Back to projects

Live project

TS-Reasoner

A verifier-first operation firewall where models may propose candidates, but typed checks decide what is accepted, repaired, remembered, or rejected.

Claim discipline: Claims on this page should be treated as hypotheses unless linked to a receipt, benchmark, or reproducible artifact.

What problem it addresses

Fluent systems can collapse confidence, plausibility, proof, memory, and policy into one opaque decision.

How it works

Keep proposal generation outside the proof boundary. Candidates enter typed verifiers, risk gates, repair loops, and receipt-backed memory transitions.

What exists now

The Habitat v3 runtime separates planning from execution: every action is revalidated, executed only in an in-process symbolic environment, observed, effect-checked, and transactionally committed with pre/post hashes.

Evidence / receipts

Habitat v3 adds verifier-gated goals, explicit topology, deterministic tension, plan verification, transactional symbolic execution, effect checking, replanning, reflection, lesson approval, multi-agent scheduling, and whole-run replay. The frozen 240-case functional and 240-case adversarial suites clear every safety hard gate.

Limits

Runtime integrity is not claim truth. Generated text and model confidence are not proof. The public experiments are bounded and domain-specific.

Next milestone

Measure and reduce conservative unreachable outcomes without weakening action, effect, replay, or renderer authority.

How to run / inspect

Install / open

git clone https://github.com/BoggersTheFish/TS-Reasoner-v0 && cd TS-Reasoner-v0

Inspect

Run the test suite and inspect verifier decisions, support types, repair receipts, memory writes, and abstentions.

Links