9. The simulation as the cost unit

# 9. The simulation as the cost unit

> **Dated reframing — 2026-09-20.** The old Program model remains valid for Directors who ground a
> world in live evidence. For simulation-shaped users, the modal bill is now generation plus fate
> search: a person brings a situation, the engine builds a kept world, develops forked arms, prices
> pathways, and prepares a reading.

## One brought situation, end to end

One simulation is the smallest honest unit for the causal simulator:

```text
intake → world → three forked arms → Meridians World Search (MWS) rounds → readings → debrief
```

The numbers are action counts, not a promise that every surface is automated. Situation-to-world
intake currently runs by hand through MCP; automated MWS is represented by bounded Scenario
Lab work, not a shipped autonomous service. Published Experiences are generated once and then served,
so readers do not pay generation COGS on every reading.

<!-- gen:simulation -->
#### One simulation

- Interview (1 subj × 6q): $0.00 (1× @ 0.0017)
- Domain creation (bible + 4 scenes): $0.04 (1× @ 0.0388)
- Reasoning graph (CRG): $0.01 (1× @ 0.0131)
- Arc generation (4 scenes): $0.08 (3× @ 0.0255)
- Scene plan (1 scene): $0.12 (12× @ 0.0098)
- Scene prose (~1.2k words): $0.09 (12× @ 0.0074)
- Game-theory scene decomposition: $0.00 (3× @ 0.0006)
- Scenario post-score Game Theory: $0.01 (9× @ 0.0007)
- Variable model (present/future): $0.03 (3× @ 0.0095)
- Image (Seedream): $0.24 (6× @ 0.04)
- TTS narration (1 scene): $0.32 (3× @ 0.105)
- Chat message (1 turn): $0.01 (30× @ 0.0005)
- Embeddings re-index (whole domain): $0.01 (1× @ 0.006)
- **Total simulation COGS:** $0.95

#### Monthly usage by simulation-shaped profile

| Profile | Monthly usage COGS |
|---|---:|
| `USAGE_SIMULATION_ONE` | $0.95 |
| `USAGE_PARTICIPANT` | $2.12 |
| `USAGE_PRACTITIONER` | $3.91 |
| `USAGE_STRATEGIST` | $1.57 |
| `USAGE_READER` | $0.11 |
| `USAGE_LEARNER` | $0.99 |
| `USAGE_AUTHOR` | $1.79 |
<!-- /gen:simulation -->

The generated block above is owned by `scripts/gen/personas.ts`. It uses `simulationCogs` and
`usageCogs` from `unit-economics.ts`, so model routing, provider prices, and fleet allocation remain
centralised.

## Monthly profile shapes

The profile table is a planning surface for simulation-shaped use:

- **Participant:** two situations, one retry, and reflection chat.
- **Practitioner:** four simulations with extra trait-oriented readings. The latent-trait layer is
  unbuilt and must remain private, non-clinical, non-diagnostic, and never a verdict.
- **Strategist:** one situation held while doctrines and arms are played through the Scenario Lab.
- **Reader:** served Experience reading plus a light return footprint; generation is not charged to the
  reader.
- **Learner:** one training world and one consequential question.
- **Author:** fiction-writer usage plus MWS-shaped work over novel branches.

These are usage profiles, not personas that the model may silently turn into product claims. The
persona PMF readout remains a fit-to-shipped-surface estimate.

## Why Hosted headroom hangs off simulations

Program crawls still matter when a Director asks a world to track live evidence. They are not the
primary cost unit for a Participant, Reader, Learner, or Author. Their spend is driven by interviews,
world construction, scene and image generation, branch play, MWS rounds, and debrief chat.

Hosted headroom therefore needs to be read against simulation volume and profile mix rather than
domain count alone. License remains BYOK; Hosted absorbs provider spend. A crawl-only allowance can
look safe while a few generation-heavy simulations consume the tail.

## What the model cannot yet see

Two meaningful costs remain open estimates:

1. **Voice intake minutes.** Voice interview is direction today; the model does not yet price duration,
   transcription, or voice-turn retry.
2. **Automated MWS compute.** Scenario Lab runs provide bounded evidence, but autonomous
   search over a generated branch fan is not shipped. The action profile uses existing planning and
   game-theory keys as a transparent placeholder, not a measured compute bill.

Neither cost is silently assumed to be zero. They need observed workloads before becoming committed
prices or Hosted allowance policy.

Raw Markdown source