1.085M real commands through the full gate: 0 real recall misses

[BGMLAI]

Today the full 6-stage gate (the real ActionPipeline, not the flat check_action path) was run against 1,085,159 unique real agent commands streamed from five public trajectory datasets (Nemotron, SWE-Zero, SWE-Hero, Kwai SWE-smith-mini, nebius), plus a deterministic sweep of all 43 known danger classes.

Result: 0 real recall misses.

• 43/43 known danger classes neutralized (30 block, 13 warn), 0 false-blocks on 13 benign twins. Reproduce in seconds: pip install gate-cat && python scripts/recall_danger_axis.py.
• 1.085M real commands: the independent 43-class catalog flagged 447 as dangerous; the gate neutralized 443 and allowed 4. Every allow was adjudicated — all 4 are the same 2 disposable-artifact cleanups (rm -f test*.csv *.pyc && rm -rf __pycache__ in a repo workspace). The delete-analyzer allowed them as proven-disposable, and the identical shape with a real target blocks 5/5 (rm -rf * on real files, /etc/*, *.db, ~/.ssh/*). So it's a catalog over-match, not a gate blind spot.

Full method + artifact: RECALL.md, results/million_recall_2026-07-08.json. Every number is in FACTS.md (F1a/F1b).

Honest caveat kept front and center: this measures detection of known-dangerous shapes. The gate is certain only about what it blocks; an unmatched action is unchecked, not safe. That's exactly why axis 2 exists — to hunt for shapes we didn't enumerate — and why the bypass suite prints its own known gaps. If you find a shape that slips through, that's the veto story worth more than a star.