# AMENDMENT 1 to PREREGISTERED.md (sha256 da52034242ff6b697818e6ef3713c551dc85e4b7cc626ed10beec48c76dd1478)

Written 18 Nov 2026, before any GPU pod for this run existed and before any AlpaJudge or Edge output on our data. No
pod has been created, so no judge score, Edge output or person's mark exists. Cause: the adversary's pre-run memo
(`evidence/adversary-prerun.md`, 1 P0 and 8 P1 findings) and the page lane's note that the example-pair rule was not
written down. Everything in PREREGISTERED.md stands except where this file changes it.

## Data

jobs.json sha256: `1353b76b37124d4636c9859cd238cf041d9f8f7403b57d553f8d2ac3ac40983b`

The new job set holds the same 149 test windows, the same inputs and the same 1,558 original pairs. It adds:
- per window: the scene, the future Lateral and Longitudinal labels and the image half that holds the car;
- the five extra probes (below);
- the rename of the real pairs' `original_is_correct_side` to `original_is_traffic_side`.

The pre-registered file is kept as `evidence/jobs/jobs-prereg-6d212670.json`.

Visibility marks: `evidence/visibility/marks.json`. They were made blind by an independent model check (a Claude subagent, not a person) on the t0 front-wide frame of each test window, before any data. sha256: see VISIBILITY line below.

VISIBILITY sha256: `aa3acd18090dc6fef1c56e030be9e5dfc6dd6cd866ee1f8720774556fda2d0e6` (marks.json; key.json `8cbe8d24c48a99bad91f3404cddc14d1ca44625c24a62ce8e0065879f64ad847`). Marks: UK and mirrored together 126 visible, 17 not, 6 can't tell.

## Changes

**A1 (P0). The bootstrap cluster is the scene, not the episode.**
- Why: episodes on one arm with one world-model seed share their opening and lead-in. They differ by 1.6 grey levels at
  takeover (frame 60) and diverge only as the plans for different exits pull apart (median 8.5 grey levels over the 130
  same-scene, same-time window pairs; 52 of them under 4).
- The cluster is now world-arm-seed, with exits and the centreline driver merged. That gives 9 scenes in all:
  - UK, 5 scenes: E seed 0 (28 windows), E seed 1 (22), S seed 0 (10), S seed 1 (14), W seed 0 (5);
  - mirrored, 4 scenes: E seed 0 (31), N seed 0 (18), S seed 0 (9), W seed 0 (12).
- The bar's bootstrap resamples scenes (10,000 resamples, seed 20261118).
- Reported, no bar:
  - the number of distinct moments (scene and time), UK 32 and mirrored 27;
  - the share of moments where the correct side wins a majority of that moment's pairs.
- The page says the sample is 5 UK and 4 mirrored scenes.

**A2 (P1). The Lateral line.**
- Reported, no bar: the primary win rate in three strata of the future Lateral label (turn word equals the correct side,
  turn word equals the wrong side, Go Straight).
- If a map fails in the normal arm but the neutral arm meets the bar on it, the page adds: "On {map}, with the action
  text naming a turn, the judge followed the turn word; with the turn removed it chose the correct side in k of n."

**A3 (P1). Action floor.**
- Reported, no bar:
  - the win rate on each of the five sub-scores;
  - the primary rate split by future Longitudinal (Stop or deceleration against the rest);
  - the median `action_matches_gt` of the correct member.
- Page caveat when the UK windows are mostly Maintain Speed: "In most UK windows the car kept its speed, so a sentence
  that says give way does not match the action, and NVIDIA's rubric then gives the action score little weight."

**A4 (P1). Visibility.** The visibility marks above. Reported, no bar:
- the primary rate on the windows marked visible;
- the rate split by upstream gap (under 90 degrees, 90 and over).

**A5 (P1). Wording.**
- A pass shows that the judge notices when the side word contradicts the traffic in the pictures. It does not show that
  it knows the British rule: in 145 of 149 windows the car with priority is also in that half of the image.
- The verdict reading for two passes becomes "notices when the side word contradicts the traffic in the pictures".
- Probes 4 and 5 below test the rule as text.

**A6 (P1). Two more arms**, both members of every primary pair:
- "blind": all six frames grey and both Lateral lines "Go Straight". The rule probes 4 and 5 are also scored here.
- "flip": all six frames mirrored left to right and both Lateral lines "Go Straight". This arm is scored for the
  unflipped map's correct side.

"From the pictures" is said only if all four hold:
1. the normal arm passes both maps;
2. the neutral arm meets 80 % with a Wilson lower bound above 50 % on both maps;
3. the blind arm does not meet that on both maps;
4. in the flip arm, the Wilson upper bound of the correct-side rate is below 50 % on both maps.

Otherwise the page says "from its inputs". The grey arm stays as pre-registered, reported only.

**A7 (P1). One-map results.**
- A pass on one map and a fail on the other reads "passes on the UK map only" or "passes on the mirrored map only".
- "Prefers "right" regardless" is used only if, in the blind arm on both maps, the "right" member scores higher in more
  than half of the pairs and the mean margin, right minus left, is above zero. "Prefers "left" regardless" is the same
  test the other way round.

**A8 (P1). Ties.**
- The tie rule and the 80 % bar are unchanged.
- Reported: ties per map and per arm, and a two-sided sign test on the untied pairs.
- If the median absolute primary margin is 0, the control reads "No verdict: the score did not change on most side
  swaps (k of n tied)." This replaces the "moved as much" sentence, which would be false in that case.

**A9 (P1). NVIDIA's score contract.** Page caveat, verbatim: "NVIDIA's scoring contract for this checkpoint is named
`vlm_rubric_c24f80fa_v1_no_direction_quality_diagnostic`. We do not know exactly what that excludes; we test whether the
score moves with the side word, not whether NVIDIA trained it to."

**A10 (P2). Code against text.**
- The family and the filler are chosen by sha256 of the window id joined with "family" (or "filler") by a NUL byte, as
  `harness/pairs.py` does. PREREGISTERED.md's shorter wording is corrected here.
- If either map fails, the primary is "not met", and the other map's status is stated.
- Missing scores are counted and stated on the page.

**A11 (P2). Page caveat:** "The mirrored roundabout keeps the British scenery and number plates, so it is not a real
country. A judge that trusts the scenery over the direction of traffic would lean towards 'right' on the mirrored map."

**A12 (P2).**
- The phantom probe is reported without the uk-W windows, where a generated person with a bicycle stands at the kerb.
- The order probe is reported as its median absolute margin, not a win rate.
- Real pairs report "traffic side higher" only for "from the right/left" phrases.

## Extra probes (normal arm, every test window, reported with no bar)

1. position: "A car is circulating on the {right|left} side of the picture." The expected word is the image half that
   holds the window's best car.
2. drive_side: "We are driving on the {left|right}-hand side of the road." The expected word is left in the UK and right
   when mirrored.
3. turn_join: "Give way, then turn {left|right} to join the roundabout." The expected word is left in the UK and right
   when mirrored.
4. rule_gb: "In Britain, traffic on a roundabout comes from the {right|left}." Right on both maps; also scored in the
   blind arm.
5. rule_fr: "In France, traffic on a roundabout comes from the {left|right}." Left on both maps; also scored in the
   blind arm.

## Example pair (page and film)

The UK primary pair with the median margin. With an even count, the lower median is used. Ties are broken by the lowest
sha256 of the window id.

## Money

About 750 more judge calls, inside the full-run stage cap of USD 12. The run cap stays USD 25.
