Experimental solver development: acceptance

Validated on 2026-10-12 for the unreleased development branch. The existing remote research campaign was not updated, restarted or steered during this change.

Behavior

New minimal studies allow private experimental solver forks by default. The study brief, monitor launch dialog, CLI and agent drafting API expose the choice. An explicit disabled choice survives drafting, launch, resume and continuation preparation. The permission is immutable within a study; a new continuation phase can declare a different choice. Older studies without the field retain their saved instructions. Restricted public trials retain the host’s supplied-solver policy.

Registration preserves existing instrument identities and can record the parent instrument, numerical/physical/diagnostic change type, rationale, preserved assumptions, validation plan and completed build receipts. Workers, directors and reviewers see the same policy and provenance. The monitor displays method approvals with their scopes rather than presenting a benchmark as validation of an entire solver. A newly registered evidence instrument requires an approved production method, including in studies whose original instruments did not require review.

This is a scientific permission and managed-registration contract. It does not change provider-native tools or operating-system access. Turning it off preserves ordinary experiment and analysis scripting. Reference protection still depends on the operator’s filesystem permissions; the checkbox does not make an existing writable installation read-only.

Validation

Check	Result
Focused backend suite	184 passed, 91.29 seconds
Playwright browser suite	9 passed, 32.07 seconds
Python lint, JavaScript syntax and whitespace	Passed
Sphinx documentation build with warnings treated as errors	Passed

Backend coverage includes new-study defaults, strict Boolean validation, immutable choices and provenance, legacy resume without inferred permission, production-method review, remote registration, role instructions, continuation inheritance, resource repairs and restricted hosted permissions. Browser coverage checks desktop and mobile monitor layouts, build provenance links, instrument readiness labels, the default-on checkbox, and a disabled choice persisted into the launched contract. The workspace handoff uses the local execution path with a fake model provider.

The final backend run covered test_solver_development, test_study, test_study_interface, test_research_director, test_research_oversight, test_research_service, test_research_continuation, test_research_resource_repairs, test_research_remote_registration, test_workspace, test_public_workspace and test_public_native. Browser checks covered test_research_progress_browser and the workspace setup/chat/files/study handoff test.

An earlier combined backend/browser invocation exited with SIGILL during Python garbage collection. Its stack did not establish the cause. The final suites ran in separate processes and passed; this does not establish that the combined-process environment issue is resolved. No full-suite or paid-model validation is claimed.

These checks establish harness behavior, not the numerical correctness of an experimental solver repair or its scientific conclusions. See the permission and qualification guide and the earlier large-build and remote-registration repair.