality Exposure Assessment
For a live deal: every channel by which deal information can reach an external AI model, and what the company could show.
See the assessment →AI governance advisory
The app suite run with your counsel and board: a route for regulated entities, a route for general companies, and standing support.
See advisory →Why the usual answers fall short
A model cannot be its own evidence.
Policies, guardrails, red-teams and audits describe the rules well, and most of them work. None of them produces a record of what the AI did that someone outside can check.
Logs are testimony from the accused
Software writes them, and software can be rewritten or bypassed at machine speed.
Guardrails share the model's privilege
Whatever compromises the model compromises the guardrail. There is no layer beneath.
Consensus is correlated
Several AI reviewers drawn from the same training share the same blind spots.
Certification is a photograph
An audit certifies a process on a date. The model keeps changing afterwards.
Three horizons
What is available, what is in build, and what comes next.
Every offer carries its horizon wherever it appears. Dates for software and hardware are planned, and say so.
Assessments, advisory and apps
The AI Obligations Atlas, AI Board Direction, the AI Risk Appetite sessions, the AI Use Assessment, the AI Confidentiality Exposure Assessment and advisory. The open-source app suite that runs them can be used today.
Software controls
The Semantic Gate, planned for December 2026. Attested Semantic Compare and the Aggregate Drift Service on customer infrastructure from mid-2027. Structured output and bounded command enforcement from December 2026.
The hardware appliance
A rules appliance gating a live model, planned by mid-2027, with first customer deployments planned for 2028. A check the AI and its operators cannot reach, for when tamper-evidence is not enough.
The control
The control sits at the boundary, not inside the model.
Where an assessment finds that a control is needed, the control checks each proposed action against a policy declared in advance, and leaves a record a third party can verify.
Untrusted
AI proposes
Any model, rented or self-hosted. Its output is treated as a proposal, not an action.
The boundary
Check
Against a policy declared in advance: rules, a grammar, a drift window or, later, meaning.
Verdict
Allow, deny or refer
No pass, no action. Safe but unfamiliar behaviour goes to a person.
Evidence
Token
Each verdict sealed and bound to the policy version, for a third party to check.
What a third party can verify. That each recorded action passed a declared check, without trusting the operator or the provider.
What it does not show. That the policy itself was right. Authoring the policy is a separate, governed step: the board's appetite, through counsel where it is a legal question.
The condition. Non-bypassable only where every consequential path passes the boundary. That is proven per deployment.
Open by design
The method is open to inspection.
Open standard
Universal Governance Protocol
One open-source format, under Apache 2.0, for constraints, verdicts and provenance, so a governance claim holds wherever it is checked.
Read the protocol →Open vocabulary
Triodian Metamodel
An open, schema-first way to describe what an organisation does and why, so its AI can be governed in its own terms.
Explore the Metamodel →How we grade claims
Four marks on every claim
Demonstrated, conventional engineering, designed but not built, or empirical hypothesis. Nothing is said to ship before it does.
See the status legend →Get in touch
Find out what your AI can prove.
Thirty minutes to test whether the evidence question is live for you, and where it bites first.
Book a 30-minute conversation →