Self-directed
Veritas
backend
decision systems
Project overview
Semi-formal business decision prover
Turns plain-language business proposals into structured decision objects, asks for missing premises, and verifies the final verdict with deterministic Python.
Date
2026
Field
Infrastructure
Stack
Python, Local web UI, Deterministic verification, LLM formalization, CLI tooling
12-case verifier battery
6 proposal transcripts
deterministic verdicts
Veritas
Context
Built as a decision-proving system for business proposals, with reproducible CLI runs, committed outputs, and a local audit UI.
The hard part
To prevent hallucination, Veritas keeps the model out of the final verdict path, so decisions are supported, refuted, or left undecidable based on supplied premises.
What it took
- Separated formalization from verification so the LLM can structure inputs but not choose.
- Built verifier routes for hiring, campaigns, channel tests, price changes, etc.
- Added audit trails with derivations, constraints, and editable JSON re-verification.