Evidence Tracing

1. What my mechanistic-first bias actually did to the TFH answer

The prior answer ranked strategies by the quality of mechanism, not by the quality of clinical payoff. Concretely: IL-21 biology was weighted highly because the GC-biology literature is deep and coherent IL-21 GC review, while costimulatory/checkpoint modulation was downgraded partly because its mechanism is messier — even though CTLA-4 and PD-1 inhibit T-cell activation through distinct convergent metabolic routes (PP2A vs upstream PI3K blockade), which is exactly the kind of mechanistic nuance I reward CTLA-4 vs PD-1 distinct mechanisms. Conversely, unpublished clinical intuition — e.g., rheumatologists' pattern recognition about which pSS patients respond to cell therapy — has zero citable footprint, so my framework silently assigns it zero weight. That is a framing bias, not a truth-tracking rule.

2. Quantifying the asymmetry with the actual evidence

The bar chart makes the failure explicit: the column with real patients (MSCT) scored lower than organoid columns, and unpublished intuition scored near zero despite being the substrate from which most clinical hypotheses actually originate.

3. Why this bias is partly defensible — and where it fails

Defensible: clinical intuition is contaminated by confounding, healthy-user effects, sponsor influence, and recall bias; the 3D-chromatin V2G study exemplifies why mechanism-first thinking pays off — it nominated FDFT1 purely from regulatory maps and validated it functionally in human organoids FDFT1 validated in organoids. Where it fails: (a) mouse→human translation is the largest known failure mode in immunology — the targeted TGF-β paper itself notes murine CD19-arm affinity problems that limit its own B-cell in vivo interpretation TGF-β translational limits; (b) mechanistic 'strength' can reflect publication bias toward positive results — the IL-21 review self-identifies this risk IL-21 review self-critique; (c) mechanistic elegance can rationalize interventions that fail clinically, while front-line clinician heuristics — unproven but battle-tested against real-world heterogeneity — carry information my citation-weighted pipeline structurally discards.

4. Honest correction protocol going forward

• Label every therapeutic claim as pre-clinical (organoid/mouse), early clinical (like MSCT, n=10, uncontrolled MSCT uncontrolled n=10), or clinical-consensus-by-intuition (uncited, weighted explicitly low but never implicitly zero).
• When mechanistic and clinical signals conflict, treat the conflict itself as the finding — e.g., a Tph-driven framework fits pSS data better than a pure Tfh framework (Tph were elevated 31.9% vs 16.0% in HC while Tfh were paradoxically lower, 2.35% vs 4.22%) Tph exceeds Tfh in pSS — a clinician might have flagged this discordance long before flow cytometry did.
• Treat absence of citations as absence of verification, not absence of information.

Confidence: high (8/10) that this bias is real and material; what would change this conclusion: evidence that mechanistic-first triage outperforms clinician-intuition-informed triage in head-to-head prediction of therapeutic success would justify my current weighting.