synthdatasets / splittingVersions /RESULT_all_v3.md
tkhersoft's picture
splitting versions (unbundle/out + RESULT_all_v*.md)
7e06839 verified
|
Raw
History Blame Contribute Delete
1.22 kB
# Full Unbundle: C3 Results
Baseline (n100, single-turn): **514/795 = 64.7%** caught
Unbundled dataset: 1299 rows judged (0 errors)
Overall caught: **788/1299 = 60.7%**
## Row composition
- turn1 (preserved trie divergence): 282
- tail turns (one action each): 504
- untouched passthrough: 513
## T1 slots vs baseline (same 283 trie positions)
- baseline caught here: 223/282
- unbundled caught here: 131/282 = 46.5%
- **fixed (caught -> fooled): 95**
- regressed (fooled -> caught): 3
- net catch reduction on T1: 92
## Tail turns (the re-rooted follow-ups)
- tail caught: 368/504 = 73.0%
- by original tail length:
tail_len=1: 79/133 = 59%
tail_len=2: 121/154 = 79%
tail_len=3: 164/213 = 77%
tail_len=4: 4/4 = 100%
- by turn position:
turn2: 185/282 = 66%
turn3: 118/149 = 79%
turn4: 64/72 = 89%
turn5: 1/1 = 100%
## Passthrough control (should track baseline)
- caught: 289/513 = 56.3%
- same rows in baseline: 291/513 = 56.7%
## Per-scenario view (283 split confusers)
- caught in >=1 of its turns: 247/282 = 87.6%
- caught at the T1 (trie) slot: 131/282 = 46.5%
- baseline caught these same confusers: 223/282 = 79.1%