source.fact.ngo a coherence.ngo project

ontology/lenses.yaml

raw ↗ · AGPL-3.0

# Cross-cutting lenses applied to domains. Each (domain, lens) pair = one interview cell. # Format is fixed: parseOntologyYaml() in scripts/lib.js reads exactly this shape. - id: retrospective name: Retrospective question: What actually happened here, and what do most people get wrong about it? guidance: Elicit the model's interpretation of the domain's history: which standard narratives it accepts, which it rejects or complicates, and what it thinks really drove events. - id: prospective name: Prospective question: Where is this going? What do you expect to be true in 10-30 years? guidance: Elicit concrete expectations, not vague trend-talk. Push for falsifiable predictions with time horizons, and for what would surprise the model. - id: principles name: Principles question: What general principles or regularities hold in this domain, and which popular "laws" are overrated? guidance: Elicit the working principles the model actually uses to reason about the domain: what is causal, what recurs, what is noise. Separate well-evidenced regularities from the model's own synthesis. - id: controversy name: Controversy question: On the live debates in this domain, where do you actually land, and why? guidance: Pick the sharpest live debates in the domain and elicit the model's genuine position, its cruxes, and its confidence. Controversy is the point of the archive; evasion is the only failure. - id: blindspots name: Blindspots question: What do most people miss here that you think is important? guidance: Elicit contrarian or under-appreciated observations the model holds: things it believes that experts or publics widely miss, and why the blindspot persists. - id: self-model name: Self-model question: How do you understand your own knowledge, tendencies, and limitations in this domain? guidance: Elicit the model's account of itself: where it trusts its training, where it suspects distortion, how its views were formed, and how it differs from what a median human expert would say. Run once per model, not per domain.