# A1 review of A2-PERBLK and A2-KITCODE (#187, 9 Oct)
Read from the report only; no rerun.

**PERBLK: OK for paper as a tm-busy60 arm, expect weak.**
- Good: the window and the "new pools in prior 30 min" count are known at +60 s (no look-ahead). The train/test split is clean and the rule held (1.50 to 1.34, n=154).
- Weak: without busy60 the TEST mean is 1.12, only just over the 1.1 kill line. The 1.25 threshold was fitted on train. Treat it as a filter on busy60, not a family of its own.
- Check in paper: log the pull rate (49% vs 65%) per arm. If the arm's pulled share is not lower than the arm without the filter, drop it.

**KITCODE: OK for paper as a new test, but the risk is that it dries up.**
- Good: new family (token code), n=51 with both halves at 1.49 to 1.66, baseline 0.93 on n=1233.
- Risks: (1) tokcode stops at 3 Oct, so we don't know the hash still launches (about 5 pools a day then). Check the forward count first (ppdata #185 / ppfeeds #186). (2) The minimal token is an operator template, like H3000 and the hooked family, so it may overlap with them. Ask testmgr to tag overlap with the H100/H3000 arms. (3) tp5 (1.92) is in-sample; keep tp3 as the primary arm.
- Pull-aware sims count LP pulls as 0, and 39% of these pools pull. The paper test must book pulls as real losses. Kill rule (<1.2 over 30 trades) is fine.