binder-lane-round

Binder Lane Round

Use one round contract to compare one or more binder-design toolchains against a fixed public target and site.

Read First

• references/docs/binder-lane-round.md
• references/docs/binder-study-decision-loop.md
• references/docs/binder-controls.md
• references/NON_CLAIMS.md
• references/docs/tooling-and-licensing.md
• references/docs/compute-backends.md

Scope

The lane validates requests, materializes plans and handoffs, runs synthetic contract checks, and runs validated fixed-argument local adapters after explicit authorization. A runtime registry below .runtime/ can describe another local tool, API client, cloud client, scheduler, or container entry point.

Required Decisions

Declare:

• the public or synthetic target accession, chain, and required residue selections
• whether reference_scope preserves the Anthropic workflow shape or checks the published tool identities on the bounded stages
• whether each arm replays those identities, replaces tools, or joins a replay-and-replacement comparison
• the generator, sequence designer, predictors, scorers, and filters
• the execution mix for each stage
• code, weight, dependency, API-term, redistribution, and use-context constraints
• the budget ceiling and runtime cap
• the planned round count
• one primary metric and a checkable stopping rule
• the required and optional controls for any measured calibration
• the expected artifacts, exact counts, hashes, validation notes, and cleanup requirements

Workflow

1. Record the target, site, study mode, tool mix, execution mix, use constraints, budget, rounds, metric, and stopping rule.
2. Run bsf binder-lane menu, then validate the request with plan-request.
3. Materialize plan.json, round-contract.json, and execution-handoff.json with plan.
4. Run preflight, then run target-check on the coordinate input before generation.
5. Use run only for a public_synthetic_demo plan.
6. Run adapters to inspect bundled records. If a selected tool has no bundled command, create a validated adapter registry below .runtime/.
7. Run prepare-execution with the exact target report and stage settings. Resolve its per-selector readiness gaps before local controller execution.
8. Dry-run each local adapter or the prepared controller request.
9. For a remote route, validate the request with remote-request. Use remote-receipt after the user-supplied dispatcher exports artifacts and verifies cleanup.
10. Before a real start, obtain approval for the named route, data posture, budget, runtime, and any paid launch, non-public upload, terms acceptance, or large or license-gated download.
11. Start an approved local adapter or controller with --authorize-local-execution. Add --authorize-network or --authorize-license-gates only when the selected adapter requires them.
12. Run closeout to count, parse, and hash the exact declared outputs before stage completion.
13. For a measured primary metric, run calibrate-controls and inspect its readiness record.
14. Run round-decision against the sequential round history. Pass the ready calibration with --calibration when the metric requires it.
15. Start another round only when the decision is continue and the remaining budget covers that round.