# FALSIFIER amendment 1 — hook-site / SAE-fidelity gate

Written: 2026-10-27T10:03:07Z (2026-10-27 04:03 MDT, America/Denver). Original FALSIFIER.md sha256 f74b1b4f7219465283cba898848db0f1a2d3d8a685481cd80ce956255c597bf3 (unchanged).

## What happened
Calibration stage ran (calibration prompts only; no test prompt generated, no steering effect computed).
The pre-registered gate in section 5 ("If both FVU > 0.5, the harness is declared NOT LOADED") tripped:
- FVU, block output: 0.529; FVU, block update (output - input): 1.803.
- Median patch norm: block output 145.8, update 106.5, Goodfire's published example_activations.pt 143.8.
- The same FVU metric applied to Goodfire's own published example activations
  (github.com/goodfire-ai/sdxl-turbo-interpretability, example_activations.pt) gives 0.447.
So under the original rule the run is NOT LOADED. That result is kept on record (logs/calib_attempt1.log).

## Why amend
The 0.5 threshold was set without a reference point. Measured the same way (per-image mean as the FVU denominator),
the SAE's own published example scores 0.447. Our block-output activations sit close to it (0.529), and the
patch-norm scale matches within 1.4%, so the hook site and input convention are the ones Goodfire used.
The update convention is ruled out (FVU 1.80).
This amendment is post hoc with respect to the FVU value. It is made before any steering effect or any
test-prompt output exists.

## New gate (replaces the last paragraph of section 5)
LOADED iff, for the convention with the lower FVU:
(a) FVU <= FVU(Goodfire example) + 0.15, and (b) median patch norm within +/-20% of the Goodfire example.
Everything else in FALSIFIER.md (prompts, metric, controls, thresholds, decision rule) is unchanged.