The net-semantics tie, INSTANTIATED — CERTIFY for a concrete trained net #
GENERATED by scripts/smoothing_net_witness_gen.py. The smoothing chain's last
informal hypotheses — hC measurability and hp interiority for the abstract
classifier — are here DISCHARGED for mlpT, the trained /128-rationalized
49→8→10 pooled-MNIST MLP (LipschitzCertInstance.lean): measurability from the
logits' continuity, interiority from ONE strict-argmax witness per class
(argmaxNet_smoothProb_mem_Ioo). Each witness is a pooled test image whose
predicted class leads every other logit STRICTLY — checked in exact integer
arithmetic at generation and re-proved in-kernel below.
smoothing_cp_certified_mlpT is Cohen's CERTIFY guarantee for THIS network's
argmax with every semantic hypothesis discharged; the demo instantiates it at
the deployed scorecard scale (N = 10112, α = 1/1000). Honest scope: mlpT is
the Lipschitz-thread pooled net, not the 784-dim smoothing-driver checkpoint
(that tie = the same generator pass at full width), and the driver's float
forward vs this real-semantics net is the FloatBridge tier's story.
Class-0 witness: pooled test image #440 (label 0); exact logit margin over the runner-up is 836677782/66846720 ≈ 12.5164.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-1 witness: pooled test image #3231 (label 1); exact logit margin over the runner-up is 476292520/66846720 ≈ 7.1251.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-2 witness: pooled test image #1895 (label 2); exact logit margin over the runner-up is 929615405/66846720 ≈ 13.9067.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-3 witness: pooled test image #1845 (label 3); exact logit margin over the runner-up is 638124398/66846720 ≈ 9.5461.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-4 witness: pooled test image #8000 (label 4); exact logit margin over the runner-up is 691947893/66846720 ≈ 10.3513.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-5 witness: pooled test image #6908 (label 5); exact logit margin over the runner-up is 644875950/66846720 ≈ 9.6471.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-6 witness: pooled test image #7843 (label 6); exact logit margin over the runner-up is 867723980/66846720 ≈ 12.9808.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-7 witness: pooled test image #5297 (label 7); exact logit margin over the runner-up is 748210150/66846720 ≈ 11.1929.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-8 witness: pooled test image #7919 (label 8); exact logit margin over the runner-up is 593814126/66846720 ≈ 8.8832.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Class-9 witness: pooled test image #2478 (label 9); exact logit margin over the runner-up is 509675629/66846720 ≈ 7.6245.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The witness bundle: one strict-argmax point per class.
Equations
- One or more equations did not get rendered due to their size.
Instances For
CERTIFY for the CONCRETE net. smoothing_cp_certified_net with the
classifier = mlpT's argmax: measurability and hp interiority both
DISCHARGED (continuity + the ten witnesses above). With probability
≥ 1 − α over the N Gaussian samples: if the vote count for y at
x comes out k₀, every ‖δ‖ < σ·Φ⁻¹(q₀) leaves y the strict
argmax of the smoothed mlpT.
The deployed-protocol demo: at the scorecard's N = 10112, α = 1/1000,
a count of 10084 certifies radius σ·Φ⁻¹(0.9952) for the smoothed
mlpT — the per-image scorecard shape with NO abstract-classifier
hypotheses left (the tail check is one kernel evaluation). The count is
SmoothingCPScorecard.lean's MNIST-MLP image-1 entry, from a 784-dim driver
run; no Monte-Carlo count has been taken on mlpT itself, so the
hypothesis is the protocol's, not a measurement of this net.