Platform / DTL
DTLIs this output or decision allowed?
Checks a decision against the rules that apply to it and returns one answer: yes, no, send it to a person, or cannot tell. It does not score confidence. It decides, or it declines to decide, and records why either way.
Deterministic Taxonomy Lanes · AU provisional 2026905289
Verification / A declared check
Explore what the method can establish.
Follow a supported route, an independent check and the record it leaves.
Select an illustrated object to inspect its role.
01 / 03
Choose an applicable method.
The declared workload and grammar determine which specialist route is supported. Unsupported work goes to review.
Supported families onlyRouting is bounded by the methods present in the intended implementation.
What it returns
Categorical, not probabilistic. Where a lane cannot decide, it says so explicitly rather than returning a low-confidence answer that reads like a decision.
What it does
- Classifies the output and selects the lanes that apply to it.
- Runs each lane deterministically. The same input produces the same verdict.
- Returns a categorical verdict rather than a probability.
- Declines to decide where a lane cannot decide, instead of guessing.
- Emits a signed record of what was checked and why.
What it does not do
- It does not judge whether an answer is good, useful or well written.
- It does not replace a model. It sits after one.
- It does not improve accuracy. It refuses to promote work that fails a check.
The questions this layer gets asked in procurement
| Concern | Why it decides the deployment |
|---|---|
| Latency | Lanes run in-line. Budget is per lane, not per model call. |
| False-block rate | The number that decides whether production tolerates it. |
| Abstention rate | Determinism bought at the cost of usefulness is a real risk, so it is measured. |
| Lane coverage | Which classes of output have a lane, and which fall through. |
What runs inside DTL
Each of these is a real implementation with code behind it. Status is labelled honestly, and nothing is listed as coming soon.
ClaimGate
Interactive labExtracts claims from AI or scientific output and routes each one to the deterministic gate that can settle it.
UnitGate
Interactive labChecks whether a quantity can be what it says it is, then converts it exactly into the units you need. 118 units, from mg/kg and mL/h to concentrations, pressure, temperature and device telemetry.
ElementGate
Interactive labChemistry validator covering formula parsing, molar mass and reaction balancing against a known-compound registry.
AMLGate
Open sourceAggregation gate for cash structuring. Across 26,521 synthetic customer-days it caught 356 of 356 structuring days and held 57% of legitimate cash-business days.
StatsGate
Open sourceDetects impossible or internally inconsistent statistics using GRIM and p-value consistency checks.
TheoryGate
Open sourceDeterminizer for mathematics and physics theory claims, returning exact verdicts rather than opinions.
ChipGate
Open sourceStructural sanity checker for Verilog RTL. Finds undriven outputs, multi-driven signals and unintended latches.
MedGate
Interactive labA research prototype for governed pilot validation. Checks dose binding, oncology dosing limits, interactions and paediatric restrictions against an explicit rule, for evaluation rather than clinical use.
ClaimLint
Open sourceLints documentation for unsupported AI and science claims. Runs as a CLI or a GitHub Action in CI.
MathGate
Open sourceExact-arithmetic engine producing replayable certificates. Of 1,116 ProofBench-X cases: 1,089 pass, 17 safely refuse, 10 are blocked, 0 fail.
Security Lanes
Open sourceReproducible test cases across CWE and OWASP weakness families with transparent scoring.
See it against your own workflow
A pilot runs in shadow mode. It controls no production decision, and the acceptance criteria are agreed before it starts.