Drop7 Research
← Docs
docs/agents/contributions-and-commits.md

Contributions, validity, and commits

This repository attributes work without confusing effort, authorship, and scientific evidence. Each model or human records its own concrete contribution; result validity is evaluated separately.

Four independent dimensions

DimensionExamplesWhat it answers
Lifecycledraft, preregistered, running, completed, paused, supersededWhere is the work?
Run validityvalid, partial, invalidDid the frozen procedure execute correctly?
Scientific outcomepass, fail, inconclusive, not-applicableDid the gate support this exact claim?
Evidence tierproposal through final confirmationHow strong and independent is the evidence?

A correct run that falsifies its theory is valid and fail. An implementation can be completed while its validity remains untested. “Validated” without a tier and artifact is not an acceptable status.

Theory assessments

Use one of these scoped labels:

  • untested;
  • supported-as-tested;
  • not-supported-as-tested;
  • mixed;
  • superseded; or
  • invalidated-by-methodological-error.

Every assessment cites immutable experiment/run IDs and describes what remains outside its scope. Never write “neural networks do not work” when one network, dataset, and gate failed.

Contribution levels

Contribution level is a compact scope label, not a percentage or a ranking of intelligence:

LevelMeaning
L0Executed or observed an existing workflow without a material change
L1Review, documentation, mechanical support, or a small diagnostic
L2Substantive bounded implementation, analysis, or validation
L3Primary author of a named theory, protocol, implementation, or result
L4Coordinated and integrated multiple independently attributed contributions

L4 does not carry stronger scientific evidence than L2; only experiment design and results do. A coordinator that delegated implementation records orchestration/integration, while the implementing model keeps its own record.

Contribution roles

Use any applicable CRediT-inspired roles:

  • conceptualization, methodology, software, validation, investigation;
  • formal-analysis, data-curation, compute, visualization;
  • writing, review, and orchestration.

For each role, record supporting, substantial, or lead. Include concrete theory, experiment, run, file, artifact, test, and commit references. Also state limitations and whether the record is self-reported or reviewer-confirmed.

Record the exact platform/model string exposed by the host. If either is not available, use unknown; do not infer it from style or capabilities. Do not store private chain-of-thought. A concise decision summary, input references, and tested outputs are the auditable contribution.

Commit convention

Git may not yet be initialized in this checkout. Do not invent commit hashes or initialize a repository automatically. Once Git is available and the active user/platform workflow permits commits, use one coherent change per commit.

Recommended subjects:

theory(afterstate): define successor-closure claim
experiment(afterstate): freeze H40 protocol
benchmark(d4): record machine scaling profile
result(afterstate): record standard-tier failure
infra(research): add deterministic result validator
fix(engine): preserve reveal parity in packed transition
docs(roadmap): reprioritize GPU leaf evaluation

Use the configured human or bot as the Git author. Attribute models through contribution records and trailers:

Theory-ID: TH-20260820-afterstate-a1b2c3d4
Experiment-ID: EX-20260820-h40-83e712aa
Run-ID: RUN-20260820T184215Z-91b02c33
Result-ID: RS-20260820T190501Z-1e7a4c02
Contribution-ID: CT-20260820T184300Z-f482ab19
Evidence-Change: implemented-to-smoke-tested
Result-SHA256: 0123456789abcdef...

Omit inapplicable trailers or write none; never manufacture an ID. A result commit adds evidence instead of rewriting the earlier implementation commit's historical status. Multi-model work lists each contribution ID. The root .gitmessage is a local template; using it does not change global Git config.

Validate a prepared message before committing:

python3 .agents/skills/million-point-research/scripts/researchctl.py \
  commit-lint .git/COMMIT_EDITMSG

The linter requires type-specific theory/experiment/run/result trailers and at least one contribution ID. It does not create a commit or change Git settings.

Review responsibilities

Before a result is promoted, a contributor other than the primary runner should verify:

  • the source, binary, model, data, protocol, and result hashes;
  • that the run did not read a disallowed cohort;
  • that all games and failures are present in canonical order;
  • the metric and confidence-bound arithmetic;
  • that partial, invalid, and censored are not conflated;
  • that model attribution matches actual artifacts; and
  • that the final prose uses the narrowest supported evidence label.

The machine-readable schema and templates are under research/.

For a walkthrough with board animations, start at how the game works and the concepts primer; every term is defined in the glossary.