Crux: PI-approved bounded computational research loops

Research applicability

A documented way to organise and iteratively test computational alternatives under a researcher-defined remit; useful for examining controlled research-loop design without claiming unsupervised science.

Getting started

Editorial, untested guidance: read the flight-plan and ordinary-versus-autopilot control documentation, define a non-sensitive bounded question/scorer and examine stop/approval conditions before any execution. Do not treat a read-only cockpit, local hash stamp or prose review gate as independent scientific validation. Keep the PI responsible for the final synthesis and merge.

Maintenance

Pinned documentation and original dated announcement inspected; maintainer support, future compatibility and operational robustness not independently audited.

License

Repository README states MIT; third-party agents, dependencies, papers and datasets have separate terms. Full licence/security audit not performed.

Workflow & human oversight

Documented-loop L4 only for a PI-approved computational autopilot with a fixed objective, researcher scope/design controls and attempt/time/worker-call limits. PI approval delegates attempts; final synthesis, answers and merges remain human decisions. Not the deterministic checker or demonstrated autonomous science. Enforcement/outcomes were not executed or verified.

Workflow level is not an evidence score; higher is not better evidence.

Problem

Researchers using coding agents need to retain questions, experiment evidence and decisions while controlling a bounded computational search.

Approach

The PI frames the question/null/bar and harness; an agent prepares checks, scored baseline and flight plan, stopping before approval. After PI approval/start, the driver edits/runs/scores repeated attempts and records hypothesis verdicts under the approved remit. Success, budget, abort and stall rules constrain the loop; final interpretation and merge remain with the PI.

Outcomes (as reported by the originating team)

Capability/protocol documentation only for this assessed scope. No Crux engine, scorer, agent, example experiment, stop enforcement or scientific outcome was executed or independently evaluated here. An announced claim that preregistration prevents p-hacking is not adopted as an established result.

Limitations

Documentation review, not implementation/security/enforcement audit. Ordinary notebook PI approval rules differ from the approved-autopilot exception; individual attempts need no new signature inside that remit. The general close path can warn about missing files, but later05.3 autopilot documentation writes and links a report for every attempt; the generic warning must not be treated as normal later-autopilot report omission. Earlier05.2 steward refusal does not generalise to later05.3 bounded steward/island delegation. Guidance is excluded from plan-hash coverage. Worker-invocation limits do not prove a financial/token cap, and fixed checks do not establish causal validity or prevention of selective analysis.

Reproducibility

Primary forum/README/control/CLI text saved with original access dates and source hashes. Selected autopilot document has API-byte/Git-blob verification. No research software, model calls or scientific findings reproduced.

Adoption requirements

A configured external coding/research agent, repository and suitable experiment/scorer, PI-chosen objective/null/thresholds, scoped writable/frozen paths and budget, plus genuine approval and validation processes. Documentation lists Python/git and optional host-specific tooling; dependencies/permissions and data rights need local assessment. No install or run was performed.