| --- |
| title: JE Validation Console |
| emoji: 🧾 |
| colorFrom: pink |
| colorTo: purple |
| sdk: docker |
| app_port: 8000 |
| --- |
| |
| # JE Validation Console |
|
|
| An RL-environment console for **journal-entry validation**: seeded, scored |
| tasks (tiers L1 deterministic checks → L2 cross-document reasoning) over a |
| simulated general ledger. Pick a task, a model, and a seed count; the agent |
| works the ledger through a fixed tool set (query, open documents, recompute, |
| disposition, submit) and is scored by deterministic verifiers on outcome, |
| process, and citation discipline. Two scripted baselines (flag everything / |
| no evidence) run free and anchor the score floor. |
|
|
| Every `(task, seed)` pair regenerates a byte-identical instance; defect |
| ground truth is held out from the agent side of the tool boundary. Citing a |
| document that was never opened counts as fabrication and fails the episode. |
|
|
| **Configuration**: LLM episodes call models through OpenRouter; set an |
| `OPENROUTER_API_KEY` secret in the Space settings to enable them. Baseline |
| runs work without a key. Run history lives on ephemeral Space storage and |
| resets on restart. |
|
|
| Ledger data is derived from the State of Connecticut's open data portal |
| (data.ct.gov, dataset `ajdm-rvz7`), snapshot 2026-08-04. |
|
|
