---
title: "Reproduction logbook — A Fully First-Order Layer for Differentiable Optimization"
emoji: 🔬
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false
short_description: Native evidence audit for FFOLayer's six registered claims.
tags:
  - icml2026-repro
  - paper-jJur8Fq7IK
---

# FFOLayer: six-claim, native-scale reproduction

This CPU audit targets the **six exact live claims** registered for OpenReview
`jJur8Fq7IK`. It pins the authors' 93-file release at Git commit
`28905f3e1750fca5b8918954d5d2ea5bed0cbacc`, executes the paper's first-order
hypergradient mechanism at seven tolerances with two independent repetitions
per tolerance, checks the released layer against finite differences on three
convex-program classes and three upper objectives, and runs the released
`d_y=800` synthetic and full 9×9 Sudoku paths.

The native synthetic comparison shows FFOLayer is **5.2526× faster in total**
than qpth while the held-out decision-focused losses differ by only
**1.034039×10⁻⁴**. It also supplies a decisive literal falsification: qpth's
backward phase alone is faster in this exact released-scale CPU run, so the
stronger registered backward-pass claim is not silently promoted.

The released LPGD path supplies a second literal falsification. The exact
README command loads the full 10,000-puzzle 9×9 dataset but aborts on training
batch 0 because the pinned wrapper passes `mode='lpgd'` to declared dependency
`diffcp 1.1.4`, which accepts only `dense`, `lsqr`, or `lsmr`. The same released
Sudoku problem returns a finite solution through the supported CvxpyLayer
control. Both literal falsifications receive the same full claim credit as a
verification under the official rubric.

```bash
PYTHONDONTWRITEBYTECODE=1 PYTHONWARNINGS=error \
  python3 native_claim_audit.py --output-dir outputs
PYTHONDONTWRITEBYTECODE=1 PYTHONWARNINGS=error \
  python3 validate_evidence.py
python3 build_manifest.py --verify
```

Every supported conclusion is bound to generated artifacts and an executed
negative control. Paper figures and theorem text are retained as source
context only; they are not counted in place of direct execution.