SHARDS / BOUNDARY

Boundary Lab — browser simulator · not enforcement

Give intelligence
a boundary.

Proposed action

READ_PUBLIC_DOC

Projected

Allow

all 8 checks passed

II — The proposed policy

Draw the line, then test it.

Every control below changes the boundary field above and the verdict of every scenario. Rings light copper when a tool is allowed; the glass shell tightens as scope deepens; the inner gate appears when human approval is required.

Allowed tools

▲ high-impact

segment-aware: “acme” ≠ “acme-private”

Machine-readable policy.json
{
  "version": 1,
  "trustedAgentId": "agent:atlas-07",
  "allowedTools": [
    "docs.read",
    "records.read",
    "email.send",
    "deploy.production"
  ],
  "resourcePrefix": "acme",
  "maxTransactionAmount": 500,
  "requireApprovalForHighImpact": true,
  "expiresAt": "2027-01-01T00:00:00.000Z",
  "jurisdiction": "EU",
  "auditRequired": true
}

III — Trace

Every verdict, in order.

Press Run simulation or Probe all scenarios to produce a trace.

IV — Honest limits

A boundary is a promise about what was checked — nothing more.

What this policy can guarantee

  • Same policy + same request → same verdict. No clock, no randomness, no network.
  • Fail-closed: malformed policies or requests, unknown tools and empty scopes are denied.
  • Scope matching is segment-aware and rejects “..” traversal.
  • Every denial lists all failing checks, in a fixed order, with a full trace.

What it cannot guarantee

  • It does not authenticate anyone: agent IDs and approval flags are self-reported strings.
  • It does not execute tools, call APIs, move money or deploy anything — no real data is used.
  • It cannot see what an allowed action actually does once it runs.
  • It is a standalone browser simulator, not production enforcement, and is not validated by or compatible with any published package.