sharky-0.5B-v3b — archived phase9 ablation (REJECTED per pre-registered gate)

Delta fine-tune from sharky-0.5B-v3 (+15k CIC-IDS2017 benign states, replay, lr 2e-5, 4,126 steps / 66,000 records). Not promoted — its purpose (repair S1 benign recall) moved the metric the wrong way: 0.36 → 0.24.

What this ablation established

• More same-corpus benign data at low LR does not fix the CIC-IDS2017 benign-web shortcut — the boundary is representational, not a data-quantity problem.
• kev's replay recipe preserved everything else (attacks improved: malicious recall 0.843→0.934, DDoS 0.230→0.676, locked 0.9836, S2b 0.9789, S4 0.9915) — perfect for adding domains, wrong tool for redrawing a boundary.

Suite	v3	v3b
S1 benign recall	0.359	0.240 ✗
S1 malicious recall	0.843	0.934
S2b unseen-benign recall	0.9718	0.9789 ✓
S4 Neris recall	0.9899	0.9899 ✓
Locked	0.9808	0.9836 ✓

Full record: phase9/REJECTED_v3b.md · gate_v3b.json · next levers in phase9/README.md and CONTRIBUTING.md.