AI that can prove it's right
Encode any rule, policy, or law as code that returns the same answer every time, with a proof of why. Language models synthesise the rules at build time. A constraint solver evaluates them at decision time.
What you can do today
Run any published public ruleset from the browser, curl, the CLI, the SDK, or an MCP-aware agent. No signup, no key, no model-provider key.
Compiling your own rulesets and publishing them is in a closed developer preview. Tell us what you are building and we will get you an invite.
Request developer access →Synthetic data only
Use made-up values. Do not enter real personal, client, or confidential information. What you type here is sent to the Aethis engine to be evaluated, and may be recorded there. How decisions are logged and retained.
Affordability check for unsecured consumer credit.
Loading the ruleset before this request can be executed…
curl -X POST https://api.aethis.ai/api/v1/public/decide \
-H "Content-Type: application/json" \
-d '{
"ruleset_id": "aethis/consumer-credit-prequalification",
"field_values": {
"credit.has_adverse_history": false,
"credit.employment_status": "employed",
"credit.gross_annual_income": 50000,
"credit.has_joint_applicant": false,
"credit.joint_annual_income": 0,
"credit.dti_percent": 25,
"credit.credit_score_band": "excellent",
"credit.product_type": "unsecured",
"credit.ltv_percent": 0,
"credit.stress_test_passed": true,
"credit.is_existing_customer": true
},
"include_trace": true,
"include_explanation": true
}'Navigation, not just decisions.
Legislation isn't a checklist. It's a forest of criteria with branches, exceptions, and alternative routes. Aethis maps the traversable paths through the rules and guides a user towards the best outcome given their situation — surfacing viable routes, not just a verdict.
Soft constraints find the path. Hard constraints hold the law.
What is Aethis?
Aethis compiles legislation, policy, and contracts into formally verified rules. Language models read the source at build time and synthesise candidate rules; subject matter experts (SMEs) write tests that define correct behaviour; the engine refines rules until every test passes. At decision time, a constraint solver evaluates the compiled rules against the supplied facts. No language model in the request path. Each decision response identifies the engine version, decision id and input hash. Request include_trace for criterion-group outcomes and include_explanation for the decision explanation; both are optional and off by default. This panel requests both and loads published rule logic separately.
- 0
- LLM calls in the request path
- 0
- Aethis keys in the public API request — copy it from the curl tab above and run it in your own terminal
What the benchmark found
On a 68-scenario adversarial suite built around three-level exception chains, frontier models scored between 88% and 100% against the benchmark's formal rule fixtures, and the failures were systematic rather than random — nested exceptions collapse, and alternative routes stop being evaluated once a primary route fails. A compiled ruleset agrees with those fixtures by construction, because it executes the same rules deterministically.
Simpson 2026 §6.3 — Confidently Wrong: Exception Chain Collapse in Frontier LLM Rule Evaluation →How does Aethis compare?
How Aethis compares to other ways of making rule-based decisions.
| Capability | Aethis | LLM only | Static rules engine | Decision tree |
|---|---|---|---|---|
| Same input → same output | Yes | No | Yes | Yes |
| Replayable decision record | Yes | No | Manual | Partial |
| Source citations on a rule | Where the ruleset publishes them | No | Manual | Manual |
| Authoring from source text | LLM-assisted | n/a | Manual | Manual |
| Handles legislation directly | Yes | Yes (probabilistic) | Hand-encoded | Hand-encoded |
| Auditable for regulated workflows | When rules are explicit and validated | Not deterministic | Yes (no citation contract) | Yes (fixed branching) |
Ready to build?
Authoring and publishing your own rulesets is invite-only during the developer preview. Tell us what you're building.
Request developer access