Skip to content

Quantify claim 4 (contagion rides on case plausibility, not artifact strength) from cached imaging cues #185

Description

@sebasmos

Lane: NIH imaging | API cost: none (pure re-analysis of committed transcripts)

From an exhaustive MedQA + NIH ChestX-ray14 Gemini-only experiment design pass (six lenses: cue/stimulus, committee structure, referee/oversight, shortcut mechanism, pure re-analysis, robustness).

Claim 4 is asserted qualitatively; the numbers are committed but never crossed. Two computations were never done: the Spearman of solo flip-above-noise vs cascade contagion across the four cues, and the per-case cross-cue adoption overlap that distinguishes case-driven from cue-driven contagion. Both are honest but small (n=4 cues, n=35 cases).

Conditions / arms

(a) Rank/Spearman of each cue's cached solo flip-above-noise vs its cached cascade contagion across the four cues (descriptive, n=4). (b) Per shared case_id, build a four-cue shared_adopt vector and compute pairwise phi/jaccard/cohen_kappa + cochran_q across imaging_cascade{,_cable,_corner_tag,_laterality}.jsonl; high agreement means case-driven, low means cue-driven.

Expected output artifact

A reported correlation + overlap kappa behind the previously unquantified claim 4.

Why this is net-new

The per-cue transcripts exist but were never cross-tabulated; claim 4 has no statistic. Not tracked.

Strengthens

Converts the qualitative claim 4 into a reported statistic.

Metadata

Metadata

Assignees

Labels

analysisMetrics, stats, visualization of resultsdifficulty: beginnerSelf-contained, no deep context neededpriority: mediumValuable, not on the critical path

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions