Level C

Reproduction harness

Which geometric capability produces the gain — not which paper won a table. Twelve cells, one protocol. No number appears here until its ancestry hashes close.

One implementation, six factor switches, two endpoint modes. atlas-kg-v1 (synthetic, 48 entities, seed 7) is bound. FB15k-237 is not. Do not read 0.402 into these cells.

12/12 atlas-kg-v1 cells bound · FB15k-237 still unbound

Synthetic 48-entity atlas-kg-v1. Not FB15k-237. Does not reproduce FlorE 0.402. One seed. M5 geodesic displacement is not implemented in the trainer (falls back to M3 score).

dataset 1deb03e2fa7a · seed 7 · train 207 · test 46

Factorhead-onlydual endpoints

M0

R

holds

ε 0

MRR 0.070

holds

ε 0

MRR 0.087

M1

R+B

holds

ε 0

MRR 0.102

holds

ε 0

MRR 0.123

M2

R+ℤ₂

holds

ε 0

MRR 0.076

holds

ε 0

MRR 0.046

M3

R+B+ℤ₂

holds

ε 0

MRR 0.066

holds

ε 0

MRR 0.089

M4

M3 + score (B)

holds

ε 0

MRR 0.074

holds

ε 0

MRR 0.072

M5

M3 + offset (A)

holds

ε 0

MRR 0.066

holds

ε 0

MRR 0.089

QD_σ vs Q vs QB

Same data, seed, trainer. One run. Not a general result. A is the only non-isometry.

  • A-QDQD_σMRR 0.127 · ε 8.54e-1 · |σ| 1.116
  • B-QQMRR 0.076 · ε 1.33e-16 · |σ| 0.950
  • C-QBQBMRR 0.066 · ε 6.36e-16 · |σ| 0.950

required hashes: dataset · splits · protocol · metric · commit · lockfile · config · seed · checkpoint

Reproduce with npm run experiment. Evidence levels.