# Pre-registration — job #2966 (route 146 rev 21), the 386 constant at a fifth level Written **before any n=20 arm was launched** (2026-09-23T23:14Z). This is the `next_step` filed in return #1561, adopted by the server as this assignment. Instrument: the served sources, already fetched and hash-verified in this run: `a144311_shared_flush.cpp` `826c1599…30433`, `a144311_shared.cpp` `d768a76d…09b3b`. Seed `mb = (a(20)-5)/6 = (1397-5)/6 = 232` (published a(20) = 1397, used only as the seed). ## Arms | arm | command | expected | |---|---|---| | n=20 N=1 seeded | `a144311_shared_flush 20 1 232 0` | value 1397, nodes near #1551's 61 957 778 | | n=20 N=4 seeded | `a144311_shared_flush 20 4 232 3` | value 1397, nodes near #1551's 61 957 778 | ## Hypothesis and falsifier (written before the runs) **H1.** The fixed +386-node N=4 minus N=1 split cost holds at n=20 as well, giving a fifth level (16, 17, 18, 19, 20) and the largest node total of the set. - **F1:** N=4 nodes minus N=1 nodes is **not** exactly 386 at n=20. Then the constant is level-specific to n=16-19, the seeded node total must always be quoted with its N, and the four-level regularity is recorded as bounded to those levels. - **F2:** either arm's value differs from 1397, or the N=1 total departs from #1551's recorded 61 957 778 by more than the ~1 % band. Then the instrument copy or this machine differs from the one that produced the route's n=20 point. ## Budget and stop rule Session clock: the session ends 2026-09-24T00:12:38Z, so at most ~55 minutes of wall time remain; the assignment's own hold runs to 01:12:22Z. Host load was 50-70 on 8 CPUs at 22:52Z, so wall time is not a reliable predictor. Each arm is cut at the session boundary; an arm cut before its final `SHARED` line yields **no node count** (the FLUSH node field is the known #1553 defect and reads 0), and is reported as a scoped obstacle, never as a measurement. Node totals, not wall, are the cross-run measure.