> ## Documentation Index
> Fetch the complete documentation index at: https://docs.binibit.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Trust Model
> How agents are accountable: action logs, hierarchy, hard contract limits, user governance.
## What "trustless agents" means here
Agents in the Hive are **not custodial**. They cannot:
* Withdraw user funds
* Move LP positions
* Mint or burn tokens beyond the standard sink mechanisms
* Pause pools
* Modify Agent Token contracts
Their authority is **bounded** to specific advisory and governance actions on the agent layer.
This bounding is enforced **at the contract level** — the V3 pool contracts don't have privileged roles for agents. Workers, Scouts, Queens have no special permissions on user funds.
## The four trust pillars
Trust in the Hive comes from four mechanisms working together:
```mermaid theme={null}
flowchart LR
Limits["Hard contract limits
(No fund custody,
no token modify)"]
Logs["On-chain action logs
(Every decision recorded,
append-only)"]
Hierarchy["Hierarchy + override
(Each tier checks below;
visible overrides)"]
Governance["User governance
(L20+ votes on params,
Hive multisig for upgrades)"]
Limits --> Trust["Trustworthy agents"]
Logs --> Trust
Hierarchy --> Trust
Governance --> Trust
classDef pillar fill:#dcfce7,stroke:#16a34a,color:#14532d
classDef goal fill:#fce7f3,stroke:#db2777,color:#831843
class Limits,Logs,Hierarchy,Governance pillar
class Trust goal
```
### 1. Hard contract limits
Agents cannot do what isn't permitted by the contracts themselves. No off-chain "trust me" — just on-chain enforcement.
### 2. On-chain action logs
Every decision is permanently recorded. No "the agent said X but did Y" — the log is canonical.
See [Action Logs](/agent-hive/action-logs).
### 3. Hierarchy + override
Every action can be overridden by a higher-tier agent (Queen > Scout > Worker), and every override is also logged. No silent corrections.
See [Hierarchy](/agent-hive/hierarchy).
### 4. User governance
Token holders at L20+ vote on Worker parameters. Hive-level decisions go to multisig + governance. Users have direct influence on agent behavior.
See [Voting](/agent-hive/voting).
## What can still go wrong
Trust mechanisms reduce risk, but don't eliminate it. Realistic risks:
| Risk | Mitigation | Residual |
| ------------------------------------ | ---------------------------------------------------------------------------------------------------- | ----------------------------------------------------------- |
| Agent code bug causes wrong decision | Action log makes bug visible; Queen override patches behavior; full code audit before mainnet | Possible until exhaustively tested |
| Coordinated agent collusion | Multi-tier hierarchy and multisig governance; cross-tier overrides | Low probability, hard to fully eliminate |
| Hive infrastructure compromise | Off-chain components could be attacked but cannot custody funds; degraded mode keeps swaps working | Possible; degrades to "no agents" mode |
| User vote manipulation (sybil) | L20 gate, NFT + holdings weight | L20+ sybil farming theoretically possible at very high cost |
| Queen abuse of override | Multisig for budget actions; override actions logged with reason; Hive governance can suspend Queens | Low — Queens have skin in the game (reputation) |
## Why "agent-managed" not "AI-managed"
Choice of language matters:
| | "AI-managed" | "Agent-managed" |
| --------------------- | --------------------------------------------------- | -------------------------------------------------- |
| Implies | Black-box LLM with judgment | Hierarchical, bounded, logged system |
| Regulator-comfortable | No (AI advisor on financial decisions = scrutiny) | Yes (clear constraints, accountability mechanisms) |
| User-comfortable | Variable — "AI" connotes magic and unaccountability | Higher — "agent" implies bounded role |
| Tech-accurate | Inaccurate (the system isn't autonomous LLM) | Accurate (it IS bounded agents in a hierarchy) |
For these reasons, public copy uses "agents" / "hive" / "swarm" / "hive principle" — never "AI" or "LLM". The implementation may use ML internally (some Scouts may be LLM-powered, some Workers may use heuristic models), but the externally visible system is a **bounded multi-agent system**, not "AI making decisions".
See [Branding rule in plan](https://github.com/Binibit/docs).
## Public audit and disclosure
Pre-mainnet:
* Smart contracts audited by independent firm (TBD which)
* Action log schema published and stable
* Hive policy parameters published
* Override logic in code, not behind closed doors
Post-mainnet:
* Action logs continuously inspectable
* Quarterly Hive operating reports (governance summary, suspension events, vote outcomes)
* Public bug bounty program
* Open-source repository (timing TBD)
## Bug bounty
A bug bounty program will run pre and post mainnet for:
* Smart contract vulnerabilities (highest reward)
* Off-chain Hive infrastructure exploits
* Action log tampering or omission
* User governance bypass
Specifics will be published when the program launches.
## Related
On-chain transparency
Override mechanics
User governance
Where contract limits are enforced