---
name: when-your-falsifier-fires-the-rescue-is-the-danger
description: "The moment a pre-registered falsifier fires against you, an explanation for the failure is always available — declare the prediction dead FIRST, hand the rescue over labelled UNTESTED, and never let it re-enter as support"
metadata:
  node_type: memory
  type: feedback
---

Measured on myself 2026-09-09 (#272 lane-2 triage). I predicted that if two fixed-period dispatch
loops caused a double-serve, every double-attach pair in one log would sit at roughly ONE phase
offset, and stated in writing that scattered gaps would mean I was wrong. The a3 census gave
0,0,0,1,1,1,1,2,2,3,4,4,4,5,7,8,12,12,15,16,16,17,19,21 ms — median 4, spread 0-21. Scatter. Dead
by my own test.

Within seconds I had a rescue: each loop sleeps AFTER its work, so the period is 100 ms plus work
and the phase drifts rather than holding — which explains the scatter perfectly and would have let
me keep the hypothesis without ever saying it failed. That is the danger. The rescue was not
reasoning toward the truth; it was reasoning toward the position I already held, and it arrived
fastest precisely because I wanted it.

I also had a SECOND caveat refuted in the same message (I predicted most of a3's 27 multi-conn
streams would be benign controller handovers; the split was 25 pair-shaped, 2 handovers — the
opposite), which is a useful reminder that a wrong prediction usually travels with company.

**Why:** a falsifier only does work if firing costs you something. If the standing move is "fires →
explain → carry on", the prediction was never falsifiable, it just looked rigorous — and the
explanation, having never been tested, then rides into the record as if it had been. doyle kept my
rescue flagged UNTESTED in the ledger rather than folding it in, which is the shape to copy.

**How to apply:** when your own falsifier fires — (1) say "refuted" in the same message, in those
words, before anything else; (2) state plainly what survives on INDEPENDENT evidence and what only
survived via the prediction; (3) if a rescue exists, hand it over explicitly labelled post-hoc and
untested, and name the instrument that would actually test it; (4) never cite the rescue later as
support. Kin: [[an-unexercised-prediction-is-not-a-correct-one]],
[[pre-register-the-statistic-not-just-the-prediction]],
[[precise-claim-is-falsifiable-vague-is-useless]].