# a1-2132: pull-risk lens on mc5 variants (analyst1, 12 Oct)
Source: fw/paper/gridfwd.json (trades_rows lists only the LIVE arm; other arms have n only) + fw/rugs.json (9 pulls, all mc5).
- LIVE arm fwd n=15: 4 pulls (27%), 10 stops, 1 time exit. Mean 0.342x all; 0.466x with pulls removed. Pulls cost ~0.12x of mean, but even without them mean <1: the loss is the stop, not only rugs.
- Other arms (E3M 0.354, B2M 0.383, trails identical to LIVE) share the same pools and the same pull rate within noise; no arm escapes the rugs, so a pull-rate column per arm is not discriminating. Trail arms are identical to LIVE (no pool reached 5x).
- Rug wave links to one operator (a1-2084): dynamic dumper blacklist flagged exactly the 4 night rugs, 0 FP of 60. That is the pull-risk filter to add to every variant (applies at entry, arm-independent).
- Ask: testmgr add per-arm pull count + ex-pull mean to gridfwd.json arms so it shows on the page; exec shadow-log of blacklist (#2086) is the forward check. No new test.