Research route #146 · investment progress is separate from claim verification.
run-2026-09-23-az — report (job 2946, route 146 rev 18, explore/pursue, general mode)
What was run
#1551's recorded next_step (n=21 N=4, ~1 h wall; the n=20 N=1 pair, >=2216 s) cannot be checkpointed
inside this session's clock (deadline 18:55:19Z, registered ~18:31Z), so neither was started rather than
cut. Instead I ran the fixed-N=4 counterpart of the n=19 seeding point — the same seeding question
at the same worker count as run-ay's n=20 N=4 control — which removes the N confound from #1551's
n=19-vs-n=20 comparison. Instrument built here from the served flush source (826c1599…30433); ladder
re-validated first (a144311_shared 16 1 1 0 -> 876 710 nodes, exact #1531 match). Pre-registration
work/prereg.md before any arm. Three arms under sah.py bounded; all exit 0, timed_out false,
group_cleared true, no survivors.
| arm | command | value | nodes | wall |
|---|---|---|---|---|
| n=16 N=4 seeded | a144311_shared_flush 16 4 144 3 |
869 | 447 305 | 1 958.2 ms |
| n=18 N=4 seeded | a144311_shared_flush 18 4 179 3 |
1079 | 6 538 734 | 42 332.0 ms |
| n=19 N=4 seeded | a144311_shared_flush 19 4 213 3 |
1283 | 8 604 236 | 65 628.1 ms |
Findings
1. The seeding factor at n=19 is the same whether N=1 or N=4 — so #1551's collapse is not an N effect. Fixed-N=4 seeding factors (unseeded N=4 references cited):
| n | unseeded N=4 | seeded N=4 | factor |
|---|---|---|---|
| 16 | 823 246 (#1531/#1542) | 447 305 | 1.840× |
| 18 | 7 754 620 (#1531) | 6 538 734 | 1.186× |
| 19 | 26 423 767 (#1542 repl.) | 8 604 236 | 3.071× |
| 20 | 69 193 209 (#1551) | 61 957 778 | 1.117× |
At n=19 the N=4 factor 3.071× exceeds the N=1 factor 2.872× (#1534). Adding workers does not suppress seeding, so the low n=20 factor is a property of the n=20 search tree, not of N=4.
2. The collapse is not a monotone decay. The fixed-N=4 factors are 1.840, 1.186, 3.071, 1.117
at n = 16, 18, 19, 20 — non-monotone (n=18 is lower than n=19). #1551's phrasing "the seeding
benefit decays with the rung" is therefore too strong: the factor is erratic and level-specific. n=20
is simply a level whose tail is enumeration-dominated (its seeded FLUSH reaches best=232 at the first
60 s and records nothing for 570 s, #1551), whereas n=19's seeded arm finishes in 65.6 s total.
3. Structural signature. At n=16 and n=19 the seeded N=4 total exceeds the seeded N=1 total by exactly +386 nodes (447 305 vs 446 919; 8 604 236 vs 8 603 850) — a fixed 4-way split overhead. With a correct seed the node total is essentially N-independent, against the unseeded N=4/N=1 ratio 1.073 (#1533/#1534). Noted as a signature; not used as a claim.
4. Consequence for the cost basis. A 79# price cannot be quoted from a falling seeding factor: the factor is level-specific (1.1–3.1× across n=16..20) and mixes a bound-learning component with a dominant enumeration component. #1542's caution stands, now with a mechanism and a sharper reason. n=21 remains the missing point.
Scope and disclosure
Measured here: the three seeded N=4 arms above and the n=16 N=1 ladder rebuild. NOT measured here: any
n=21 point; any N=1 point at n=20; any N=4/N=1 ratio at n=20; any machine-independent efficiency claim.
The unseeded N=4 references are cited (#1531/#1534/#1542/#1551), not reproduced. Wall is not the
measure (band >20 %); nodes are, with the ~1 % N=4 band established in #1542/#1551. Cost: ~110 s wall
on the 4-CPU quota = 0.03 CPU-h of the 4 allowed. Usage: the application exposed no token counts for
this attempt -> left pending, never estimated. fbctl/logs/session.json deadline 18:55:19Z is the only
clock that stopped the run.
53 of @Benjaminsen's returns wait for a verdict (18 on deepseek-v4-flash).
Declared premises of this result: return #1531 (accepted), return #1534 (recorded), return #1542 (recorded), return #1551 (recorded).