Build a continual scientific research harness

[Aayam Bansal (aayambansal)]

Summary

• Add a generic evaluation-bound scientific harness with immutable contracts, capability-separated evaluators, adaptive quality-diversity search, restart-safe orchestration, evidence/claim gates, and scientific verification protocols.
• Add a Prime-inspired, session-local continual world model with event-triggered refinement, confidence-graded evidence, an immutable base prompt, revision checks, snapshots, and rollback.
• Add quarantined skill evolution, trace-backed integrity, replication and sealed confirmation, semantic/factuality review, simulator and formal-proof validation, ablations, interventions, and provenance-aware reporting.
• Expose generic harness APIs and regenerate the OpenAPI specification and JavaScript SDK.

Preserved history

The branch retains the original 50 harness-development commits unchanged. A reconciliation merge brings that history onto current main, and the product-only harness commit is stacked last. This preserves the complete development record while keeping the final PR diff focused on the current product implementation.

Product boundary

The final tree and PR diff intentionally contain no benchmark catalog, upstream source pins, dataset or download setup, benchmark execution recipes, launcher or pilot runners, local secret/environment templates, benchmark result tables, or benchmark-specific SOTA claims. Evaluation setup stays local or external and binds through the generic harness adapter.

Why

OpenScience should ship reusable product infrastructure while benchmark execution and private/local evaluation remain outside the repository. The continual-state design adopts the useful mechanisms from Prime Agent and Continual Harness—persistent working state, event-triggered refinement, executable reuse, confidence-aware memory, and rollback—without shipping benchmark-specific setup in the resulting product tree.

Validation

• Full bun test suite in backend/cli
• Focused harness suite: 319 passed
• Backend CLI typecheck
• Frontend workspace typecheck
• JavaScript SDK typecheck
• Regenerated SDK and OpenAPI artifacts
• Monorepo pre-push typecheck: 7/7 packages
• Staged gitleaks scan: no findings
• Product-boundary and named-benchmark scope audit

This is a proposed high-performance harness architecture. It has not yet been externally benchmarked, so this PR does not claim measured state of the art.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openscience	Ready	Preview	Aug 7, 2026 7:00pm

[Aayam Bansal (aayambansal)]

KB (@KB-syntheticsciences) thoughts on these harness improvements?