Bounded verification

Small checks. Specific claims.

These checks exercise simplified mathematical or systems mechanisms using Python’s standard library. They do not run the archived frameworks, train a model or certify the original implementations. Article-specific proposed tests remain proposals unless named here.

Executed 2026-09-19 · 14 checks passed.

  1. Patch boundary arithmetic

    An 8×4 patch on 32×31 pixels gives 32 SAME-padded patches versus 28 complete VALID patches.

  2. Attention head dimensions

    A 32-channel total split over four heads gives eight per head; four 32-channel heads give 128 projected channels.

  3. Topology survives softmax only with a mask

    Zero logits assign positive mass to nonedges; a separate exclusion mask removes it.

  4. Pooled statistics do not freeze old predictions

    The scalar ridge optimum changes 0.5→0 while old squared error rises 0.25→1.

  5. Floating addition is not bitwise order invariant

    Changing positive-term summation order changes a binary64 result.

  6. Nominal bits versus packed bytes

    Three four-bit values need two whole payload bytes; an int32 workspace needs twelve payload bytes.

  7. Embedded header and timestamp width

    A packed little-endian H/I/b/H header has nine bytes; uint32 microseconds wrap after about 71.58 minutes.

  8. Modulo pointer updates do not wrap a memcpy

    A twelve-byte packet at offset sixty in a 64-byte ring crosses the allocation even though the equality overflow check does not fire.

  9. Bucket identity differs from emission time

    A bar in bucket zero is emitted only when a later bucket arrives; absent buckets cannot be inferred as observations.

  10. Single-tone recurrence

    The independent sinusoid recurrence recovers a synthetic 20 kHz tone within 1e-7 Hz; this is not a noisy physiological test.

  11. Causal versus unmasked attention

    An unmasked earlier query mixes in a future value; a causal mask excludes it.

  12. Blockwise reverse chain rule

    A two-layer scalar adjoint sweep matches finite-difference parameter gradients within 1e-9; no PyTorch equivalence is claimed.

  13. Checkpoint input expansion

    Zero sensor columns preserve an actor affine map; relocating the critic action columns preserves its old input/action map.

  14. Integrity is distinct from accuracy

    A payload mutation changes its digest. This says nothing about whether the stored scientific answer is correct.

Back to ML Engineering