Skip to content
How it works

How a rehearsal runs.

Four steps. The longest one is launching it. The hardest one is reading the result honestly.

01 / Build

Build the scenario from your sources.

Pose the decision in plain language. Pick its archetype. M&A, capex announcement, crisis response, activist play, policy package, GTM repricing. Upload the documents behind it: the draft press release, the deal model, the brief, the legislative text. Glasshouse builds the world from your sources, not from training data.

  • Six archetypes; closed registry, no marketing-invented event types
  • Dossier upload: PDF, DOCX, MD, plain text, one or many
  • Stakeholder cohorts inferred from the dossier; editable before launch
02 / Configure

Set the population, the surface, the duration.

How many personas. Which simulated platforms: analyst desks, financial press, retail forums, regulator-side comms, internal channels. How long the simulated window runs. The defaults are reasonable; the controls are explicit.

  • Population size: 25–80 personas, with cohort balance preserved
  • Engagement platforms: per-archetype defaults, fully customisable
  • Simulated duration: 24h, 72h, 7d, or custom up to 30 days
03 / Backend

Pick where it runs.

Local on the developer machine for sensitive work. Cloud for capacity. Per-stage assignment: heavy stages (report generation, persona-fidelity scoring) on the capable backend, lighter stages on the fast one. The backend is configuration, not architecture; the same scenario runs anywhere.

  • Local: llama.cpp on Mac / Linux workstation
  • Cloud: Vertex AI / Gemini, with Pro for capability-limited stages
  • Hybrid: per-stage assignment via configuration, no code fork
04 / Launch

Launch. Watch honest progress.

One button. The build takes minutes to tens of minutes; the simulation rounds run for ten more. The progress UI shows three honest layers: stage stepper, phase strip, micro signal (a node arriving, a post landing). It never invents a percentage the backend cannot justify.

  • 34 entities extracted, not 47%
  • Past stages stay clickable; you can jump back into the graph at any time
  • The Diagnostics section in the report exposes run-level quality before you act on it

Now see what comes out.

Inside a sim See an engagement