MLLoop

A scientific-method harness for AI-driven machine learning.

Coding agents (Claude Code, opencode, ...) can already write training code and run ten variants overnight. What they don't do by themselves is science: diagnose why a model underperforms, form falsifiable hypotheses, run discriminating experiments, and — when the data itself is the problem — produce evidence strong enough to convince stakeholders.

MLLoop is an MCP server that sits between the agent and your training code and enforces that loop at the tool layer, not via prompts:

• Experiment ledger — every run, hypothesis, and decision recorded in SQLite plus an append-only JSONL event log, all under .mlloop/ in your project.
• Hypothesis gate — run_start refuses any experiment that doesn't test a registered, falsifiable hypothesis. No hypothesis, no run.
• Artifact contract — each run writes standardized predictions.parquet + meta.json; diagnostics never read your training code, so any framework works.
• Diagnostics battery — after every run: error slices, bootstrap noise floor ("what delta counts as evidence"), confusion/residuals, calibration, the operating curve (overkill vs catch rate, with degenerate-prediction detection), a SHAP explanation of missed positives (are they feature-limited or learnable?), and the overfit gap. Diagnosing the previous run is itself a gate: no diagnosis, no next experiment.
• Data Verdict Report — when runs stagnate, forensics_run interrogates the dataset with independent probes (shuffled-label signal check, confident-learning label-noise estimation, conflicting-duplicate bound, learning curve, per-feature signal) and report_generate renders a stakeholder-readable HTML verdict: is the ceiling set by the data or by the modeling? Demo: inject 20% label noise into a clean dataset — the report catches it, quantifies it, and lists the suspect rows.
• Domain context — context_register records what columns MEAN in domain terms (learned from dataset docs, domain MCP servers/skills, or the user); error slices and reports become domain-readable, and every report ships a data dictionary.
• FE-opportunity probe — fe_probe prices feature engineering before you spend runs on it: screens arithmetic combinations and stacked-model features for incremental signal with a paired, multiple-testing-adjusted significance bar. The probe generates hypotheses; the ledger tests them.
• Ensemble probe & paired comparisons — ensemble_probe prices combining finished runs with zero training (from their stored predictions); compare_runs and run_finish resolve small-but-real deltas with paired bootstrap significance on shared rows, far sharper than the single-run noise floor.
• Exploration discipline — stopping requires evidence (target met, high-confidence data-limited verdict, or budget exhaustion); until then status keeps the pressure on and stagnation suggests concrete pivots. Budgets cover both run count and wall-clock training time, and HPO sweeps are first-class runs.
• Custom metrics — a domain metric (AMS, weighted cost, ...) plugs in as a python file defining metric(predictions) -> float (goal_define(metric_script=...) or metric_register); the noise floor is then computed in the metric's real units. goal_define also refuses task-mismatched metrics and flags accuracy-on-imbalance with an advisory.
• Dashboard — "The Lab Ledger": lineage tree with hypothesis-labeled edges, metric journey with target line and noise-floor band, a narrated overnight log, evidence rail, and per-run dossiers — built for the morning-after review of an overnight autonomous session. The MCP server auto-opens it in your browser on the first tool call (MLLOOP_NO_DASHBOARD=1 to disable); mlloop dashboard serves it manually.

Status: Phase 2 — ledger, gates, diagnostics, forensics, reports, and dashboard. Full design: DESIGN.md. Agent setup (Claude Code / opencode / Codex): docs/integrations.md.

Quickstart

pip install -e .
cd your-ml-project
mlloop init --agent claude    # or opencode / codex / all — writes the MCP config

Then tell your agent to train a model. The enforced workflow:

status shows the current state and allowed actions at any time; ledger_query restores full context after an agent restart or context compaction.

Contributing

Issues, design feedback, and pull requests are welcome — see CONTRIBUTING.md. Please note the Code of Conduct.

License

Apache-2.0