Correction: As-of feature computation

Sharpened the wording on As-of feature computation after someone pointed out it was overstated.

The original read as though the pattern always applies. It does not. Every feature must be computable from information available at prediction time. Write down each one's timestamp, and the leakage disappears.

What it should have said — and now does — is where it stops applying. Every technique has a range of validity, and a pattern library that states rules without stating their limits produces exactly the cargo-culting it was meant to prevent.

The revised page names the case where the simple version is fine, so you are not paying the cost of the careful version on a table with four rows.

Corrections go in the page, with a note about what changed. No silent edits. The course will contain errors — it is thirty-six lessons about a field where the standard advice is frequently wrong — and the only defensible approach is to fix them visibly.

The repository takes issues, and several pattern pages are sharper than what we shipped because someone argued.