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