Auditable local SimpleFold-3B inference on Apple Silicon

Working Claim

This paper is not proposing a new protein model. It asks whether SimpleFold-3B local inference is operationally viable on a 128 GB M3 Max under one direct CLI route and a frozen 8-target protocol, with enough public artifacts for review.

Current Abstract

Evidence Already Available

Phase 0 Results Snapshot

The latest public-tunnel run used deployed Workers and science.eigenatlas.com to reach the M3 Max server. It produced a clean Durable Object run with 3 predictions, 3 scores, recipe, verdict, and 41 audit entries.

Phase 1 Preparation Snapshot

Technical rehearsal results are separated from final benchmark claims. They currently show that inference route is itself an experimental variable: public synchronous HTTP, local synchronous HTTP, and direct CLI execution fail or succeed differently before model quality is even evaluated.

Repeat/cache-state control

The three 13CZ:A SimpleFold-3B CLI smokes are now summarized as a formal reporting rule. Same-target structure scores stayed close while confidence and runtime moved enough that one-off timing and pLDDT values must be labeled as preliminary unless repeated controls exist.

No-fixture Reader bridge

The runner now has a no---recipe-fixture SimpleFold Reader path through a manual command adapter. It produced reader_raw.txt, recipe.json, verdict.json, and card.json, extracting 70 CASP14 target labels, 13 quote entries, and the CASP14 SimpleFold-3B median LDDT value 0.709. The validator status is successful_with_caveat because this was a dry-run protocol-curation artifact, not inference.

Target review triage

The 24 candidate chains now have a reproducible triage layer: 8 provisional keeps, 14 human-review rows, and 2 replacement-recommended rows. This makes biological approval auditable, but it still does not freeze test_set_v1.csv.

First benchmark slices

Three provisional SimpleFold-3B targets now have real benchmark cards: 11BC:C, 13EE:A, and 13FT:A. All completed and scored. Median wall-clock time was 390.28 s, median peak RSS was 38.54 GB, median sequence-aligned lDDT was 0.8582, and median TM-score was 0.9212. The hard-bin 11BC:C slice reached 41.4060 GB peak RSS and aligned 63 reference residues, so alignment coverage must be reported alongside lDDT. This is early target-level evidence, not a final cross-model claim.

Phase 1A frozen protocol

The first paper is now scoped to an 8-target SimpleFold-3B benchmark: 13CZ:A, 13EE:A, 13FT:A, 11BC:C, 12OY:A, 9H8N:A, 9K9X:A, and 9K9Y:A. All eight targets have now been scored. The frozen manifest is public as CSV and JSON; results are public as summary CSV and summary JSON.

Phase 1A Results

The completed frozen set produced 8/8 scored targets. Across all targets, median wall-clock time was 395.96 s, median peak RSS was 34.08 GB, median sequence-aligned lDDT was 0.8597, and median TM-score was 0.9218. The four prospective targets run after freeze had median wall-clock time 726.63 s and median peak RSS 34.08 GB.

The companion SimpleFold-3B runtime card records the M3 Max hardware, macOS build, package versions, source commits, checkpoint SHA256 hashes, direct CLI route, scoring implementation, and public environment JSON. This is now part of the evidence chain for the first-paper claim.

The Phase 1A biological review packet marks all eight targets as acceptable for operational viability while preserving explicit caveats: 11BC:C is a coverage-limited scoring probe, 12OY:A is a membrane/complex stress case, the DcrB pair should not overweight one family, and 9K9Y:A should not be interpreted as a ligand-quality benchmark.

The selective repeat controls now add n=3 evidence for 11BC:C and 9K9Y:A. The hard/coverage target preserved exactly 63 aligned residues across three runs, while the long-bin target preserved high lDDT/TM-score with variable wall-clock time.

Draft Structure

Figures And Tables Planned

• Figure 1: Portable lab architecture: Workers, Durable Object, Tunnel, M3 Max server, Reader, Executor, Validator.
• Table 1: Phase 0 public-tunnel smoke results, including lDDT, TM-score, RMSD, wall-clock, and audit count.
• Table 2: Phase 1A frozen-target matrix: run status, memory, throughput, failure tag, aligned residues.
• Figure 2: Phase 1A target-level lDDT/TM-score panel for SimpleFold-3B.
• Figure 3: Throughput versus quality tradeoff on Apple Silicon.

What Cannot Be Claimed Yet

Immediate Writing Tasks

1. Tighten the Phase 0 validation narrative inside the Pandoc manuscript.
2. Apply the repeat/cache protocol to final timing and confidence claims.
3. Polish the completed Phase 1A figures and manuscript tables for the chosen venue.
4. Run the final citation audit and keep every numeric claim traceable to card/summary artifacts.
5. Configure one credential path for autonomous Reader extraction only if the paper claims autonomous paper reading.

Draft Links

• Current Pandoc-rendered workshop draft
• Current PDF render
• Submission-readiness checklist
• Phase 1 technical rehearsal packet
• Direct long-target recovery packet
• SimpleFold-3B short-smoke packet
• SimpleFold-3B mid CLI-smoke packet
• SimpleFold-3B telemetry-smoke packet
• SimpleFold-3B execution metadata-smoke packet
• SimpleFold-3B repeat/cache protocol
• No-fixture manual Reader packet
• SimpleFold-3B benchmark-slice results
• SimpleFold-3B Phase 1A results
• SimpleFold-3B model/runtime card
• SimpleFold-3B runtime environment JSON
• Phase 1A biological review packet
• Phase 1A biological review JSON
• Publication strategy and 200-hour sprint memo
• Target review triage
• Difficulty log
• Phase 1 protocol
• Phase 1A frozen protocol
• Phase 1A frozen CSV
• Phase 1A frozen JSON
• Phase 1 target candidates CSV
• Phase 1 model matrix CSV
• Abstract draft
• Latest public-tunnel card.json
• Current remaining gates