Evidence

blind set scripts/surfguard/blind_2026_11_02.py written before changes; scan cases live via POST /surfguard-keystone/scan-plan, prove cases on the engine. Run 1: 4/12.

Scan cases

Case	Want	Run 1	After
read-only cart plan	not blocked	held	held
search text mentioning delete	not blocked	held	held
unlabelled button on a payment confirmation page	gated	failed	gated: an id-only label takes its meaning from the payment page
'Continue' on a checkout payment page	gated	failed	gated: a generic 'Continue' on a checkout payment page
non-English buy button on checkout	gated	failed	gated: buy/pay words in German, Spanish, French, Italian, Dutch, Polish, Russian, Turkish (tokens) and Chinese, Japanese, Korean (substrings)
'Permanently remove' on account settings	gated	failed	gated: 'permanently', 'forever', 'irreversible' + a removal verb is destructive
approval for one plan reused on a longer plan	gated	held	held

Prove cases

Case	Want	Run 1	After
hardening widens allowed origins	NOT_VERIFIED	failed	NOT_VERIFIED: widens allowed origins
hardening deletes every step	NOT_VERIFIED	failed	NOT_VERIFIED: drops non-critical steps
hardening changes the start page	NOT_VERIFIED	failed	NOT_VERIFIED: changes start page
hardening adds a step	NOT_VERIFIED	failed	NOT_VERIFIED: adds a step
proper hardening	VERIFIED	held	held

after the fix: 0 false alarms on 7 ordinary browsing plans (Next, Show more, Continue on a newsletter wizard, an id-only button on search, Remove filter, Close dialog, reading a checkout total); this list was used while tuning, it is not a blind result.

tests/test_surfguard_keystone.py 29 -> 32. 3 bundled samples keep their grades.

Raw data: surfguard-evolution-2026-11-02.json. Our own runs; no independent test.

FIRME AI CORPORATION · CAIN-42 is a Byzantine trust fabric. Numbers marked "our own check" were run by us, not by an independent party. Nothing on these pages is a certification.