Decision rules¶
Every public conclusion has one owning layer and a ceiling set by the weakest required evidence. These rules prevent execution, scientific acceptance, grounding, recommendation, feasibility, and authorization from collapsing into one vague “successful” state.
Route the question to its owner¶
| Question | Owning layer | Required record | Cannot establish alone |
|---|---|---|---|
| did the scientific transformation satisfy its contract? | Core | accepted and rejected records, policy, provenance, benchmark result | reproducible execution or biological truth |
| what ran and can the run be reopened? | Runtime | environment, state history, artifacts, hashes, replay result | scientific acceptance |
| what supports or contradicts the claim? | Knowledge | source lineage, evidence graph, confidence policy, review revision | recommended action |
| which action is defensible and stable? | Intelligence | candidates, policy, alternatives, challenge, calibration, regret, posture | laboratory authorization |
| can the proposed work proceed responsibly? | Lab | design, readiness, controls, resources, authority, custody | successful physical execution |
| may the repository publish the combined claim? | release governance | all required owner records and current blockers | stronger evidence than those records contain |
flowchart LR
C["Core acceptance"] --> R["Runtime replay"]
R --> K["Knowledge grounding"]
K --> I["Intelligence posture"]
I --> L["Lab feasibility and outcome"]
L --> G["release decision"]
C -. no shortcut .-> G
R -. no shortcut .-> G
Decision rules¶
- Use the weakest required layer. A strong benchmark cannot compensate for a refused rerun, unresolved contradiction, unstable recommendation, or infeasible follow-up.
- Preserve non-success states. Rejected, refused, failed, held, and inconclusive records are valid outcomes and remain visible in synthesis.
- Keep versions joined. A public conclusion names the exact scientific result, run bundle, evidence revision, decision policy, and Lab record it combines.
- Do not borrow authority. A package may consume another package’s artifact without inheriting ownership of its meaning.
- Append changed decisions. New evidence or policy produces a superseding record; it does not rewrite the prior decision.
- Narrow before extrapolating. Missing transfer, calibration, consequence, or external evidence lowers the claim instead of becoming an optimistic assumption.
Common invalid inferences¶
| Observed fact | Invalid inference | Required next evidence |
|---|---|---|
| command exited successfully | scientific result is accepted | Core acceptance and rejection record |
| artifact bytes are stable | workflow is scientifically valid | benchmark and scientific contract evidence |
| source is cited | claim is supported in this context | Knowledge edge, context, and review state |
| candidate ranks first | action should be taken | challenge, regret, authority, and consequence |
| batch is ready | experiment succeeded | external observation and QC |
| one workflow family is strong | repository-wide authority is strong | independent evidence for every claimed family |
Use Decision Support to follow the complete record and Workflow Families for the current family-specific ceilings.