File size: 1,306 Bytes
f67de90 59db361 f67de90 59db361 f67de90 1c43ec2 59db361 1c43ec2 59db361 1c43ec2 59db361 1c43ec2 59db361 1c43ec2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 | ---
title: JE Validation Console
emoji: 🧾
colorFrom: pink
colorTo: purple
sdk: docker
app_port: 8000
---
# Journal-Entry Validation — Centific AI Research
Can an AI agent audit like an auditor? This environment issues journal-entry
validation cases over real government ledger data. An agent works each case
with professional tools — querying the ledger, looking up reference data,
opening supporting documents — and is scored on three axes: **Outcome** (found
the right entries, rightly labeled), **Process** (worked like an auditor:
evidence opened, citations honest, no wasted steps), and **Rationale**
(reasoning quality, graded against human experts — upcoming).
Every case replays byte-for-byte, citing evidence the agent was never shown
fails the episode automatically, and every model must beat two built-in scripted
baselines — flag everything, answer without evidence — before its score means
anything.
**In the app:** start on **Home** for the story, **Datasets / Tasks /
Scoring** for the details, then the **Console** to configure and launch runs.
Baseline runs are free and need no key; LLM runs call models through
OpenRouter (`OPENROUTER_API_KEY` secret in the Space settings). Run history
lives on ephemeral Space storage and resets on restart.
© Centific AI Research
|