2026-11-13
Task50 results ledger
In plain English: this table lists every retraining run from this round, one row per run, next to the model that's live (the v39-ref row). The main columns to read are "normal 327" (photos, out of 327, where every line is right) and "camberErr" (how far off the sail's depth measurement is; lower is better). The short story is that no new run clearly beats the live model, because identical reruns already differ by about 9 photos. The readable summary is at https://reports.sailscanapp.com/report50/.
Source: R2 scratch/t50/
| run | seed | ckpt | normal 332 | normal 327 | strict 327 | camberErr med / p90 | draftErr med / p90 | entryErr med / p90 | exitErr med / p90 | twistErr med / p90 | distMeanPx med / p90 | luffEndErr med / p90 | leechEndErr med / p90 | lengthErr med / p90 | wall_s | cost_usd_est |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| v39-ref | best | 266 | 264 | 108 | 0.30 / 1.30 | 2.40 / 5.80 | 2.00 / 8.80 | 1.20 / 4.50 | 0.20 / 1.20 | 1.46 / 4.10 | 0.79 / 2.64 | 0.80 / 2.88 | 0.01 / 0.05 | |||
| t50-T0-s0 | 0 | best | 274 | 271 | 116 | 0.30 / 1.20 | 2.30 / 5.80 | 2.00 / 8.80 | 1.10 / 4.50 | 0.20 / 1.30 | 1.45 / 4.07 | 0.76 / 2.65 | 0.85 / 3.14 | 0.01 / 0.05 | 6693.00 | 8.16 |
| t50-T0-s1 | 1 | best | 274 | 271 | 116 | 0.30 / 1.20 | 2.30 / 5.80 | 2.00 / 8.80 | 1.10 / 4.50 | 0.20 / 1.30 | 1.45 / 4.07 | 0.76 / 2.65 | 0.85 / 3.14 | 0.01 / 0.05 | 6846.00 | 8.35 |
| t50-T0-s2 | 2 | best | 264 | 262 | 109 | 0.30 / 1.20 | 2.30 / 5.80 | 2.10 / 9.20 | 1.20 / 4.50 | 0.20 / 1.20 | 1.41 / 4.08 | 0.78 / 2.64 | 0.81 / 3.08 | 0.01 / 0.05 | 6627.00 | 8.08 |
| t50-T0-s0 | 0 | last | 278 | 274 | 113 | 0.30 / 1.20 | 2.30 / 5.80 | 2.10 / 9.10 | 1.10 / 4.50 | 0.20 / 1.30 | 1.45 / 4.39 | 0.74 / 2.76 | 0.84 / 2.99 | 0.01 / 0.05 | 6693.00 | 8.16 |
| t50-T0-s1 | 1 | last | 278 | 274 | 113 | 0.30 / 1.20 | 2.30 / 5.80 | 2.10 / 9.10 | 1.10 / 4.50 | 0.20 / 1.30 | 1.45 / 4.39 | 0.74 / 2.76 | 0.84 / 2.99 | 0.01 / 0.05 | 6846.00 | 8.35 |
| t50-T0-s2 | 2 | last | 271 | 269 | 116 | 0.40 / 1.20 | 2.30 / 5.70 | 2.00 / 8.80 | 1.10 / 4.20 | 0.20 / 1.20 | 1.42 / 3.92 | 0.73 / 2.68 | 0.77 / 2.86 | 0.01 / 0.05 | 6627.00 | 8.08 |
| t50-T3-s0 | 0 | best | 271 | 270 | 108 | 0.30 / 1.20 | 2.30 / 5.70 | 2.00 / 9.00 | 1.00 / 4.50 | 0.20 / 1.20 | 1.41 / 4.00 | 0.77 / 2.48 | 0.75 / 2.93 | 0.01 / 0.04 | 6122.00 | 9.00 |
| t50-T3-s1 | 1 | best | 271 | 270 | 108 | 0.30 / 1.20 | 2.30 / 5.70 | 2.00 / 9.00 | 1.00 / 4.50 | 0.20 / 1.20 | 1.41 / 4.00 | 0.77 / 2.48 | 0.75 / 2.93 | 0.01 / 0.04 | 6103.00 | 8.97 |
| t50-T3-s2 | 2 | best | 267 | 264 | 107 | 0.40 / 1.20 | 2.30 / 5.70 | 2.10 / 8.70 | 1.10 / 4.30 | 0.20 / 1.20 | 1.42 / 3.82 | 0.73 / 2.42 | 0.79 / 2.80 | 0.01 / 0.04 | 6053.00 | 8.89 |
| t50-T3-s0 | 0 | last | 264 | 261 | 112 | 0.30 / 1.20 | 2.20 / 5.70 | 1.90 / 8.80 | 1.10 / 4.20 | 0.20 / 1.10 | 1.39 / 3.85 | 0.73 / 2.31 | 0.80 / 2.82 | 0.01 / 0.04 | 6122.00 | 9.00 |
| t50-T3-s1 | 1 | last | 264 | 261 | 112 | 0.30 / 1.20 | 2.20 / 5.70 | 1.90 / 8.80 | 1.10 / 4.20 | 0.20 / 1.10 | 1.39 / 3.85 | 0.73 / 2.31 | 0.80 / 2.82 | 0.01 / 0.04 | 6103.00 | 8.97 |
| t50-T3-s2 | 2 | last | 274 | 269 | 110 | 0.40 / 1.20 | 2.30 / 5.80 | 2.10 / 8.70 | 1.00 / 4.30 | 0.20 / 1.20 | 1.44 / 3.66 | 0.72 / 2.37 | 0.78 / 2.80 | 0.01 / 0.04 | 6053.00 | 8.89 |
| t50-T1-s0 | 0 | best | 273 | 271 | 89 | 0.40 / 1.30 | 2.30 / 5.70 | 2.10 / 8.70 | 1.20 / 4.50 | 0.20 / 1.30 | 1.54 / 4.35 | 0.78 / 2.61 | 0.82 / 3.12 | 0.01 / 0.05 | 1383.00 | 1.69 |
| t50-T1-s5 | 5 | best | 273 | 271 | 89 | 0.40 / 1.30 | 2.30 / 5.70 | 2.10 / 8.70 | 1.20 / 4.50 | 0.20 / 1.30 | 1.54 / 4.34 | 0.77 / 2.61 | 0.82 / 3.12 | 0.01 / 0.05 | 1456.00 | 1.78 |
| t50-T1-s6 | 6 | best | 273 | 271 | 89 | 0.40 / 1.30 | 2.30 / 5.70 | 2.10 / 8.70 | 1.20 / 4.50 | 0.20 / 1.30 | 1.54 / 4.34 | 0.77 / 2.61 | 0.82 / 3.12 | 0.01 / 0.05 | 1469.00 | 1.79 |
| t50-T1-s0 | 0 | last | 259 | 256 | 88 | 0.40 / 1.50 | 2.50 / 6.10 | 2.40 / 9.60 | 1.30 / 4.60 | 0.20 / 1.40 | 1.67 / 4.79 | 0.90 / 2.83 | 0.97 / 3.41 | 0.02 / 0.05 | 1383.00 | 1.69 |
| t50-T1-s5 | 5 | last | 258 | 256 | 83 | 0.50 / 1.50 | 2.60 / 6.20 | 2.40 / 9.70 | 1.30 / 4.80 | 0.30 / 1.50 | 1.72 / 4.86 | 0.91 / 2.81 | 0.95 / 3.41 | 0.02 / 0.05 | 1456.00 | 1.78 |
| t50-T1-s6 | 6 | last | 269 | 267 | 85 | 0.40 / 1.30 | 2.40 / 6.30 | 2.50 / 8.80 | 1.40 / 4.80 | 0.30 / 1.40 | 1.85 / 5.40 | 0.90 / 2.58 | 0.98 / 3.27 | 0.01 / 0.05 | 1469.00 | 1.79 |
| t50-T1T3-s0 | 0 | best | 256 | 254 | 100 | 0.30 / 1.30 | 2.40 / 5.90 | 2.10 / 9.20 | 1.20 / 4.50 | 0.20 / 1.20 | 1.46 / 4.41 | 0.84 / 2.68 | 0.83 / 3.06 | 0.01 / 0.05 | 1367.00 | 1.67 |
| t50-T1T3-s5 | 5 | best | 272 | 270 | 127 | 0.40 / 1.30 | 2.30 / 5.80 | 2.20 / 9.40 | 1.30 / 4.60 | 0.20 / 1.30 | 1.55 / 4.37 | 0.89 / 2.59 | 0.85 / 2.83 | 0.01 / 0.05 | 1648.00 | 2.01 |
| t50-T1T3-s6 | 6 | best | 256 | 254 | 100 | 0.30 / 1.30 | 2.40 / 5.90 | 2.10 / 9.20 | 1.20 / 4.50 | 0.20 / 1.20 | 1.46 / 4.41 | 0.84 / 2.68 | 0.83 / 3.06 | 0.01 / 0.05 | 1475.00 | 1.80 |
| t50-T1T3-s0 | 0 | last | 259 | 258 | 74 | 0.40 / 1.50 | 2.40 / 5.90 | 2.60 / 9.30 | 1.40 / 4.70 | 0.20 / 1.30 | 1.84 / 5.51 | 0.96 / 2.75 | 0.97 / 3.04 | 0.02 / 0.05 | 1367.00 | 1.67 |
| t50-T1T3-s5 | 5 | last | 262 | 259 | 72 | 0.40 / 1.40 | 2.40 / 6.30 | 2.40 / 9.50 | 1.30 / 4.80 | 0.30 / 1.30 | 1.76 / 4.56 | 0.91 / 2.75 | 0.94 / 3.07 | 0.01 / 0.05 | 1648.00 | 2.01 |
| t50-T1T3-s6 | 6 | last | 250 | 249 | 78 | 0.40 / 1.40 | 2.40 / 6.10 | 2.30 / 9.50 | 1.30 / 4.60 | 0.20 / 1.40 | 1.65 / 5.25 | 0.86 / 2.57 | 0.88 / 3.10 | 0.01 / 0.05 | 1475.00 | 1.80 |
1. Per-arm summary (distinct seeds, best checkpoint)
Mean of per-run medians. d = arm minus comparison (negative = lower error).
| arm | seeds | camberErr | luffEndErr | leechEndErr | normal-327 (per seed) | range | camber d vs T0 | luff d vs T0 | leech d vs T0 | camber d vs v39-ref | luff d vs v39-ref | leech d vs v39-ref |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T0 | 0,2 | 0.30 | 0.77 | 0.83 | 271, 262 | 262-271 | +0.00 | +0.00 | +0.00 | +0.00 | -0.02 | +0.03 |
| T3 | 0,2 | 0.35 | 0.75 | 0.77 | 270, 264 | 264-270 | +0.05 | -0.02 | -0.06 | +0.05 | -0.04 | -0.03 |
| T1 | 0,5,6 | 0.40 | 0.77 | 0.82 | 271, 271, 271 | 271-271 | +0.10 | +0.00 | -0.01 | +0.10 | -0.01 | +0.02 |
| T1T3 | 0,5,6 | 0.33 | 0.86 | 0.84 | 254, 270, 254 | 254-270 | +0.03 | +0.09 | +0.01 | +0.03 | +0.07 | +0.04 |
| v39-ref | - | 0.30 | 0.79 | 0.80 | 264 | 264 |
2. Winner by camber+end
Score = camberErr median + (|luffEndErr| median + |leechEndErr| median)/2, lower wins. Distinct best-ckpt runs (incl. v39-ref).
- t50-T3-s0: 1.06
- t50-T0-s2: 1.09
- v39-ref: 1.09
Full ranking: t50-T3-s0 1.06; t50-T0-s2 1.09; v39-ref 1.09; t50-T0-s0 1.11; t50-T1T3-s0 1.14; t50-T1T3-s6 1.14; t50-T3-s2 1.16; t50-T1-s5 1.20; t50-T1-s6 1.20; t50-T1-s0 1.20; t50-T1T3-s5 1.27
3. Footnotes
- t50-T1-s1..s4 and t50-T1T3-s1..s4 are replicas of their s0 (ultralytics 8.4.92 ignores args.seed in the dataloader; patched from s5 on). They are not listed as rows.
- t50-T0-s1 and t50-T3-s1 are also replicas of s0; kept as rows because they prove recipe determinism. They are excluded from the per-arm summary and ranking.
- All rows were scored by the same raw-model R1 harness (conf 0.45), NOT the production chain.
- v39-ref = shipped v39 weights run through this same harness (best only, no config).
- luff/leech end err are aggregated on |value| per the harness convention.
4. Total GPU cost
TOTAL cost_usd_est over 20 configs: $75.94 (includes the 8 replica runs t50-T1-s1..s4 and t50-T1T3-s1..s4, $13.75 of the total, which are not shown as rows.) Fixed overhead not captured in per-run wall: pods also spent setup time of roughly 25-40 min/pod across 8 pods.