What remains after Phase 1A benchmark runs

Public Tunnel

The public tunnel path is now active. The final route is science.eigenatlas.com -> 127.0.0.1:8001, because port 8000 is occupied by omlx serve on this workstation.

See the completed public tunnel smoke result packet.

No-Fixture Reader

The first autonomous no-fixture Reader attempt used a legacy child-process adapter and failed before extraction because provider subscription access was disabled. This is recorded as a credential/configuration blocker, not a paper-reading failure. The current benchmark route uses the generic command adapter with a narrow manual SimpleFold reader, preserving the same Recipe schema and audit artifacts without using --recipe-fixture.

PIPELINE FAILED: reader failed: external child-process reader exited 1:
provider subscription access disabled

The completed manual bridge is public at phase1-no-fixture-manual-reader. It extracted 70 CASP14 target labels, 13 quote entries, and CASP14 SimpleFold-3B median LDDT 0.709. The validator status is successful_with_caveat because no inference was executed.

To upgrade this from manual bridge to autonomous Reader-agent evidence, configure READER_ADAPTER=command and READER_COMMAND='node scripts/codex-agent.mjs' through a local Codex/OpenAI credential path.

READER_ADAPTER=command \
READER_COMMAND='node scripts/codex-agent.mjs' \
CODEX_AGENT_FLAGS='--full-auto' \
OPENAI_API_KEY=... \
bun run pilot --doi 10.48550/arXiv.2509.18480 \
  --url https://arxiv.org/pdf/2509.18480 \
  --model simplefold-100m

Credential check on 2026-05-13 and 2026-05-14: no usable local pipeline API credential was present for child-process Codex, and Codex CLI reported Missing OpenAI API key. The adapter is implemented; autonomous extraction waits on one credential path.

Target Review

The Phase 1 candidate target manifest exists and passed technical fetch checks for all 24 rows. The target review triage identifies 8 provisional keeps, 14 human-review rows, and 2 replacement-recommended rows. The 8 provisional keeps are now frozen as Phase 1A, and the biological review packet marks all 8 as acceptable for operational viability while preserving context/coverage caveats before formal sign-off.

Resume

The runner now supports --resume. It loaded 3 predictions and 3 scores from the real smoke packet, skipped inference, and regenerated the final card.

Long-Run Telemetry

SimpleFold-3B now has short, mid, telemetry, and execution-metadata technical smokes. The repeat/cache protocol defines how timing, pLDDT, cache state, and execution metadata should be interpreted before the full benchmark. The current gaps are documented in the difficulty log.

Benchmark Slices

The first SimpleFold-3B benchmark slices completed on three provisional targets, including one hard-bin complex-context target. They provide real target-level evidence and exposed two scoring-instrumentation requirements: lDDT must be sequence-aligned when predictions and PDB references use different residue numbering, and aligned-residue coverage must be reported for complex-derived references.

Completed Phase 1A

The completed Phase 1A SimpleFold-3B result packet now reports 8/8 scored frozen targets. The SimpleFold-3B model/runtime card records hardware, package versions, source commits, checkpoint hashes, and route boundaries. The next scientific gates are external biological target review, citation audit, solo-author submission statements, and venue-ready packaging.

Decision

Phase 1A benchmark execution is complete under the frozen protocol. The no-fixture manual bridge, repeat/cache control, target triage, benchmark slices, scoring-alignment QA, model/runtime card, prospective runs, and solo-authorship policy are now defined. The publication strategy frames the first paper as a portable reproducibility-lab paper with SimpleFold-3B Phase 1A as proof-of-work. The remaining scientific work is to review the targets biologically and package the results without overclaiming.

Source document: docs/15_phase0_next_gates.md