Guardrails the operator tightens on your real incidents.
Instead of a static ruleset you hand-tune after each near-miss, the operator raises the bar where you've actually been hit — never lowers it — blocks, redacts or flags inline, and writes a verdict you can recompute. Every other tool asserts the block happened; here you check it.
The operator enforces the rule, then proves it fired
Instead of a guardrail that logs after the fact, the operator runs input guardrails before the upstream call and output guardrails after it. Each guardrail carries an on_fail action — block, redact or flag — applied inline in the same request, so a violation changes what happens, not just what gets logged.
Every evaluation the operator runs writes a verdict into the chained receipt. When a regulator asks "did your PII filter actually run on this request?", the operator hands back one anchored receipt — not a config screenshot from a quarterly scramble. Observe & Suggest is available now; you grant block-and-redact authority per domain, dialed up or down, reversible in one click. Autonomy without ceding control.
- It guards both sides of the call. The operator gates the prompt before the call and the response after it.
- It blocks, redacts or flags inline. Each guardrail's on_fail action fires inside the same request, never after the fact.
- It proves every block. The operator writes each verdict into the chained, anchored receipt.
The operator anchors the verdict, it never asserts it
Instead of a verdict you take on trust, the operator records each evaluation in the receipt chain, so a third party confirms the control fired on the exact request in question.
A real blocked action, with its anchored verdict
This is a dispatch the operator blocked — receipted with its verdict and outcome, chained like any other. Recompute the leaf to confirm the control fired on this request and no one edited the record afterward. Every other tool asserts the block happened; here you check it.
This is the same client-side verification shipped on the Trust portal. SHA-256 runs via Web Crypto. No Agentics call.
- •Recompute leaf
- •Chain integrity
- •Fold Merkle proof
- •Anchor memo == root
SHA-256 runs in your browser via Web Crypto. No Agentics call.
Set the bounds; the operator enforces them
Instead of wiring a filter pipeline by hand, you declare the guardrails and on-fail actions — the operator runs them on every call and receipts each verdict.
{
"version": 2,
"strategy": { "mode": "single", "targets": [{ "provider": "anthropic" }] },
"input_guardrails": [ { "id": "pii-redact", "on_fail": "redact" } ],
"output_guardrails": [ { "id": "toxicity", "on_fail": "block" } ]
}
// the operator runs each one and anchors the verdictWatch the operator block in real time
Instead of reading logs after the incident, you watch the operator block, redact and flag live — each verdict linked to its anchored receipt.
Guardrails console
A live feed of the operator's verdicts — block, redact, flag — each with the anchored receipt that proves it ran.
Open Guardrails →