Operating principles

The short list. Print it. Pin it next to your monitor. Apply on every task.

The ten

1. Plan before pressing run. A one-page plan saves a one-week detour.
2. Baseline first, model second. No model is "better" without a number to beat.
3. One change per experiment. Two changes, learn nothing.
4. Reproducible or it didn't happen. Seed, version, command, hash. All four.
5. Numbers measured, never estimated. Empty cell beats fabricated number.
6. Read every diff. The agent's summary is intent. The diff is evidence.
7. Read the data with your own eyes. Twenty raw rows before trusting any model.
8. Verify splits at every refactor. Leakage hides in shuffles.
9. Doc updates in the same commit. Code without doc, doc without code — both wrong.
10. Own the output. The agent will not face the user. You will.

Why these and not others

Most "ML best practice" lists drown you in 50 items. These ten are the non-negotiables. Everything else is a special case of one of them.

Anti-principles (sound smart, aren't)

• "We'll add reproducibility later." Later does not come. The state that produced the result is gone within hours.
• "The agent already tested it." It ran the test. You did not check what the test asserts.
• "It's just a quick prototype." The prototype becomes the demo. The demo becomes the product.
• "We'll fix the metric next sprint." A wrong metric this sprint poisons every decision built on it.

How this list evolves

This is the v1 list. When a new failure mode catches you twice, add it. When a principle here has not bitten anyone in two quarters, retire it. The list should breathe with the team's actual experience, not freeze into a wall of platitudes.