# Item 1: features 4045, 4071, 9820 under the amendment-3 protocol (internal)

Prereg: FALSIFIER_FEATURES_2.md sha256 9c60a66fa2fe2d8fcc4f36cdda4adcebebb50a4ef990972bc3e04ae9bb485d51
(mtime 05:44:01 MDT; first image written 05:46:58 MDT, so the prereg is earlier than any generation). Run exactly as written; not edited.
Generation finished 07:10 MDT (FEATURES2_GEN_DONE); analysis by harness/analyze_features2.py (CPU).
Calibration picked s = 4.0 for all three (m = 583.23 = 4 x median patch norm 145.807). 100 pairs per feature (20 prompts x seeds 0-4), 20 random features + planted fake each.

| Feature (label) | View | CLIP b_F [95% CI] | CLIP rank / max random | SigLIP b_F [95% CI] | SigLIP rank / max random | Rules | Verdict |
|---|---|---|---|---|---|---|---|
| 4045 putting hands on face in distress | crop 2.0x ("person head. human face.") | 6.71 [5.77, 7.64] | 1 / 5.93 (rand2281) | 5.31 [4.57, 5.99] | 1 / 3.59 (rand2447) | R1-R4 pass, both | PASS |
| 4071 hugging | full image | 10.61 [9.49, 11.64] | 1 / 3.63 (rand6867) | 6.31 [5.47, 7.11] | 1 / 2.51 (rand6388) | R1-R4 pass, both | PASS |
| 9820 squaring up to fight | full image | 8.88 [8.15, 9.68] | 3 / 10.37 (rand9921) | 6.70 [5.69, 7.79] | 2 / 7.02 (rand9921) | R2 (rank 1) fails, both | FAIL |

Gates: F pairs dropped 0/0/0 (coverage gate OK); planted fake did not pass for any feature (not INVALID); scorers agree on every verdict.
4045 detection rate 1.0 (baseline, F, planted; 0.997 overall); full-image sensitivity also PASS (CLIP 6.38, SigLIP 5.35, rank 1).
CLIP-SigLIP correlation of per-pair dS_F: 0.67 / 0.66 / 0.65.

## Caveats
- Each feature had one or more random features that would pass R1/R3/R4 on their own (listed in verdict_F*.json "random_false_positives"); the text-pair scorers respond to generic changes, so R2 (rank 1 of 21) carries most of the specificity.
- 4045 margin over the best random is modest on CLIP (6.71 vs 5.93; CIs overlap); clearer on SigLIP (5.31 vs 3.59).
- 4045 area ratio vs baseline (geometric mean) is 2.82: steering enlarges/changes the head region. The size term c (-0.33) and the matched-size subset (n=19, mean 5.77) are in the prereg protocol, but the size confound is larger than for 7446.
- s = 4.0 is the top of the grid for all three; subject drop at s=4 was 3.0-3.5, so subjects are partly degraded.
- 9820: a random feature (rand9921) scores higher on both scorers, so the effect is not specific under this protocol. That is a FAIL, not evidence the feature does nothing.
- Scorer text pairs are ours (CLIP/SigLIP), not human ratings. No human rating was done for these features.

## Paths
DGX ~/honesty-tools/runs/goodfire-features2/ (images in raw/img stay on the DGX); box mirror /workspace/honesty-tools/runs/goodfire-features2/ (raw/*.json, raw/*.jsonl).