Current status
Phase 1A complete. SimpleFold-3B has run and scored all 8 frozen targets on the M3 Max.
Last verified run
exec
SimpleFold-3B cards now preserve checkpoint, cache, progress, and RSS evidence.
Next gate
Submit to a workshop/artifact venue; source snapshot and Zenodo DOI are complete.
What is being built?
This project is a portable computational biology lab for a single researcher. The goal is to make protein-prediction benchmark runs inspectable: every paper identifier, recipe, target, prediction, score, verdict, and failure gets written down.
Why Cloudflare?
| Component | Role in the lab |
|---|---|
| Workers | Expose the orchestrator and MCP services as deployable HTTP services. |
| Durable Objects | Hold per-paper state, audit history, recipe, score, and verdict records. |
| R2 / KV | Cache PDFs, PDB structures, and future result artifacts. |
| Cloudflare Tunnel | Connect the M3 Max model server to the pipeline without directly exposing the workstation. |
Public artifacts
Phase 0 findings
The scientific lab note for the first engineering smoke run.
E2E runbook
Operational steps for deploying Workers, starting the M3 server, and running the pipeline.
Smoke result packet
Card, recipe, predictions, scores, and verdict from the deterministic smoke run.
Real SimpleFold smoke
Real SimpleFold-100M predictions and lDDT/TM-score records for three engineering targets.
Public tunnel smoke
The same real smoke routed through deployed Workers and science.eigenatlas.com.
Phase 1 protocol
Pre-benchmark scientific protocol: minimum publishable target set, model tiers, run order, and claim boundaries.
Phase 1A frozen protocol
Eight frozen SimpleFold-3B targets, per-target fixtures, decision rules, and public CSV/JSON manifests for the first publishable benchmark unit.
Phase 1A SimpleFold-3B results
Completed 8-target frozen benchmark: 8/8 scored, median wall clock 395.96 s, median peak RSS 34.08 GB, median sequence-aligned lDDT 0.8597.
SimpleFold-3B model/runtime card
Hardware, OS, source commits, package versions, checkpoint SHA256 hashes, scoring route, and limitations for the completed Phase 1A run.
Publication strategy
Decision memo for the 200-hour sprint: publish the portable reproducibility-lab paper first, with SimpleFold-3B Phase 1A as proof-of-work.
Reviewer packet
Claim-to-evidence map showing exactly what the first paper claims, what it does not claim, and which gates remain before submission.
Paper PDF
Headless-Chrome PDF render of the current Pandoc manuscript for workshop-style review packets.
Submission readiness
Final publication posture: completed source snapshot and DOI, remaining biological/venue gates, and recommended venue path.
Second committee response
Tracked response to the latest Weak Accept / Major Revision light critique: narrower title, claim table, target rationale, and scorer smoke test.
Submission statements
Draft contribution, AI-assistance, competing-interest, funding, ethics, data, and code statements for preprint/journal packaging.
Public release audit
Repository visibility and release checklist: the clean public source snapshot is released; Zenodo archive DOI is complete.
Repeat controls plan
Selective hard/coverage and long-bin repeat plan to strengthen runtime, confidence, and scoring interpretation before submission.
Repeat controls results
Completed n=3 controls for 11BC:C and 9K9Y:A, showing stable structural scores and variable wall-clock time.
Score recomputation
Reviewer path that recomputes lDDT/TM-score from saved predictions: 14/14 cards recomputed with zero score deltas.
Solo authorship policy
First-submission governance: Raúl as solo author/corresponding author by default, with external review acknowledged unless full authorship criteria are met.
Phase 1A biological review
Structured caveat packet: 8/8 targets acceptable for operational viability, with explicit coverage/context limits before final biological interpretation.
Phase 1 rehearsal
Three candidate targets through the public path. Short and mid ran; long target exposed a synchronous timeout gate.
Direct long run
Direct local prediction completed the long target, proving the 524 was a route-orchestration limit.
SimpleFold-3B smoke
Cached-checkpoint SimpleFold-3B completed one short candidate target through the direct local route.
SimpleFold-3B mid CLI
Direct SimpleFold CLI mode completed one mid-length candidate after HTTP timeout failure modes were isolated.
SimpleFold-3B telemetry
Repeated mid-target CLI smoke with progress logging and approximate peak RSS recorded in the result card.
SimpleFold-3B execution metadata
Repeated mid-target CLI smoke with machine-readable checkpoints, auxiliary caches, and progress samples.
SimpleFold-3B benchmark slices
First target-level benchmark evidence: 3/3 provisional targets scored, median wall clock 390.28 s, median peak RSS 38.54 GB, median sequence-aligned lDDT 0.8582.
Target review
RCSB-enriched review worksheet for deciding which candidate chains enter the final manifest.
Target review triage
Reproducible freeze recommendations: 8 provisional keeps, 14 human-review rows, and 2 replacement-recommended rows.
Run policy
Direct-local inference policy for final Phase 1 runs, with Workers retained for audit and publication.
Difficulty log
Operational failures, ambiguous outcomes, mitigations, and paper implications captured as negative evidence.
Repeat/cache protocol
Three-run SimpleFold-3B control defining how Phase 1 treats runtime, confidence, cache state, and execution metadata. Data: CSV / JSON.
No-fixture manual Reader
SimpleFold recipe extracted without --recipe-fixture: 70 CASP14 targets, 13 quote entries, and dry-run validator status successful_with_caveat.
Next gates
Manual Reader bridge complete; Phase 1A is run; remaining work is external biological review, citation audit, venue packaging, and optional Phase 1B expansion.
Workshop draft
Living manuscript draft with Phase 0 public-tunnel evidence, scientific boundaries, and Phase 1 benchmark gates.
Paper checklist
Submission-readiness checklist showing completed artifact release, completed Phase 1A evidence, and remaining external review/venue polish.
Boundary
This site intentionally distinguishes engineering evidence from scientific evidence. Phase 0 proves the public lab can run, score, and audit real predictions on a small engineering set. Phase 1A now provides a completed SimpleFold-3B Apple Silicon benchmark; Phase 1B remains the broader multi-model expansion.