Phase 1 / technical rehearsal

Three-target rehearsal on candidate PDB chains

A pre-benchmark engineering run through the public Worker/Tunnel path. This validates plumbing and exposes timeout behavior; it is not a final Phase 1 result.

Short target

ok
9PFE:A predicted and scored.

Mid target

ok
13CZ:A predicted and scored.

Long target

524
13BB:A hit public synchronous route timeout.

Run Metadata

Field	Value
Run ID	`phase1-rehearsal-2026-05-13-simplefold-100m-3targets`
Model	`simplefold-100m`
Route	deployed Workers -> `science.eigenatlas.com` -> M3 Max server
Durable Object audit entries	29 after resume/re-score
Scope	Technical rehearsal only, not a biological benchmark claim.

Target Outcomes

Target	Bin	Outcome	pLDDT mean	Wall clock	lDDT	TM-score	Failure tag
`9PFE:A`	short	predicted and scored	95.9711	117.69 s	0.7395	0.0482	n/a
`13CZ:A`	mid	predicted and scored	71.5778	115.44 s	0.8532	0.9004	n/a
`13BB:A`	long	public synchronous prediction timed out	n/a	n/a	n/a	n/a	`orchestration_timeout`

Technical Findings

PDB:CHAIN labels work through the runner, PDB fetcher, structure predictor, and scorer.
The lDDT scorer now uses single-chain sequence alignment when real PDB residue numbering does not overlap predicted residue numbering.
HTTP 524 from the public synchronous route is now tagged as orchestration_timeout, not dependency_missing.
Long targets need async orchestration or direct local M3 execution before the final 24-target benchmark.

Artifacts