a builder's codex
codex · patterns · No-such-thing-as-long-feedback-loop vs. one-year holdouts evaporate

No-such-thing-as-long-feedback-loop vs. one-year holdouts evaporate

Position A — Find a correlated short signal; there is no such thing as a long feedback loop

Position B — 30-40% of growth experiments with short-term lift evaporate at one year

Conditions distinguishing them

Resolution / synthesis

Genuine tension. Duke's prescription (find a short signal) is precisely the move Abrams's data shows fails 30-40% of the time. They cannot both be right unless the correlated-short-signal claim is conditioned on validation.

Resolution: Duke is right if the correlation is validated against a long-term holdout; Abrams's data shows that most teams skip the validation. The synthesised rule:

1. Find a candidate short signal (Duke).

2. Validate the correlation against a one-year holdout before using it to ship decisions (Abrams's evidence).

3. Re-validate quarterly because correlation drifts.

Without step 2, Duke's advice produces Abrams's evaporation. With step 2, the two cards are compatible: short signals work, but only the validated ones, and validation requires a long loop somewhere in the system. The genuine contradiction is whether most teams should trust their short-term signals — Duke implicitly yes, Abrams explicitly no.

Open the interactive view →