Counting V12 round 2 -- frozen before generation, 2026-11-08
Question: find nonisomorphic six-letter Boolean binary targets with identical
complete labelled five-position pattern laws (equivalently all loopless tests
on at most five letters). No claim to exhaust all six-letter relations.

Families, independent of outcomes:
A. All n=6 directed relations (loops allowed) of constant OUT-degree d=0,1,2,3.
   Enumerate all C(6,d)^6 labelled matrices; remove isomorphism by complete S6
   orbits. Relation complementation covers d=4,5,6; equality of complete laws
   is invariant under complement. Different degrees separated by edge count.
B. All 2^21 symmetric binary relations on six vertices, loops allowed.
Duplicates across families are removed exactly by canonical representative.

Expected labelled cases A: 1 + 46656 + 11390625 + 64000000 = 75437282.
Expected labelled cases B: 2097152. These counts overlap, not distinct types.

CPU: orbit generation and exact isomorphism deduplication. A hash table only
indexes full 36-bit keys; comparisons check exact keys, no hash certification.
GPU: sorted integer pattern codes for every ordered assignment in {0..5}^k.
Each complete sorted vector is an exact law (multiplicities retained).
Group at k=2,3,4,5; singleton groups need no larger tests. A Python bytes dict
compares full vectors exactly; hashing does not replace equality.
Warmup and benchmark GPU vs CPU at k=5 on up to 256 fixed input graphs, 3 runs.
Independent numpy generation must match all benchmark vectors. Use GPU only
if measured wall time including transfer is less than CPU on this benchmark.
Verify all surviving five-position collision pairs independently on CPU,
check nonisomorphism by a separate permutation implementation, and seek exact
witnesses at k=6,7,8 as needed. If still unseparated, report only a lower bound.
Audit 64 fixed-seed pairs separated at k=5 (or all if fewer), including their
prior equality and actual distinguishing hom counts.
Controls: known n=4 seed collision at k=4 but not k=5, padded to n=6;
relabelled graph must have exactly the same laws. No floating-point decisions.

Stop after these finite families. No model inference/training, no automatic
renewal, deletion or shutdown. Report absent collisions only within families.
This is post-V12 exploration; preserve released V12 manuscript and TMLR files.
