Decision support¶
Decision support begins after a scientific result exists. It combines a fixed scientific result, execution record, evidence revision, decision policy, and consequence assessment into an advisory posture. It does not turn a successful run into truth or a ranking into authority.
Decision record¶
flowchart LR
S["Core result and acceptance"] --> R["Runtime run bundle"]
R --> K["Knowledge evidence revision"]
K --> I["Intelligence decision record"]
I --> L["Lab readiness and consequence"]
L --> H{"required human review"}
H -->|accept| A["authorized bounded handoff"]
H -->|revise| N["new evidence or policy record"]
H -->|reject| X["refused action"]The chain remains reviewable only when each artifact identifies its inputs. Later observations can change the next decision while leaving the prior evidence and rationale intact.
Minimum review packet¶
| Record | Required contents |
|---|---|
| scientific basis | workflow family, input level, accepted/rejected counts, policy, provenance, benchmark ceiling |
| execution basis | run mode, environment, state history, artifact inventory, hashes, replay and refusal results |
| grounding basis | exact claims, supporting and contradicting evidence, source lineage, freshness, confidence policy, unresolved gaps |
| decision basis | candidate universe, exclusions, normalized policy, alternatives, challenge, sensitivity, calibration, regret, posture |
| consequence basis | proposed assay, controls, burden, readiness, authority, observed outcome, QC, promotion status |
| decision identity | stable identifier, input revisions, reviewer, time, supersession link, allowed action |
Decision postures¶
| Posture | Meaning | Appropriate trigger |
|---|---|---|
| recommend | one bounded advisory action survives required evidence and challenge | evidence, stability, and consequence are adequate for the stated scope |
| downgrade | a weaker action or claim remains defensible | contradiction, sensitivity, calibration, or burden weakens the stronger option |
| hold | decision-critical evidence or review is absent but recoverable | named evidence or authority can close the gap |
| escalate | the decision exceeds automated or package authority | human, domain, safety, or operational judgment is required |
| refuse | a required precondition is violated or uncertainty is unacceptable | no responsible action exists inside the declared policy |
Attribute A Recommendation Change¶
Two decisions are comparable only when the changed inputs are explicit. Keep the scientific result fixed when testing evidence or policy sensitivity; keep the evidence snapshot fixed when testing policy sensitivity. If several axes change together, record the revision as a new composite decision rather than claiming a single cause.
| changed axis | hold constant | conclusion that can be drawn |
|---|---|---|
| comparator or literature evidence | scientific result, policy, burden model | whether grounding pressure caused the posture change |
| decision policy or threshold | result, evidence snapshot, burden model | whether the call is policy-sensitive |
| assay burden or feasibility | result, evidence snapshot, policy | whether the same analytical belief still justifies action |
| observed outcome | retained prior decision and its complete inputs | whether the next recommendation should narrow, hold, or widen |
A recommendation diff must name the old and new decision identifiers, changed inputs, unchanged inputs, posture transition, and exact driver records. A prose summary without that comparison is an interpretation, not attribution.
Follow the limiting evidence¶
- Workflow Claim Grounding exposes support, contradiction, freshness, and context.
- Workflow Recommendation Confidence exposes challenge, calibration, overconfidence, and regret.
- Workflow Consequence Maps connects every family to downstream burden and its weakest permitted posture.
- What Changed The Recommendation compares evidence, policy, burden, and outcome changes against the retained prior decision.
- Lab Consequence evaluates whether follow-up is feasible and informative.
- Outcome Learning Loops returns observed outcomes without rewriting decision history.
Claim ceiling¶
Widen a decision only when the same versioned chain has stronger scientific acceptance, replay evidence, grounding, challenge performance, calibration, and feasible consequence. If the limiting layer cannot be named, the decision is not ready for downstream use.