diff options
| -rw-r--r-- | MATCHING_RESULTS.md | 18 |
1 files changed, 16 insertions, 2 deletions
diff --git a/MATCHING_RESULTS.md b/MATCHING_RESULTS.md index ff33077..10c55f1 100644 --- a/MATCHING_RESULTS.md +++ b/MATCHING_RESULTS.md @@ -206,8 +206,22 @@ automorphisms is all there is to find. pure-quadratic descent scores 0.21, because the descent discards it immediately and wanders back into the decoy region. Pinned through the Frank-Wolfe iterations, the same descriptor scores 0.84. That distinction is the entire -result, and it is why amplification — which is also pure-quadratic — plateaued -at 0.508 no matter how large the pool grew. +result. + +It also settles the amplification line, which is pure-quadratic and therefore +inherits the same blind spot. Its accuracy does climb with pool size and does +not saturate — but it buys the climb with compute, and it is still short of the +ceiling at sixteen times the cost: + +| amplification pool | 40 | 160 | 640 | | fused unary+quadratic | +|---|---|---|---|---|---| +| accuracy | 0.167 | 0.508 | 0.644 | | **0.837** | +| descents | 720 | 3,840 | 7,680 | | ~40 Frank-Wolfe steps | +| wall clock | 12 min | 45 min | ~2 h | | **1.4 s** | + +Five thousand times faster and better. The compute was buying its way around a +missing term in the objective, which is the most expensive way to not fix +something. Negative energy gaps on rank8, synth-full and natural say something different and worth keeping: there the solver finds states **below** the truth, so what |