QIWSign in

Four NSCLC drug targets, ranked from evidence

A small, real run: four non-small cell lung cancer drug targets (ALK, EGFR, KRAS, MET) compared pairwise on 20 criteria by one LLM judge, using only the evidence dossiers collected beforehand. 240 judgements, about a fifth of them abstentions where the dossiers held no usable evidence. The ranking is a Bayesian Bradley-Terry-Davidson posterior with intervals; every judgement, rationale and cited source is open below. Not a clinical recommendation.

Content in the run’s language (EN)

Candidates
4
Judgements
240
Abstention rate
22.1%
Criteria
20
Cost
$0.84
Duration
4m
Judges
openai:gpt-5.6-luna
Models
gpt-5.6-luna
Dossier status
Sep 9, 2026
Published
Sep 15, 2026

Context

Tissue
Lung
Disease
non-small cell lung cancer
Modality
Small Molecule
Cell_Type
mucus-producing glandular cells
Biomarker_Set
Genomic biomarkers
Model_Context
Patient-Derived Xenograft
Patient_Segment
Newly diagnosed
Backbone_Therapy
Chemotherapy
Standard_of_Care
Chemotherapy
Experimental_Context
In vivo proof-of-concept study
Direction_of_Modulation
Inhibition

Leaderboard

RankCandidatePosterior μIntervalObservations
1ALK *0.91[0.51, 1.32]89
2MET0.42[0.02, 0.82]90
3KRAS *-0.54[-0.94, -0.13]89
4EGFR-0.79[-1.20, -0.39]84

* Not decided: the 94% interval overlaps with the next rank, the two are not separated.

Explanation

Front group: ALK and MET; back group: KRAS and EGFR. The separation between the groups is robust (the intervals do not overlap); within a group the order is not decided.

Of 240 judgements, 53 are abstentions. Q12 and Q17: almost only abstentions, no contribution to the order. Q3 and Q4: at least half abstentions. Not in the overall rating: Q12. For Q3, Q12, and Q17 this suggests a run context that does not match the dossiers (run: Small Molecule, dossiers: small-molecule inhibitor). For Q4 this suggests a gap in the dossier schema, not thin dossiers or the context.

For Q1 and Q8 the judge changed its answer in 3 of 6 mirrored pairs: position, not evidence. For Q18 and Q19 the judge changed its answer in 4 of 6 mirrored pairs: position, not evidence.

The order only compares these 4 candidates with each other, only applies in the context Lung, non-small cell lung cancer, Small Molecule and 8 more and is based on the dossiers as of 09/09/2026.

What the ranking does not prove

  • The ranking is only relative between these candidates, not an absolute value.
  • The explanation only holds in this context, not in general.
  • Dossier as of Sep 9, 2026

Self-consistency

Shows how often a judge reaches the same verdict when the order is swapped, without rating the ranking.

Overall

Overall
pairs judged twiceshare flipped
below 60%0
60–79%5345.3%
80–94%4513.3%
95% or above220%

Uninformative criteria

  • Combination potentialThe run’s context differs from the dossier. (Not included in the overall score.)