About
About VerifiQuant
The paper
VerifiQuant: Governance-by-Execution for Financial LLM Workflows. Yu-Chi Ko, Mi-Yen Yeh and Shou-De Lin. In Proceedings of the 7th ACM International Conference on AI in Finance (ICAIF '26), Milan, Italy, November 14–17, 2026. ACM. doi:10.1145/3858063.3858688. Published under CC BY 4.0.
What this site shows
- Paper replay. Recorded outputs of R-PAL, R-PAL+ and VerifiQuant on all 80 traps and three clean questions, taken from the runs behind the paper's tables and scored with the paper's scorer. The replay makes no model calls.
- Verified repair. The loan contract's declared annuity-due transform, checked live by the same verifier the paper evaluates. It makes no model calls.
- Try it live. The full funnel on the paper's 250-contract library with the top 4 contracts retrieved, on Gemini 2.5 Flash. Selection, binding and ambiguity detection are LLM steps, so a live result can differ from a recorded one.
What the system guarantees, and what it does not
- Arithmetic stays outside the model. A number comes only from a contract's own code, after every gate passes.
- Inputs are checked for schema and bounds, not for provenance. A gate cannot tell whether a bound value came from the right field.
- Records replay as stored. Deterministic steps re-execute exactly from the stored binding. A fresh run from the raw text can take a different path.
- Repairs are bounded numerical validation, not formal equivalence: an edit must agree with a declared invariant on sampled bindings.
- Out of library it gives a scope signal, not more answers. Out-of-library means a formula family absent from the library, within the same benchmark.
Setup
BackboneGemini 2.5 Flash, every compared system
Contract library250 FICs, built offline, fixed at run time
RetrievalBM25 over contract metadata, top 4
LLM calls per questionSelect, bind, ambiguity
VerificationAST bounds + numerical invariant checks
ServingPython 3.11, Flask, SQLite, Google Cloud Run
Acknowledgments
This work was supported in part by the National Science and Technology Council, Taiwan, R.O.C., under Grant 113-2628-E-001-003-MY4; the National Taiwan University and Academia Sinica Innovative Joint Program under Grant AS-NTU-114-06; and the Institute of Information Science and the Research Center for Information Technology Innovation, Academia Sinica.