Console - Monte Platform

The Console shows what the Ledger already knows. It is strictly read-only: the CLI is the only write path, and there is no monte console command. The server imports monte.api and nothing else. Reading Ledger files directly is banned, there as everywhere.

Run it locally

The Console is two processes:

# API server
uv run uvicorn routes:app --app-dir console/server

# Web app, proxies /api to 127.0.0.1:8000
cd console/web && npm run dev

The server reads MONTE_ROOT for the data root. The web app finds the server through MONTE_API_ORIGIN, default http://127.0.0.1:8000. On a laptop with an empty data root, seed demo data first:

MONTE_ROOT=~/monte-console-dev uv run python console/seed.py

Seeding runs mock and lands in seconds. A second seed against the same root refuses, because a Measurement is frozen at init. Point MONTE_ROOT elsewhere or delete the directory.

Pages

Training and Evaluations: the Measurements, each opening its Ledger table: Run rows with a score column, the Split each Run scored, and an Evidence toggle that hides smoke rows.
Run detail: the row in full: plan settings, base model, serving stack, checkpoint relations, live events, and for training Runs the reward curve.
Per-task results: every Trace of an eval Run with its score, expected answer, and extracted answer.
Environments: a task browser for each installed Environment.

Current limits

A Trace does not carry its Task’s question text. Cross-reference the task ID in the Environment browser.
Per-task results are not paginated server-side. The test Split ships whole and paginates in the browser.
Training stats show reward only. Loss and KL from the training log are not parsed yet.
There is no archive or delete. The Evidence toggle is the only filter.

⌘I