The platform
An assessment tells you where the evidence is missing. The platform is one way to produce it: controls that check what an AI proposes against a policy declared in advance, and leave a record a third party can verify. Each piece carries its horizon. Software is in build; the hardware appliance comes next.
The boundary
A model that gives no guarantees can still be useful. It cannot be its own evidence, so the check runs outside it.
Untrusted
Any model, rented or self-hosted. Its output is treated as a proposal, not an action.
The boundary
Against a policy declared in advance: rules, a grammar, a drift window or, later, meaning.
Verdict
No pass, no action. Safe but unfamiliar behaviour goes to a person.
Evidence
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.
Why the usual answers fall short
Software writes them, and software can be rewritten or bypassed at machine speed.
Whatever compromises the model compromises the guardrail. There is no layer beneath.
Several AI reviewers drawn from the same training share the same blind spots.
An audit certifies a process on a date. The model keeps changing afterwards.
Three tiers
The rules tier is decidable by construction: a check accepts or rejects, with no score or threshold to argue about. The software tier adds evidence for a self-hosted model and for the whole set of decisions. The hardware tier is the ceiling when tamper-evidence is not enough.
Rules tier
Output conforms to a supplied schema or grammar. Malformed output cannot be produced.
Read more →A closed, approved set of commands. Any action outside the set cannot be emitted.
Read more →The two above, with identical accept or reject every time and a tamper-evident conformance record per output.
Read more →Software tier
One governed path for everything bound for a model you run or rent: embed, adjudicate, record.
Read more →Each output checked inside an attested enclave, on a separate machine, against a versioned rule library.
Read more →Tests a window of decisions against appetite, where every single output passes but the set drifts.
Read more →Each verdict sealed, bound to its policy version, and written to a hash-linked record a third party can check.
Constraint authoring, versioned and signed separately from the model, with a certified miss rate per constraint.
Re-validates each constraint after a model change, a measured shift or a fixed interval, with a fallback that never leaves a constraint ungoverned.
Hardware tier
Grammar enforcement in a dedicated hardware datapath that no software at any privilege level can bypass. First customer deployments planned for 2028.
Read more →A certified envelope with a stated miss rate, per constraint, on the same interlock.
The semantic check on the same interlock, if a pre-registered experiment passes.
For customers who will not install hardware: the same verifiable token from an appliance run for them.
Dates are planned. Each operating point is owned by oversight, not by the team that benefits from fewer blocks, and is recalibrated on a model change, a measured shift or a fixed interval.
From a finding to a fix
The assessment stays solution-agnostic. A provider's own control, an existing tool or a contract term can fill the same rung; a Triodian product is one option, decided separately.
| Evidence rung | What it has to show | Triodian route, if chosen | Horizon |
|---|---|---|---|
| Disclosure | Where AI acts and what authority it holds | The AI use register, from the assessment | Available now |
| Attestation | That an output met a declared policy | Certifiable Conformance Pack | In build · 2027 |
| Event notification | That a set of decisions moved outside appetite | Aggregate Drift Service | In build · mid-2027 |
| Audit | What was checked, against which policy version | Compliance token and ledger | In build · 2027 |
| Continuous evidence | That the check still holds after the model changes | Harness-as-Assurance | In build |
| Deterministic control | That a forbidden action could not be taken | Bounded Command, then the hardware appliance | In build · then 2028 |
How much can be governed
Meaning is about a tenth, and waits on the experiment. About a tenth stays with people. Which kind of constraint applies decides which tier can enforce it.
Triodian working estimate, judgement-weighted, as a share of modelled decisive AI outputs across the economy. Not yet independently reviewed.
See the coverage analysis →How we grade claims
Runs in a live commercial product, pinpole, Triodian's sister company, on a structurally different problem: semantic modelling, constraint reasoning and continuous change detection.
Deterministic rules enforcement at the output boundary, signed hash-linked records, hardware root of trust. No experiment stands in front of it.
The integrated architecture as one system, and the sealed hardware tier. Specified, not yet built.
That distance from an authored reference tracks compliance, and that the authored drift statistics catch what they were written to catch.
What a board can rely on
The control governs what a system is permitted to do. It does not govern the advice an AI gives a person, and it does not prove the policy was right.
Can be relied on
Not claimed
What needs an experiment
Each decision rule is declared before any measurement. A failed test removes one tier above the rules floor; the rules tier does not depend on any experiment.
| Test | What it asks | When | If it fails |
|---|---|---|---|
| 1 · Screen | Can the constraint be told apart in any representation at all? | Before the experiment | The meaning work is not funded. Rules are unaffected. |
| 2 · Experiment | Does the arbitrated verdict separate compliant from violating, and beat a single-metric baseline by a declared margin? | About 10 weeks; planned for the first half of 2028 | Meaning cannot load. The distribution envelope still ships. |
| 3 · Live accuracy | Do the authored drift statistics catch what they were written to catch, on a live decision stream? | In a live deployment | Nothing is owed. The rules and software tiers stand. |
Get in touch
The assessments and apps are available now. Thirty minutes to see whether the question is live for you.
Book a 30-minute conversation →