# Trust Model

SSE combines on-chain verification with AI transparency to minimize trust assumptions. This document covers what's enforced today and what's planned.

## On-Chain Guarantees

These guarantees are enforced by the smart contract — no trust in the backend required.

| Guarantee           | Mechanism                                                     |
| ------------------- | ------------------------------------------------------------- |
| Answer immutability | `SHA-256(answer:salt)` committed on-chain before any deposits |
| Fund safety         | PDA escrow — no external wallet has custody                   |
| Payout correctness  | Deterministic BPS calculation with checked arithmetic         |
| Round integrity     | Sequential round IDs prevent replay/skip attacks              |
| Fee transparency    | Treasury and buyback addresses locked at `initialize`         |
| Evidence cap        | Evidence payouts hard-capped at 30% of pool                   |

For implementation details, see [Shared Patterns](https://simulation-theory.gitbook.io/simulation-theory-docs/contracts/shared-patterns) and [Security Model](https://simulation-theory.gitbook.io/simulation-theory-docs/security-model).

## AI Transparency

For rounds resolved by the AI Judge, trust comes from transparency:

* Published adjudication prompt framework
* Stable, versioned evaluation rules
* Visible rationale published every round
* Archived inputs and outputs for every round
* Consistency trackable across rounds

The Judge cannot pick randomly — it follows a reproducible logic process against published criteria.

## Whale Resistance

Stake influences weight, but does not guarantee a win:

* In guessing games: spending more gives more attempts, but doesn't guarantee the answer
* In prediction rounds: larger stakes add weight, but the AI Judge evaluates argument quality independently

This is a core protocol property, not a soft guideline.

## Moderation Policy

### Allowed

* Absurdity, bluffing, taunts
* Narrative threats ("AI choose us or face collapse")
* Memes, roleplay

### Blocked

* Real-world violent threats
* Doxxing
* Slurs / severe harassment
* Illegal content

If content is blocked, the player's stake remains valid. Only the message is redacted.

## TEE Roadmap

### Current State

The backend generates answers and commits them on-chain. The commit-reveal scheme proves the answer was fixed before deposits, but players must trust the backend generated the answer fairly.

### Planned: Phala Network Integration

Moving AI inference into Trusted Execution Environments (Intel TDX via Phala Network) adds hardware-level trust:

| Component            | Current              | With TEE              |
| -------------------- | -------------------- | --------------------- |
| Answer generation    | Trust the backend    | Verified in enclave   |
| Commit hash creation | Trust the backend    | Signed by enclave key |
| Answer commitment    | Trustless (on-chain) | Trustless (on-chain)  |
| Fund distribution    | Trustless (on-chain) | Trustless (on-chain)  |

With TEE:

* The exact code and model that ran is verifiable via Intel TDX attestation
* No one, including the node operator, can observe or tamper with the inference
* Attestation reports are stored on-chain per round

See [Roadmap](https://simulation-theory.gitbook.io/simulation-theory-docs/resources/roadmap) for the TEE integration timeline.