Causal Geometry Detail
This demo is about how a causal research system avoids pretending that a collection of local claims is already a settled global causal model. Each paper, endpoint, population, and mechanism gives a partial view; the framework asks where those views agree, where they fail to glue, and what latent structure might explain the obstruction.
Infinitesimal Causality supplies the geometric lens: local intervention effects behave like directions, and non-closing directions expose missing variables, incompatible mechanisms, or under-specified causal scope. Democritus is the current recipe for turning literature into those typed local causal sections.
Infinitesimal Causality Diagnostic
From observed causal variation to a latent obstruction that should remain qualified.
TICKET Transformer
TICKET(democritus_prometheus_run) - Topos Integration using Causal Kan Extension Transformers
Expression: TICKET(democritus_prometheus_run) -> compose(scientific_challenge, evidence_argument, causal_structure) -> glue(extraction_drift_audit)
FSQL: FROM prometheus_runs SLICE BY family("democritus") TICKET BY target_foundry
Admission policy: preserve drift, off-topic extraction, weak causal joins, and unsupported claims as obstructions
Infinitesimal Causality: GLP-1 Run
Democritus is the current computational recipe: it extracts scoped causal PSR cells, preserves uncertainty, and exposes obstructions without turning hypotheses into settled interventional proof.
Admission
democritus_minimum_viable_run_admitted_with_scope_qualification: Add synthesis_auditor and scylla_writer after typing dose/duration on causal edges.
Role Outputs
| Role | Status | Promoted | Qualified | Open Questions |
|---|---|---|---|---|
| source_scout | pass | 1 | 0 | Should broader real-world evidence be admitted in a later run or kept separate? |
| claim_extractor | pass | 5 | 0 | Should maintenance and cardiovascular outcome claims be separated into different foundry cells? |
| causal_mapper | qualified | 0 | 5 | Should causal edges encode dose and duration as first-class typed fields? |
Audit Gates
| Gate | Status | Rationale | Corrections |
|---|---|---|---|
| source_quality_gate | pass | The fixed source manifest preserves study metadata and limitations. | none |
| claim_grounding_gate | pass | Promoted and qualified claims preserve article ids or article anchors. | none |
| causal_scope_gate | qualified | Causal edges preserve population and limitations, but dose/duration should become typed fields before full promotion. | promote dose and duration into first-class causal edge fields |
| handoff_compatibility_gate | pass | Executed handoffs preserve required upstream artifacts. | none |