# V3 Formal 70-Cell Detailed Comparison

Source: `runs/real-v3-extra-hard10-gate-score-70-20260709-03/final-scores.csv`.

Gate Score is comparable across L1/L2/L3. Native run status and posthoc score are separate signals.

## By Run And Replicate

`mean +/- sigma` uses population standard deviation across the ten cases in each replicate.

| Level / replicate | Run success | Gate mean +/- sigma | Gate >= 4 | Structure | Semantic visual | Geometry | Abs. flagged | Hidden | Large media | Holistic | Mean duration |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| L1 r1 | 10/10 | 2.254 +/- 0.400 | 0/10 | 3.124 | 3.876 | 0.292 | 8/10 | 2/10 | 4/10 | 3.000 | 1.2m |
| L2 r1 | 10/10 | 2.345 +/- 0.418 | 0/10 | 3.310 | 3.688 | 0.252 | 8/10 | 0/10 | 3/10 | 3.200 | 3.0m |
| L2 r2 | 10/10 | 2.116 +/- 0.382 | 0/10 | 3.123 | 3.241 | 0.167 | 9/10 | 2/10 | 2/10 | 2.900 | 3.9m |
| L2 r3 | 10/10 | 2.282 +/- 0.386 | 0/10 | 3.061 | 3.680 | 0.289 | 10/10 | 2/10 | 3/10 | 3.100 | 3.5m |
| L3 r1 | 10/10 | 3.015 +/- 0.657 | 1/10 | 3.625 | 4.314 | 1.302 | 3/10 | 0/10 | 4/10 | 2.700 | 111.7m |
| L3 r2 | 10/10 | 2.912 +/- 0.440 | 0/10 | 3.560 | 4.502 | 0.878 | 4/10 | 0/10 | 5/10 | 2.700 | 86.4m |
| L3 r3 | 10/10 | 2.876 +/- 0.413 | 0/10 | 3.310 | 4.502 | 0.908 | 5/10 | 0/10 | 3/10 | 2.900 | 118.3m |

## By Case

L2/L3 values aggregate three replicates. Sigma is population standard deviation across those replicates.

| Case | L1 Gate | L2 Gate mean +/- sigma | L3 Gate mean +/- sigma | L3 - L2 | L3 run success | L3 Gate >= 4 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| case-h1 | 1.920 | 2.337 +/- 0.174 | 2.473 +/- 0.019 | +0.137 | 3/3 | 0/3 |
| case-h2 | 1.620 | 2.250 +/- 0.393 | 2.660 +/- 0.201 | +0.410 | 3/3 | 0/3 |
| case-h3 | 2.120 | 1.940 +/- 0.233 | 3.310 +/- 0.086 | +1.370 | 3/3 | 0/3 |
| case-h4 | 2.440 | 2.147 +/- 0.107 | 2.757 +/- 0.208 | +0.610 | 3/3 | 0/3 |
| case-h5 | 2.430 | 2.750 +/- 0.185 | 2.717 +/- 0.128 | -0.033 | 3/3 | 0/3 |
| case-h6 | 1.810 | 1.440 +/- 0.049 | 2.457 +/- 0.339 | +1.017 | 3/3 | 0/3 |
| case-h7 | 2.120 | 2.353 +/- 0.308 | 3.147 +/- 0.203 | +0.793 | 3/3 | 0/3 |
| case-h8 | 2.370 | 2.360 +/- 0.099 | 3.170 +/- 0.778 | +0.810 | 3/3 | 1/3 |
| case-h9 | 3.020 | 2.630 +/- 0.121 | 3.873 +/- 0.066 | +1.243 | 3/3 | 0/3 |
| case-h10 | 2.690 | 2.270 +/- 0.255 | 2.780 +/- 0.267 | +0.510 | 3/3 | 0/3 |

## Raw 10 x 7 Cell Matrix

Format: `Gate Score (passed/total)`. `*` means native run status is not success but a scoreable artifact exists.

| Case | L1 r1 | L2 r1 | L2 r2 | L2 r3 | L3 r1 | L3 r2 | L3 r3 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| case-h1 | 1.92 (13/112) | 2.09 (16/112) | 2.46 (23/112) | 2.46 (23/112) | 2.46 (18/112) | 2.50 (17/112) | 2.46 (18/112) |
| case-h2 | 1.62 (10/66) | 2.29 (14/66) | 1.75 (10/66) | 2.71 (24/66) | 2.94 (22/66) | 2.56 (17/66) | 2.48 (15/66) |
| case-h3 | 2.12 (13/37) | 1.88 (11/37) | 1.69 (10/37) | 2.25 (13/37) | 3.21 (21/37) | 3.42 (22/37) | 3.30 (21/37) |
| case-h4 | 2.44 (15/49) | 2.25 (14/49) | 2.19 (13/49) | 2.00 (12/49) | 2.76 (18/49) | 2.50 (15/49) | 3.01 (20/49) |
| case-h5 | 2.43 (16/52) | 3.00 (18/52) | 2.56 (15/52) | 2.69 (16/52) | 2.55 (17/52) | 2.74 (17/52) | 2.86 (19/52) |
| case-h6 | 1.81 (11/34) | 1.50 (9/34) | 1.38 (9/34) | 1.44 (9/34) | 2.00 (12/34) | 2.56 (15/34) | 2.81 (17/34) |
| case-h7 | 2.12 (12/29) | 2.68 (15/29) | 2.44 (14/29) | 1.94 (11/29) | 3.30 (19/29) | 3.28 (19/29) | 2.86 (16/29) |
| case-h8 | 2.37 (13/34) | 2.50 (13/34) | 2.29 (12/34) | 2.29 (12/34) | 4.27 (28/34) | 2.62 (14/34) | 2.62 (14/34) |
| case-h9 | 3.02 (30/84) | 2.70 (28/84) | 2.46 (25/84) | 2.73 (29/84) | 3.96 (58/84) | 3.80 (58/84) | 3.86 (66/84) |
| case-h10 | 2.69 (17/49) | 2.56 (15/49) | 1.94 (12/49) | 2.31 (13/49) | 2.70 (19/49) | 3.14 (22/49) | 2.50 (15/49) |

## Key Readouts

- L2 is nearly tied with L1: Gate delta -0.006.
- L3 leads L2 by +0.687; semantic visual delta is +0.903, and computed text-geometry delta is +0.793.
- Gate >= 4: L1 0/10, L2 0/30, L3 1/30.
- Mean cell duration is 1.2m for L1, 3.5m for L2, and 105.4m for L3.
- Geometry is no longer scaffold-saturated: every candidate is measured with Playwright computed DOM boxes against one shared grounding bank per case.
- Gate Score v2 is not numerically interchangeable with the earlier scaffold score; historical values are archived under the run's evaluation-archives directory.