Core claim
bounded
SimpleFold-3B local inference on one M3 Max, one CLI route, and one frozen 8-target set.
Proof-of-work
8/8
SimpleFold-3B completed every frozen Phase 1A target.
Not claimed
bounded
No new model, no CUDA comparison, no biological generality.
Manuscript Title
Auditable local SimpleFold-3B inference on Apple Silicon: an 8-target reproducibility benchmark
Claim Map
| Claim | Evidence | Limitation |
|---|---|---|
| SimpleFold-3B completed Phase 1A on Apple Silicon. | 8/8 frozen targets produced predictions and scores in summary CSV/JSON, cards, and public result page. | One model, one M3 Max, one direct CLI route. |
| Long-bin targets from 565-604 aa completed without OOM. | 9H8N:A, 9K9X:A, 9K9Y:A cards and summary table. | RSS is approximate sampled process RSS, not full unified-memory profiling. |
| Selected structural scores are repeat-stable. | n=3 controls for 13CZ:A, 11BC:C, and 9K9Y:A. | Five Phase 1A targets remain single-run observations. |
| lDDT/TM-score scoring is auditable. | Sequence-aligned scorer, aligned-residue counts, pseudocode, minimal scorer smoke test, and 14-card score recomputation. | External implementation comparison remains future work. |
| The atlas is auditable by reviewers. | Cards, runtime card, environment JSON, checkpoint hashes, public site, score recomputation packet, and clean source snapshot path. | Zenodo archive DOI is complete: https://doi.org/10.5281/zenodo.20275055. |
Non-Claims
- No new protein-structure model.
- No claim that Apple Silicon beats CUDA.
- No claim that AF3-class models generally work on Apple Silicon.
- No SimpleFold CASP14 headline reproduction.
- No biological generality for membrane proteins, transporters, enzymes, or ligand-bound structures.
- No ligand placement or affinity claim for
9K9Y:A.
Submission Gate Status
| Gate | Why it matters | Output |
|---|---|---|
| Selective repeat controls | Reduces concern that runtime/pLDDT claims are one-off artifacts. | Complete for 11BC:C and 9K9Y:A; 13CZ:A already has n=3. |
| External biological review | Confirms caveats and prevents biological overclaiming. | Reviewer note or explicit limitation. |
| Citation audit | Protects reference-level credibility. | Citation audit log. |
| Archive DOI | Makes the packet citable beyond the live site. | 10.5281/zenodo.20275055. |
| Public source release | The build repository remains private because of pre-release planning history; the public artifact should not expose that history. | Clean exported v0.1.2-phase1a snapshot is public and archived. |
| Venue formatting | Turns the working draft into a submission packet. | PDF, captions, reviewer README. |
Second Committee Response
The latest hardening pass retitled the manuscript, moved Claim/Evidence/Limitation into the paper, expanded the target-set rationale, documented the lDDT scorer, added a minimal scorer smoke test, and made the remaining public-release gates explicit.
Reading Order
- Read the abstract and conclusion in the Pandoc draft.
- Check the Phase 1A result page and summary JSON.
- Inspect the SimpleFold-3B runtime card.
- Inspect biological caveats before judging scope.
- Read the limitations section.
Source document: docs/38_reviewer_packet.md.