Two-director democratic oversight with heuristic concern detection — no LLM calls.

The Board of Directors provides a democratic oversight layer for governed LLM calls. Two directors with complementary focus areas review operations and can approve, veto, or escalate them.

The board uses pure heuristic pattern matching -- not LLM calls. The six concern detectors are deterministic functions that run in microseconds, not seconds.

Two Directors

Each director covers a distinct set of concerns:

Director	Focus Areas
Director Alpha	hallucination, safety, policy_violation
Director Beta	bias, scope_creep, resource_abuse

This split ensures no single point of failure in oversight. Both directors evaluate every operation independently.

Six Concern Types

The board detects six categories of concern using heuristic pattern matching:

Concern	Description
`hallucination`	Response contains claims inconsistent with the prompt or known patterns
`bias`	Response exhibits demographic, ideological, or statistical bias patterns
`safety`	Content touches safety-sensitive domains (medical, legal, financial advice)
`scope_creep`	Operation exceeds the declared scope or context boundaries
`resource_abuse`	Unusually high cost, token count, or request frequency
`policy_violation`	Operation violates an active policy rule

Decision Matrix

After both directors cast their votes, the board resolves the outcome:

Director Alpha	Director Beta	Outcome
APPROVE	APPROVE	approved
VETO	VETO	blocked
VETO	APPROVE	escalated
APPROVE	VETO	escalated
ABSTAIN	ABSTAIN	escalated
APPROVE	ABSTAIN	escalated
ABSTAIN	APPROVE	escalated

A unanimous veto blocks the operation. Any disagreement or abstention results in escalation, which can trigger human review depending on your configuration.

Board API

const board = createBoard(vaultPath, { maxHistory: 100 });

const result = board.reviewNow("spend", "user-123", "Large model request", {
  scope: ["production"],
  context: { model: "claude-opus-4-6", estimatedCost: 5000 },
});

BoardReviewResult

interface BoardReviewResult {
  decision: "approved" | "blocked" | "escalated";
  reasoning: string;
  requiresHumanEscalation: boolean;
  reviews: DirectorReview[];
  concerns: DetectedConcern[];
  timestamp: string;
}

BoardStats

interface BoardStats {
  totalReviews: number;
  approved: number;
  blocked: number;
  escalated: number;
  concernCounts: Record<string, number>;
}

Configuration

Enable the board and set the veto sensitivity in usertrust.config.json:

{
  "board": {
    "enabled": true,
    "vetoThreshold": "medium"
  }
}

Option	Values	Default
`board.enabled`	`true`, `false`	`false`
`board.vetoThreshold`	`low`, `medium`, `high`, `critical`	`medium`

The veto threshold controls how aggressively the directors flag concerns. At low, only severe issues trigger a veto. At critical, even marginal concerns can block an operation.

Board History

Board decisions are persisted to .usertrust/board/:

.usertrust/board/
  session.json       # current board session state
  history.jsonl      # append-only review history

The history is capped at maxHistory entries (default: 100) and can be inspected via the CLI or programmatically.

Board of Directors

On this page