Workflow Recommendation Challenges

bijux-proteomics-intelligence owns recommendation posture, so it also owns the surfaces that can embarrass that posture.

These challenge pages exist to stop elegant decision prose from sounding more trustworthy than the hidden evidence it depends on. The repository now ships blinded recommendation challenges for every flagship workflow family that currently carries outsider-facing recommendation posture.

What The Challenge Actually Does

The sequence is simple and deliberate:

1. freeze the recommendation posture while some companion or perturbation evidence is still hidden
2. reveal the withheld pressure packet
3. publish whether the original recommendation landed as a hit, a miss, an overconfidence event, or an underconfidence event

That is why this surface matters. It tests whether the intelligence layer stays honest when the easy-looking view is no longer the full view.

What Ships

• one blinded recommendation challenge per flagship family: dda_blinded_recommendation_challenge.json, dia_blinded_recommendation_challenge.json, lfq_blinded_recommendation_challenge.json, ptm_blinded_recommendation_challenge.json, targeted_blinded_recommendation_challenge.json
• one cross-family overconfidence audit at artifacts/intelligence/benchmark-decisions/workflow_overconfidence_audit.json
• one cross-family underconfidence audit at artifacts/intelligence/benchmark-decisions/workflow_underconfidence_audit.json

Current Family Outcomes

• dda: 1 hit, 1 overconfidence, 0 misses
• dia: 1 hit, 1 overconfidence, 0 misses
• lfq: 1 hit, 1 overconfidence, 0 misses
• ptm: 1 hit, 1 overconfidence, 0 misses
• targeted: 1 hit, 1 overconfidence, 1 miss

How To Read Those Outcomes

workflow family	strongest current reading
dda	the family can survive one blinded hit, but cross-engine reveal pressure still shows that import-backed confidence can sound cleaner than the harder adjacent package earns
dia	the recommendation layer still leans too optimistically on library-conditioned evidence when the harder matrix-shift view is revealed
lfq	recommendation posture remains bounded because sparse-contrast and missingness pressure still create visible overconfidence debt
ptm	localization can look recommendation-ready before ambiguity-aware consequence pressure pulls the sentence back down
targeted	the family shows the sharpest practical failure because interference and carryover pressure can collapse an action that still looked viable from the cleaner brief

The targeted miss matters most right now. The hidden interference and carryover perturbation collapses the follow-up lane that still looks actionable when only the cleaner decision brief is visible.

What These Challenges Prove

• recommendation posture is no longer judged only by summary confidence prose
• the repository can show where hidden evidence creates measurable overconfidence debt
• recommendation errors are published as first-class review surfaces instead of being folded back into vague caution language

What They Do Not Prove

• that the benchmark packet itself is scientifically complete
• that one blinded hit authorizes broader public wording
• that intelligence can widen past grounding, runtime, or lab consequence boundaries on its own

Best Next Routes

• Open Workflow Recommendation Confidence when the question becomes cross-family overconfidence, underconfidence, or regret rather than one blinded reveal.
• Open What Changed The Recommendation when you need to see how comparator removal, literature removal, or burden changes move the sentence.
• Open Workflow Consequence Maps when the question is whether the challenged recommendation still survives the downstream assay burden it would trigger.

Why This Belongs Here

These surfaces are intelligence work rather than core or knowledge work.

The benchmark packages, literature surfaces, and contradiction dossiers still own truth. This package owns what the recommendation layer does with those surfaces when it must choose, stay bounded, or refuse.