Safeguard Work-Order Agent Ecosystem

Spec · Validator

Agent Spec — Validator

Status: implemented Source: src/agents/validator.mjs Owner interface: validate(order, classification, routing, ctx) -> validation

Purpose

The safety boundary of the ecosystem. Decide whether an order is safe to action automatically, must be handed to a human, or should be rejected — and record exactly why. The validator is deliberately conservative: when in doubt it returns NEEDS_REVIEW so a human handles the exception.

Interface

validate(order, classification, routing, ctx) -> {
  status:      'VALID' | 'NEEDS_REVIEW' | 'REJECT'
  violations:  Array<{ code, severity: 'blocking'|'review', detail }>
  checks:      Record<string, boolean>
  fingerprint: string   // stable hash for duplicate detection
}

ctx = { seen: Map<fingerprint, lastTimestampMs>, config } carries the duplicate-detection memory and threshold overrides.

Rules

#	Rule	Violation code	Severity	Effect
1	Location/zone resolvable	`MISSING_LOCATION`	blocking	REJECT
2	Description present (≥10 chars)	`MISSING_DESCRIPTION`	blocking	REJECT
3	Region mappable for routing	`UNRESOLVED_REGION`	review	NEEDS_REVIEW
4	Category confidence ≥ threshold (0.55)	`LOW_CATEGORY_CONFIDENCE`	review	NEEDS_REVIEW
5	Priority confidence ≥ threshold (0.55)	`LOW_PRIORITY_CONFIDENCE`	review	NEEDS_REVIEW
6	Estimated cost ≤ auto-approval limit ($5,000)	`OVER_COST_LIMIT`	review	NEEDS_REVIEW
7	Not a duplicate within 24h window	`DUPLICATE`	review	NEEDS_REVIEW

Disposition: any blocking violation → REJECT; else any review violation → NEEDS_REVIEW; else VALID.

Thresholds (configurable via `ctx.config`)

minCategoryConfidence = 0.55
minPriorityConfidence = 0.55
autoApprovalCostLimit = 5000
duplicateWindowMs     = 24h

These are policy knobs, not magic constants. Raising minCategoryConfidence trades a higher auto-action rate for fewer over-escalations and vice-versa.

Verified guarantees (`verify.mjs`)

100% of missing-location orders are blocked from auto-dispatch.
100% of seeded duplicates are detected and held.
100% of over-cost-limit orders are held for human approval.
Exception-detection recall = 1.0; false-auto-action rate = 0.0 on the

synthetic corpus (see proof/LIMITATIONS.md for scope).