Autoresearch V5 Results

Public snapshot of V5 adaptive protein binder search compared with official Proteina-Complexa seeded baselines. The page exposes result curves, operator diversity, and the recorded reasoning attached to each search node.

Generated: Strict: iPAE < 7/31, pLDDT > 0.90, RMSD < 1.5 A Public data: metrics and hypotheses only

Discovery Curves

Cumulative Discovery

X-axis counts completed Proteina-Complexa child GPU-hours. Qwen/vLLM controller time is not included in the curve axis.

Target Snapshot

Per-Node Yield

Each point is one completed child job or official baseline repeat. X is the chronological node index within each run, not tree depth or GPU time. Y is the selected per-node yield; use Discovery Curves and Run Table for cumulative performance.

Operator Mix

Run Table

Evidence Graph

This graph shows public evidence provenance, not a literal scheduler tree. Edges are inferred from explicit MethodCard evidence references; nodes without a recorded evidence parent are attached to root. Root children therefore include bootstrap nodes plus independent exploration/bandit proposals. Baselines are independent fixed seed-sweep repeats.

Select a node

Click a node to inspect the public hypothesis, trigger, recipe, evidence references, and measured outcome.

Wet-Lab Panel Candidates

When a V5 wet-lab panel is available, this table shows top selected candidates without sequences or local paths.

Definitions And Caveats

Global-New Strict

Strict successes after sequence-level global deduplication within a run.

Foldseek Unique

Structure-unique strict success is shown only when run-level Foldseek clustering is available. Pending Foldseek is not inferred from sequence diversity.

Public Reasoning

The evidence graph includes MethodCard hypothesis, metric trigger, falsification rule, operator, lane, sanitized recipe, and measured outcome. Root-attached nodes are independent or missing explicit evidence parents, not necessarily first-generation descendants.