How We Grade Claims: Four Marks, Three Horizons · Triodian

How we grade claims

We flag our own claims. Here is the key.

Every claim on this site carries one of four marks, and every offer one of three horizons. Nothing is described as available before it is. Dates for software and hardware are planned, and say so.

Four marks

How sure we are that a claim holds.

Demonstrated

Running in a live product

Semantic modelling, constraint reasoning and continuous change detection run in pinpole, a live commercial product, on a structurally different problem.

Conventional engineering

No experiment stands in front of it

Deterministic rules enforcement at the output boundary, signed hash-linked records, a hardware root of trust. Established methods, built to a specification.

Designed, not built

Specified, not yet built

The integrated Deterministic Governance Architecture as one system, and the sealed hardware tier.

Empirical hypothesis

Waiting on a pre-registered test

That distance from an authored reference tracks compliance, and that the authored drift statistics catch what they were written to catch.

Three horizons

When you can have it.

Available now

You can engage it today

Delivered now, or demonstrable today.

In build

Being built, with a planned date

Each carries the date it is planned for. Plans can move; when they do, this page changes first.

Next

The step after

The hardware appliance and its configurations, each behind its own gate.

Available now

What you can engage today.

OfferShapeHorizon
AI Obligations AtlasFree to use; web app, REST API and MCP serverAvailable now
AI Board DirectionOne workshop and one board meetingAvailable now
AI Risk Appetite sessionsTwo board sessionsAvailable now
AI Use AssessmentThree to four weeksAvailable now
AI Confidentiality Exposure AssessmentTwo weeksAvailable now
AI governance advisoryThe app suite run with you, by route; standing supportAvailable now
The app suiteOpen source under Apache 2.0, free to self-host: Obligations Atlas, Board Direction, Risk Appetite, Use Assessment, Confidentiality ExposureAvailable now · open source

Products

Every product, its horizon and its mark.

The rules tier does not depend on any experiment. The meaning tier does, and says so.

ProductHorizonClaim mark
Semantic GateIn build · planned December 2026Range check: conventional engineering. Meaning check: empirical hypothesis.
Attested Semantic CompareIn build · mid-2027Demonstrated in a prototype. Distance check: empirical hypothesis.
Aggregate Drift ServiceIn build · mid-2027Mechanism: conventional engineering. Whether the statistics catch what they were written to catch: empirical hypothesis, settled on a live stream.
Structured Output AssuranceIn build · December 2026Conventional engineering.
Bounded Command and Action GovernanceIn build · December 2026Conventional engineering.
Certifiable Conformance PackIn build · December 2026Conventional engineering.
Compliance token and ledgerIn build · December 2026Conventional engineering.
PACE authoringIn build · certification late 2027Designed, not built. The miss rate is to be certified per constraint.
Harness-as-AssuranceIn buildDesigned, not built.
Rules applianceNext · prototype mid-2027Designed, not built. A check the AI and its operators cannot reach. First customer deployments planned for 2028.
Distribution applianceNext · after certification, late 2027Designed, not built.
Meaning applianceNext · decided first half 2028Empirical hypothesis, decided by a 10-week pre-registered experiment.
Hosted applianceNext · 2028Designed, not built.

Horizons as at 6 November 2026, from the current roadmap.

The experiment

Each decision rule is declared before any measurement.

No failure removes enforcement. It removes one tier above the rules floor. Alongside sit cheap diagnostics built to falsify the claim early.

TestWhat it asksWhenIf it fails
1 · ScreenCan the constraint be told apart in any representation at all?Before the experimentThe meaning work is not funded. Rules are unaffected.
2 · ExperimentDoes the arbitrated verdict separate compliant from violating, and beat a single-metric baseline by a declared margin?About 10 weeksMeaning cannot load. The distribution envelope still ships.
3 · Live accuracyDo the authored drift statistics catch what they were written to catch, on a live decision stream?In a live deploymentNothing is owed. The rules and software tiers stand.

What a board can rely on

Repeatable is not the same as correct.

Can be relied on

  • A rules verdict is decidable and repeatable
  • Each verdict is bound to a policy version
  • The record can be checked by a third party
  • Failure modes are declared in advance

Not claimed

  • That a semantic score proves correctness
  • Non-bypassability outside an enclosed deployment
  • A bound that survives conditions it was never calibrated on
  • That the patent will be granted

Get in touch

Ask us where anything sits on this page.

If a claim here would not stand in front of your board, your insurer or your supervisor, that is the most useful thing we can hear.

Book a 30-minute conversation →