Simulation
Methodologies
Grounded in evidence,
not vibes.
Actual Simulation is opinionated about how it reasons over evidence. Every technique is grounded in approved evidence, cited back to source, and framed as directional — never as prediction.
Evidence grounding & provenance
Every persona trait, claim, panel rating, and briefing finding is tied to verbatim evidence chunks — nothing is asserted without a traceable source. The source-approval gate is the trust boundary.
Contradiction mapping
Rather than smoothing disagreement away, synthesis surfaces where the evidence conflicts with itself — across personas, the knowledge graph, and the research briefing.
GraphRAG-lite knowledge graph
Evidence is extracted into an entity/relationship graph. Retrieval embeds the query, ranks entities, and expands one hop to pull the most relevant chunks; “contradicts” edges become surfaced tensions.
Evidence-grounded STORM briefing
Adapts Stanford OVAL’s STORM (NAACL 2024) — multi-perspective question asking — but run over your real evidence, not the model’s priors. Each lens retrieves its own evidence, contradictions are mapped, findings synthesized, and a peer-review pass grades every finding 1–10. The self-critique step STORM itself flagged as missing.
Synthetic panels
A persona or evidence-sampled panel is run against measures to estimate a population’s distribution of responses — reported as rating distributions with variance, and calibratable against real human data.
Synthetic focus groups
Several evidence-grounded agents discuss a moderator’s questions over rounds, reacting to each other — a multi-agent take on the same multi-perspective principle, with a cited readout.
Cross-interview sense-making
Per-conversation and per-interview summaries, plus a method-neutral project overview — recurring themes, coverage, tensions, gaps — that organizes the evidence without prescribing a research method.
The stance
The app organizes and points; the researcher interprets. Outputs are explicitly directional and evidence-cited, with reliability levels where it matters — never presented as prediction.