DETERMINISTIC REASONING / CPU ONLY

Evidence in.
Accountable action out.

A conclusion should come with a reason. Debtaur turns explicit premises into inspectable decisions — and stops when the evidence doesn’t hold.

Explore the reasoning lab No model weights. No sampling.

01 / DERIVEFollow every dependency

02 / CHALLENGEExpose gaps & conflicts

03 / REPLAYVerify the same decision

01 / INTERACTIVE PLAYGROUND

Put the evidence to the test.

LIVE PUBLIC MIRROR

Start with a case, then change its evidence. The same server-side rules evaluate every run. Registration is simulated.

Loading cases…

Your evidence

0 / 4 required

Toggle a premise to withdraw it. Change its source to test provenance.

Introduce a contradiction“The system is not a high risk AI” · statedAdd an unreleased standardA taught interpretation with a source mismatch

Changes are evaluated when you run the evidence.

Decision & derivation

READY

Make the reasoning visible.

Choose a case or evaluate your evidence to see what follows — and what doesn’t.

Challenge the whole decision boundary.

Run six cases: complete, missing, conflicting, provenance risk, repaired, and taught evidence.

02 / THE MECHANISM

A decision you can take apart.

Assert the world

Explicit facts carry a source: stated, taught, or dreamt. “Stated” means asserted in this example world, not independently verified reality.

Derive the consequence

Four conjunctive rules connect disclosure to a complete submission. Each derived step lists the premises it depends on.

Gate the action

Missing facts refuse. Contradictions and provenance mismatches escalate. Only a complete, grounded chain unlocks the simulated tool.

SCOPE

This is a self-contained public mirror of a bounded workflow, not a connection to the private Debtaur engine. Rules are illustrative workflow policies, not authoritative legal text. Every evaluation derives afresh; withdrawing a dependency removes its consequences. The hashes are reproducibility checksums, not signatures.

03 / HISTORICAL ENGINE BENCHMARK

Bounded reasoning. Measured throughput.

RECORDED 2026-09-01

Original production-box measurements, retained from the previous demo. These are historical figures for the private engine, not a benchmark of this browser session.

311,543lines in the mixed corpus

49.2 smixed corpus ingest

0.6 msrecorded warm query · mixed

Production box · 1 September 2026 · reproduction command: make scale
Corpus	Lines	Propositions	Ingest	Lines / second	Warm query
college	3,182	2,940	0.35 s	9,078	0.3 ms
fairy	103,598	13,973	16.5 s	6,291	0.2 ms
fable	197,662	12,445	28.4 s	6,964	0.1 ms
mixed	311,543	12,907	49.2 s	6,335	0.6 ms

Evidence in.Accountable action out.