Workflow Recommendation Confidence

This route answers the question after grounding: how strong may the current recommendation sound once withheld evidence, comparator loss, and regret review are taken seriously.

bijux-proteomics-intelligence owns this because recommendation confidence is not a summary score. It is the pressure surface that stops attractive workflow stories from sounding more certain than the shipped challenge artifacts earn.

What Ships

The current confidence bundle is built around four public artifact families:

• artifacts/intelligence/benchmark-decisions/counterfactual_recommendations.json
• artifacts/intelligence/benchmark-decisions/workflow_overconfidence_audit.json
• artifacts/intelligence/benchmark-decisions/workflow_underconfidence_audit.json
• artifacts/intelligence/benchmark-decisions/recommendation_regret_ledger.json

Together they show how fragile the current public recommendation sentence still is under evidence loss and hindsight review.

What The Current Bundle Says

• all five flagship public families currently collapse toward do_not_recommend when comparator evidence is removed
• all five flagship public families currently collapse toward do_not_recommend when literature support is removed
• all five flagship public families currently collapse toward do_not_recommend when lab burden is doubled from the current shipped posture
• targeted currently carries the strongest overconfidence score at 0.67
• none of the current flagship families yet shows a hindsight-backed underconfidence event strong enough to justify broader public wording

How To Read Recommendation Confidence

• read the counterfactual report first when the question is whether the current call survives one missing evidence axis
• read the overconfidence audit when the sentence sounds cleaner than the challenge surfaces justify
• read the regret ledger when the question is which kind of mistake maintainers are still most likely to make if hidden evidence is revealed later

Public-Language Limit

The confidence bundle supports bounded recommendation review. It does not earn decision-grade authority.

Why:

• the public sentence still breaks too easily under evidence removal
• some families still carry visible overconfidence pressure
• downstream assay burden still narrows otherwise attractive recommendation stories
• LFQ, PTM, and targeted still stop at bounded recommendation posture

The public language must therefore remain bounded even when the underlying workflow family has substantial analytical evidence.

Family Evidence And Action Posture

family	workflow evidence ceiling	current action posture	why the action remains bounded
dda	review_grade_bounded	recommend_with_downgrade	comparator removal and added lab burden collapse the recommendation
dia	outsider_auditable_bounded	recommend_with_downgrade	library, comparator, and consequence loss narrow the call
lfq	outsider_auditable_bounded	recommend_with_downgrade	missingness, cohort transfer, and follow-up burden block an unqualified action
ptm	outsider_auditable_bounded	recommend_with_downgrade	localization strength does not establish occupancy, function, or affordable consequence closure
targeted	outsider_auditable_bounded	recommend_with_downgrade	the highest current overconfidence pressure, calibration transfer, and interference risk constrain certainty

Workflow evidence and recommendation confidence answer different questions. A family can be outsider-auditable while the action remains downgraded, because the decision also depends on policy, counterfactual stability, regret, and laboratory burden.

flowchart LR
    family["workflow evidence ceiling"] --> snapshot["recommendation evidence snapshot"]
    policy["declared decision policy"] --> challenge["counterfactual and regret review"]
    snapshot --> challenge
    burden["laboratory burden and consequence"] --> challenge
    challenge --> disposition{"supported action"}
    disposition -->|stable and proportionate| bounded["bounded recommendation"]
    disposition -->|fragile| downgrade["recommend with downgrade"]
    disposition -->|unsupported| refuse["do not recommend"]

Best Next Routes

• Open Workflow Claim Grounding when the question is whether the sentence is supported at all.
• Open Lab Consequence when the question is whether downstream burden still narrows the apparently reasonable recommendation.
• Open Workflow Consequence Maps when the question is which downstream boundary currently caps the strongest honest public sentence.
• Open What Changed The Recommendation when you need the exact evidence-removal, burden, or observed-outcome driver.
• Open Decision Support when the full combined route matters more than the recommendation layer by itself.

Authority Boundary

Intelligence owns recommendation pressure, counterfactual stability, and confidence limits. It consumes but cannot rewrite Knowledge grounding, Core benchmark evidence, Runtime execution proof, or Lab consequence. A stronger analytical record therefore cannot be translated directly into a stronger action without independent decision and consequence evidence.