Provenance trace
Audit trail · tofacitinib → JAK-STAT pathway → Alzheimer's disease
14 source papers, 14 extracted claims, scored by Swanson ABC chain.
Backfilled retroactively from DB state on 2026-05-13. The pipeline_run this row references was created by the backfill script, not by the actual generation orchestrator. Source papers, claims, scoring, and embedding config are recovered; generation-time metadata (exact git commit at gen time, host, exact command) is not recoverable.
1Source papers14
2Extracted claims14
Each row is one structured triple extracted by deepseek-v4-flash (prompt v1_2026_05_07,
source: configured). Confidence is the LLM's self-rating.
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibsuppressesJAK-STAT pathwayA→B
- TofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- TofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- tofacitinibdownregulatesJAK-STAT pathwayA→B
- tofacitinibinhibitsJAK-STAT pathwayA→B
- JAK-STAT pathwaycorrelates_withAlzheimer's diseaseB→C
No claims match this filter.
3Bridging logicSwanson ABC
- edge_support_log_ab
- log(ab + 1) =
2.639 - edge_support_log_bc
- log(bc + 1) =
0.693 - novelty_factor
1.000(no direct drug→outcome evidence — full novelty)- direct_a_to_c_count
0
4Embedding modelbge-m3
Embeddings are used at the L1 filtering stage to prune off-topic abstracts before claim extraction. They do not directly score the hypothesis.
- model
bge-m3- dim
1024- max_length
8192tokens- similarity_scores
- No per-hypothesis cosine scores stored.
5Final scoringhigh · 0.333
- method
swanson_abc_log_evidence_with_novelty_penalty- formula
score = min((log(ab+1) + log(bc+1)) * novelty / 10, 1.0); novelty = 1.0 if direct==0 else 0.7 / (direct+1)- raw score
0.3332- outreach score
0.3332- quality tier
high
Horizontal exploration
Similar hypotheses
Closest 8 of 1,618 hypotheses, ranked by hybrid score (50% bge-m3 embedding cosine · 40% drug/mediator/outcome overlap · 10% shared evidence papers).
- Cross-indication0.87tofacitinib→JAK/STAT pathway→post-operative recurrence
Same drug, different disease — potential alternate indication
- Cross-indication0.87tofacitinib→JAK/STAT pathway→anti-inflammatory environment
Same drug, different disease — potential alternate indication
- Cross-indication0.85tofacitinib→IFN-γ→recurrence
Same drug, different disease — potential alternate indication
- Cross-indication0.85tofacitinib→JAK2→poor prognosis
Same drug, different disease — potential alternate indication
- Cross-indication0.85tofacitinib→IFN-γ→NMOSD
Same drug, different disease — potential alternate indication
Show 3 more
- Cross-indication0.84tofacitinib→IFN-γ→ALS
Same drug, different disease — potential alternate indication
- Cross-indication0.84tofacitinib→STAT3→UC
Same drug, different disease — potential alternate indication
- Cross-indication0.84tofacitinib→IFN-γ→surgical relapse
Same drug, different disease — potential alternate indication
How this hypothesis evolved
First captured May 12, 2026, latest May 13, 2026. 2 snapshots.