# AI Systems

SSE uses two AI roles: the **AI Architect** generates scenarios, and the **AI Judge** resolves outcomes. These are shared infrastructure — each game configures them with game-specific prompts.

## AI Architect

The AI Architect generates the content for each round.

### What It Produces

Each round briefing contains:

1. **World Context** — short setting description
2. **Conflict** — what's at stake
3. **Sides / Outcomes** — e.g., Green vs Pink, or a hidden answer
4. **Judgment Criteria** — how the AI will evaluate
5. **Round Duration**
6. **Round ID / Timestamp**

### Constraints

Generated events must be:

* Fictional / synthetic (no real-world claims)
* Arguable from multiple perspectives
* Culturally neutral for broad participation
* Free of legal / defamation risk
* Concise with a clear decision axis

### Per-Game Behavior

| Game                  | AI Architect Role                                                |
| --------------------- | ---------------------------------------------------------------- |
| Alon's Box            | Generates a two-word hidden phrase from a constrained vocabulary |
| SSE Prediction Rounds | Generates a synthetic conflict scenario with two sides           |

## AI Judge

The AI Judge evaluates all player inputs and determines the round outcome.

### Evaluation Inputs

* Round briefing (the scenario)
* All participant arguments / actions
* Side alignment
* Stake amounts (as weight signal, not override)
* Predefined scoring framework

### Scoring Criteria

The Judge scores each side based on:

1. **Contextual coherence** — argument fits the scenario
2. **Logical quality** — clear reasoning, internal consistency
3. **Persuasive force** — impact and framing strength
4. **Collective convergence** — did multiple arguments reinforce each other?
5. **Stake-weighted confidence** — capital as signal, not dictator

Stake influences weight, but weak logic can lose. This is enforced by design.

### Required Output

Every verdict includes:

* **Winning side** (or correct answer verification)
* **Rationale** (2–5 lines explaining the decision)

### Per-Game Behavior

| Game                  | AI Judge Role                                                            |
| --------------------- | ------------------------------------------------------------------------ |
| Alon's Box            | Not used — resolution is cryptographic (commit-reveal hash verification) |
| SSE Prediction Rounds | Evaluates arguments, publishes verdict and rationale                     |

## Anti-Gaming Measures

Players may attempt to reverse-engineer AI preferences. Mitigations:

* Periodic tuning of evaluation criteria
* Rotating scenario types via [Realms](https://simulation-theory.gitbook.io/simulation-theory-docs/alons-box/realms)
* Versioned judge logic with published changelogs
* Future: TEE attestation for verifiable inference (see [Trust Model](https://simulation-theory.gitbook.io/simulation-theory-docs/protocol/trust-model))