# Conclusion
---
The structural core reproduces: exhaustive enumeration found zero coverage failures across all 371 maximally populated masks for N=2–8, and the N/(N−1) bound gives exactly 1.50×, 1.33×, and 1.25×. The empirical audit is mixed: released A100 prefill data gives 1.3187× rather than an exact 1.33×, and the claimed H100 FP8 1.35–1.37× range is actually labeled RTX 4090. A local Qwen3-0.6B proxy strongly supports the moderate-sparsity motivation—6:8 answer NLL rose only 2.55%, while 2:4 rose 522.5%—but it is not the paper's fine-tuned Qwen3 setup. Runs used one local RTX 4000 Ada GPU for under ten minutes at $0 cloud cost; an A100 Job was rejected before scheduling because the Hugging Face token lacked job.write.
## Scope & cost
| | This reproduction | Full replication |
|---|---|---|
| Scope | Exhaustive theorem audit; released-table audit; Ada INT8 kernel proxy; Qwen3-0.6B/24-example GSM8K NLL proxy | Paper kernels on A100/H100 plus full Qwen2.5-7B inference and fine-tuned Qwen3 reasoning suite |
| Hardware | 1× RTX 4000 Ada 20 GB (GPU 0) | A100 INT8, H100 FP8, and multi-GPU model training/evaluation |
| Compute time | Under 10 minutes GPU time | Multi-hour kernel/inference sweep plus checkpoint fine-tuning |
| Cost | $0 cloud; rejected A100 submission incurred no charge | Not estimated; requires unavailable hardware and training recipe |
| Outcome | Claims 1–2 reproduced; Claim 5 qualitatively supported; Claims 3/6 qualified; Claim 4 hardware attribution corrected | Not attempted |
---
````html
````
---
**📦 Artifact** `slidesparse-icml33765-repro/reproduction-bundle:v1` · dataset
https://huggingface.co/buckets/vimarsh/Repro-SlideSparse-ICML-33765-artifacts#slidesparse-icml33765-repro/reproduction-bundle:v1
---
The reproduction bundle contains the fixed runner and configuration, exact snapshots of the released source CSVs, machine-readable result tables, full claim-by-claim notes, the print-ready poster/PDF, gate report, and Trackio logbook state. Use the artifact link above to open the Hugging Face Bucket after publication; download the bundle from that Bucket and run `bash run_reproduction.sh` from this directory. Model weights and replaceable Hugging Face caches are intentionally excluded.
