# DEVIATION_2: rater exposure before blind rating (Kevin Russell)

Recorded: 2026-10-27 05:52 MDT (America/Denver). This was written before any ratings were received or scored.

## What happened
These times are as reported to Ford by Abe and have not been independently confirmed:
- About 05:42 MDT: Data sent Kevin the labelled contact sheet `contact/contact_sheet_seed0.png` (sha256 431e243c7af55921a93f8caa99ee4de33f324ea6f069bac99673b1354f2425e6).
- About 05:50 MDT: Data sent Kevin the /goodfire explainer video.

The seal and the prereg came first. RATING_PREREG.md (sha256 6c0262f62199b1282447039d91ad091b529710ab631d3ba6df4e97c451f1bbfc) was written at 05:33 MDT, and SEAL.md was written before 05:42.

## What he saw
The contact sheet shows seed 0 only, for six animals (dog, wolf, shark, rabbit, eagle, otter). It has four rows, each labelled with its condition: baseline, F7446, rand3199 and planted. From it a viewer can learn what the steered effect looks like, a wide-open mouth. The baseline and F7446 images of those six animals at seed 0 are the same source images used in 6 of the 100 real blind pairs. The blinded copies were re-encoded at 384 px, but the images are recognisable.

The explainer video's contents weren't inspected for this record. Assume it shows steered and baseline examples.

## The rule it triggers
RATING_PREREG.md line 23 says: "Raters should not be the person who built the packet. Anyone who has seen the contact sheets is noted as an unblinded-exposure caveat."

As written, the prereg only notes exposure as a caveat and doesn't exclude the rater. This deviation applies a stricter rule, decided before any data was seen: Kevin's ratings will NOT count as the preregistered blind test. The packet, key, seed and pass rule are unchanged.

## Pending decision (Kevin)
Kevin chooses one of these:
1. Use a new rater who hasn't seen the contact sheet, video or results. The prereg test runs on that rater as written.
2. Score Kevin's ratings, labelled NOT BLIND / EXPLORATORY. They would be reported with this deviation and could never be cited as the prereg result.
3. Drop the human rating arm. It would be reported as not run.

Until he decides, nothing is scored as the prereg test.