Fresh-seed BIB-CCS comparison
Open track · seeds 29, 31, 37, 41, 43 · 12 agents · horizon 40
Production defaultCell-aware adapterReplay tradeoffBIB specialist
Cell-aware replay: multi-seed confirmation
The general-action guardrail replicated; the external BIB gain did not
Evidence decision
Latest classification of what the experiments establish
General action preservation
Demonstrated
75.31% mean exact action; 75.00% unseen; 66% worst-cell floor across training seeds.
Robust external BIB gain
Not demonstrated
Mean gain was .0013, smaller than cross-seed variation; only seed 42 improved.
Production selection
Balanced checkpoint retained
The BIB specialist and replay adapters remain experimental checkpoints.
Checkpoint tradeoffs
| Checkpoint | Fresh BIB-CCS | V1.2 exact action | Status |
|---|---|---|---|
| Balanced base | .7741 | 75.25% | Production default |
| BIB action-only | .8479 | 44.04% | Experimental specialist |
| Uniform replay, w16 | .8033 | 73.87% | Experimental tradeoff |
| Cell-aware KD-256, 3-seed mean | .7753 | 75.31% | Preserved action; BIB gain not robust |
BIB-CCS v1 open-track artifacts updated 13 July 2026. Scores describe synthetic benchmark evidence and do not by themselves establish construct, ecological, or operational validation.