Route → check → receipt
ClaimStack demonstrates dimensional checks, explicit refusals and sealed records using Python's standard library.
Run ClaimStack →JARVI3 / ECOKURE
Inspect the check.
Keep the evidence.
Explore the public tools behind an approach to AI verification: scoped checks, explicit unsupported outcomes, evidence receipts and replay. Run a small example, inspect its limits, then evaluate it in your workflow.
Exact arithmetic
DTL MathGate evaluates supported arithmetic with exact values. Its public corpus and regression tests expose the implemented scope.
Inspect MathGate →Replay and inspect drift
ReplayGate replays supported evidence packs and compares results. Inspect the implementation and its execution limits.
Explore ReplayGate →Public evidence record
This page records development evidence and submission status. It is not an external leaderboard. ProgramBench PR #27 remains open as of 1 October 2026.
| Entry | Track | Result | Status | Evidence |
|---|---|---|---|---|
| DTL MathGate lite Public arithmetic suite |
Development corpus and regressions | 192 exact answers 48 expected refusals 240 cases · 0 failures 18 unit tests pass |
Reproduced 1 Oct 2026 Windows/Linux CI passed. Public tests with a separate in-repository oracle. |
Run record Code and regressions |
| Jarvi3: Themis-G ProgramBench stored candidates |
Official benchmark submission | 2/200 resolved 1,037/1,037 tests passed |
Scores reproduced 1 Oct 2026 Official evaluator; registry PR pending. Generation provenance remains unresolved. |
Fresh reproduction Package Official PR #27 |
| DTL/SuperMath AIME-style public seed |
Case-specific arithmetic replay | 14/30 correct overall 0 incorrect 16 abstain |
Development evidence Known case IDs and problem-specific calculations; not an unseen-problem reasoning score. |
Full report |
The public lite tools implement bounded checks. Passing a check does not establish general scientific truth, production readiness, or validation of every E-stack layer.