Research records

The research area is a small index over standard BrainlessLab records. It does not add a database or a separate research runner.

Current contract

Every repeated run uses an operation plan and writes one portable record:

record-id/
├── record.toml
├── request.toml
├── resolved.toml
├── seeds.csv
├── data/
├── summary/
├── figures/
├── report/index.html
└── DONE

request.toml retains the exact submitted configuration. resolved.toml retains defaults and derived values. The seed ledger, raw tables, summary, Git revision, Julia version, checksums, and completion state make the run reviewable and shareable.

An ExperimentSpec groups operation plans under one question and version. Each operation still writes its own standard record. A benchmark is one operation type, not the parent of all research.

Public indexes

• Accepted runs reads the committed contribution catalogue.
• Benchmarks lists versioned benchmark protocols and records.
• Experiments lists versioned questions and their evidence states.
• Experimental lists software capabilities, not research results.

The accepted-run catalogue lists accepted contributions and explicit pre-pipeline compatibility records. It does not aggregate outcomes, rank results, or plot parameter-performance relations.

Two-stage contribution path

First, merge the protocol and any required software through ordinary review. This stage fixes the scientific question, operation plans, evidence state, and code that will execute.

Second, generate the public records from a clean Git commit that is reachable from main. The second pull request contains records only.

A contributor submission and its maintainer replay are linked roles within one contribution. They are not independent evidence or additional replicates. Replay checks execution and portability within the declared environment.

Only a human maintainer accepts a contribution. Automated checks can validate the bundle, compare linked roles, and generate the catalogue. They cannot promote evidence.

Publication limits

Public contribution records must not contain private or sealed material. Aim for at most 1 MiB of text-only files in the complete contribution. The hard limit is 5 MiB. This Git-native pipeline does not accept larger files, and large datasets remain out of scope.

Accepted records are immutable. Add an annotation or superseding protocol when an interpretation changes. Do not rewrite the accepted bundle.

The catalogue can include an explicit pre-pipeline compatibility entry for older material. Such an entry appears only when it is indexed and does not imply retrospective acceptance.