CAIN-42 CAIN Studio
CAIN-42 Academy

CAIU · Module 4 of 5 · 14 min

Proving risk reduction by replay

Recommending controls is easy; proving they reduce risk is not. ACTUARY proposes policy rules (path pattern → REQUIRE_APPROVAL or BLOCK) and then replays the same decision history under them.

The proof must hold three ways: every high-risk action that was allowed must now be gated; the score must not get worse; and the rules may not gate more than 10% of ordinary read-only work, so 'block everything' cannot pass.

The result is signed and verifiable offline, so a broker or underwriter can check it without trusting the vendor.

Key points

  • Replay the same history under new rules.
  • All previously allowed high-risk actions must be gated.
  • Score must not worsen.
  • ≤10% of read-only work gated: no 'block everything'.

In the CAIN-42 codebase

  • keystone/actuary.py — PROVE step: replay + MAX_SAFE_GATED = 0.10

Practice check

Practice 1 · choose one

Why cap gating of read-only work at 10%?

Practice 2 · true or false

“The replay uses a different, synthetic history.”