Architecture
A planner proposes one step at a time. The runner checks every step against a policy, runs it through a tool server, checkpoints the result and writes an audit entry. Nothing the planner says can bypass the runner.
What is enforced, and where
- Resumable. Every step is checkpointed; a hard-killed run resumes at the first unfinished step.
- Budgets. Steps, wall time and tokens are limited. When one is exhausted the run jumps to a bounded wrap-up (compare, review, refit, predict, report) and still returns the best model so far.
- Guardrails. A code policy, deterministic reviewer checks, an untrusted-data sanitiser and a structural action guard that does not depend on the model behaving.
- Approvals. Deployment needs a single-use token bound to the action, the run and the model;
the run pauses as
awaiting_approvaluntil an admin approves it. - Two planners. A deterministic scripted planner (used for the example runs here) and an LLM planner that only chooses which model family to try next. The LLM never sees column names or cell values.
- Honest evaluation. Benchmark numbers on the results page are generated from result files, with the planner and LLM mode stated in every row.
What the live demo would disable
A hosted demo runs the same runner with uploads, deployment and the code sandbox switched off, the scripted
planner only, a 90 s hard wall clock (the agent itself stops searching at a shorter soft
wall and wraps up, which is why the example runs list both), a 20 s AutoML ceiling, one run at a time and per-address rate limits.
That mode exists in the repository (DEMO_MODE=1), is covered by tests and runs live on a
free Render instance at https://junior-ml-dev-demo.onrender.com/. The free instance has a throttled CPU, so
it skips the leakage detectors (recorded as a partial check), turns FLAML off and keeps a warm worker process.
The source repository is currently private, so file references below (result JSONs, code) are names only, not links.