Rules
ExactFailure patterns your experts already recognise, encoded as explicit traversals. If one matches, the answer is not an estimate.
Decision intelligence
Decision intelligence for AI agents operating in high-stakes domains.
Hetu computes the decision, verifies it, and scores its own confidence — before your agent is allowed to act on it.
A wrong answer and a right answer arrive in the same font, with the same confidence, at the same speed.
Hetu removes the model from the part where that matters. Every figure an agent sees comes from a deterministic rule or a fitted causal model. The model only writes the sentence — and only about what was already proved.
How it answers
Three attempts. Each one knows what it is allowed to claim.
Causal model
AttributedA fitted graph spreads the anomaly across candidate causes, with intervals it will not round away.
44% · CI 37–51
26% · CI 19–33
Language, last
HypothesisThree ranked guesses at most, each with a test that would settle it. Forbidden from naming a cause.
See how it works
Watch one drop all the way to the bottom.
A question falls through the strictest method first.
Advancing automaticallyStep 1 of 3
24 rules your experts wrote downeach square is one failure pattern
No square matchedThis combination was never written down, so nothing here may answer.
The question drops to Tier 2 — and the missing rule gets logged as a gap to write.
The guard
Nothing reaches the agent until it passes every check.
Not a score bolted on afterwards. Failing one check refuses the conclusion rather than downgrading it.
The guard
one shared service✓Temporal ordering
✓Magnitude consistency
✓Gate validity
✕Sample-size floorn=3 of 5
–Label match · not reached
Conclusion refused
The first failure ends the call. Nothing downstream runs, and the label is not quietly downgraded.
- Action held
- Falsification test returned
- Written to the log
Five checks in sequence, called by every tier, so there is one place to audit what verification did.
Autonomy
Agents don't earn autonomy by being confident. They earn it by being measured.
The envelope starts closed. It widens only where measured outcomes prove the system right, and narrows on its own when they stop.
01 Constrain
What counts as a fault, what gates what, and what is never permitted, written before anything runs.
02 Circuit-break
Irreversible, fat-tailed decisions never auto-execute, whatever the confidence score says.
03 Record
Every decision generated, approved, deferred, executed and measured, in one immutable log.
04 Calibrate
Outcomes update the weights. Types that prove out move inside. Types that don't stay gated.
Go deeper
PlatformOne call, before the action. Either it verifies, or it refuses.Financial servicesForty agent pilots and nothing in production that touches a real decision.Commerce & growthYour growth agent is right most of the time. That is exactly why it's dangerous.ConsultingYour first agentic decision is too important for a self-serve trial. We build it with you.DeploymentWhen the blocker isn't one decision, it's getting anything live at all.Open sourceAn eval you can run against your own logs, before you talk to us.
Design partners
Bring the decision you don't currently let an agent make.
We work with a small number of enterprise teams where a wrong action is expensive. We map your constraints, seed the graph with your experts, and show you the first refusal.