# Binder campaign design and interpretation

The [world-model paper](https://www.biorxiv.org/content/10.64898/2026.06.03.729735v1) and [Biohub release](https://biohub.ai/esm/protein) report experimentally validated binder-design results. Reproducing useful hit rates requires the released protocol, selection, and wet-lab screening—not a single generated sequence.

## Plan

1. Define target construct/species/domain boundaries and experimental assay.
2. Select minibinder or scFv modality, framework/template set, and sequence constraints. Preserve exact starting sequences and digests.
3. Establish negative/off-target controls and any liabilities to filter.
4. Choose seeds/templates to yield hundreds and often around 1,000 candidates.
5. Set bounded image build/cold start, GPU hours, persistent storage, concurrency, candidate-count, and cost-ceiling parameters. Refresh current provider pricing and payment readiness.
6. Show the frozen scope and obtain separate explicit current-turn confirmation before any material GPU spend. A CLI confirmation flag is only an execution backstop after this consent.
7. Run deterministic, pinned jobs within the confirmed bounds and preserve failed/partial outputs.
8. Rank with multiple signals: fold/interface confidence, critic evidence, structural plausibility, diversity, sequence liabilities, and controls.
9. Select a diverse experimental panel. Model confidence is not affinity, specificity, expression, stability, or safety.

## Smoke versus campaign

One seed with batch size one proves only that image build, weight load, inference, persistence, and result retrieval connect correctly. Label it `non-representative integration smoke`. Do not use its outcome to estimate hit rate or claim the workflow is scientifically validated.

## Boundaries

Follow Biohub AUP and institutional biosafety review. Do not circumvent managed sequence restrictions or assume open weights permit prohibited work. Treat designed sequences as hypotheses requiring synthesis, biochemical/biophysical assays, specificity controls, and any additional validation appropriate to the intended use.