Replay & diff

Test every rule
against real history.

Edit a policy. Before you save, The Relic replays it against the last 1, 7, or 30 days of your fleet's actual decisions and tells you exactly which verdicts would have changed, with sample run IDs to inspect.

policy editor · candidate

# candidate policy

version: "1"

mode: enforce

rules:

- id: allow-search

target: "web_search"

action: allow

- id: deny-fetch-with-creds

target: "web_fetch"

when: "params.url contains 'token='"

action: deny

Replay · last 7 days

Would deny

42

previously-allowed actions

Would allow

0

previously-denied actions

Unchanged

8,124

verdicts identical

3 sample flipped runs ↓

  • 01JD8K2X… · code-assist
  • 01JD8K7M… · data-ingest
  • 01JD8KAQ… · research

The trust loop

Three steps.
Same engine as production.

The replay engine is the same code that enforces policy on the agent's host. No second implementation that could drift. The diff cannot lie about what production will do.

01

Edit the rule

Open the policy editor, draft the rule. Nothing is saved. Your fleet keeps running the old policy.

02

Replay against history

Pick 1d / 7d / 30d. The Relic re-evaluates every recorded action and reports the deltas, usually under ten seconds.

03

Inspect the flips

Click the diff badge. See sample flipped runs side-by-side. Catch unintended consequences before they ship.

What the badge tells you

Three numbers.
Every commit.

Would-deny

Actions the rule would have blocked

The headline for tightening changes. The count of previously-allowed actions that would now fail, with sample runs linked.

Would-allow

Actions the rule would have unblocked

For broadening changes. Tells you whether you're letting through denials you didn't intend to.

Unchanged

Verdicts that stay the same

The reassuring number. Zero flips means the edit is a true no-op. Exactly what you want when refactoring.

Engineering note

The simulator can't lie.

The replay engine vendors its source from the runtime, and a CI job refuses any drift. The diff badge is computed with the exact bytes that enforce policy on the agent's host.

Your action history is stored alongside your traces. The replay runs against the same Postgres + S3 that the platform uses for live observation, so nothing crosses a trust boundary that the trace data didn't already cross.

Back to story