Per-pair LipSDP scorecard — unconstrained /128 net (mlpT) #
REDUCED CERTIFICATE MODEL — this file's concrete net is the 4×4-pooled 49-dim
MNIST family (width-8 hidden, /128–/256 rational weights), NOT the canonical
784→512→512→10 mlpVerified; chosen so every margin/norm/SOS check is exact rational
arithmetic in-kernel. Canonical surface: Proofs/MlpCanonical.lean.
The tighter-Lipschitz-constant pass over the SAME first-100 MNIST test
subset and SAME ε = 1/10 as LipschitzCertScorecard.lean: replacing the
global √2·∏‖Wᵢ‖ criterion by per-pair LipSDP certificates lifts the
count from 1/100 to 63/100 — no retraining, no new data, just a
less lossy constant on each pairwise logit gap. Each class pair carries
a PSD witness checked by linarith from the 8 LDLᵀ column squares
(hS* — the exact sum-of-squares certificate, d ≥ 0 weights found by
linarith itself), turned into the squared gap bound by pair_sq_bound;
each image evaluates its logits once (logit*_eval) and then needs only
9 rational margin checks against Lp·ε (certifiedSU<i>). PGD bracket:
69/100 — the cert ≤ TRUE ≤ PGD sandwich is nearly closed.
Theorem vs. measurement — read this before quoting a number. Soundness
lives in the ENGINE (LipschitzCertPairSDP.lean), proved once — kernel-
checking the 57th image buys nothing the 56th didn't. The count above is an
exact-rational MEASUREMENT over the first 100 images; the first 8 certifying
images (test-set order — an unbiased, reproducible rule) each carry a
CertifiedAt THEOREM, and scorecard_sdp_uncon below states only those.
Capping the images also caps the class pairs that need an LDLᵀ witness,
which is what this tier costs on every proof push.
Generated by scripts/lipschitz_cert_pair_sdp.py; weights/images/
certificates are DATA (SDP solved off-line, verified exactly here).
Equations
- One or more equations did not get rendered due to their size.
Instances For
MNIST test image #2 (digit 1), exact pixel sums /4080.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
MNIST test image #4 (digit 4), exact pixel sums /4080.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
MNIST test image #6 (digit 4), exact pixel sums /4080.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
MNIST test image #11 (digit 6), exact pixel sums /4080.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
LipSDP-certified witnesses on the unconstrained net: (subset index, image, class).
Equations
- One or more equations did not get rendered due to their size.
Instances For
The LipSDP scorecard — MEASURED 63/100, vs 1/100 under the
global √2·∏-norm criterion, same net, same ε, same images; the constant
was the bottleneck, not the network. The 8 below are the emitted
witnesses carrying CertifiedAt proofs, not that measurement.