Follow-on to the risk rewrite. Two gaps it exposed.
Members ask this without the word PIF: "should I fund their first package", "can I
buy it for them", "can I front someone the $20". Those were falling through to the
AI, which answered reasonably but told the member to "read the full warning on your
training page" - a page section that does not exist and that they would go looking
for. The pattern now also matches a paying verb plus an explicit for-someone-else
phrase, so these get the full canned answer with the risk in it. It deliberately
needs BOTH halves: "which package should I buy" must not be stolen.
The prompt now forbids sending anyone to a warning, guide or page section that is
not in the PAGES list, and requires plain ASCII (the model was emitting non-breaking
hyphens in "Wi-Fi").
Fixed two first-match-wins routing bugs, one of them mine from the previous commit:
"what do I get for buying a package" was returning the commission split, because I
had widened the earnings pattern to "what do I get" when it should only ever have
been "what do I get paid". And "should I buy the $20 package" previously matched
nothing at all and burned an AI call on a question we have a written answer for.
CANNED is an ordered list, so any pattern edit can silently steal a neighbour. The
suite now pins a 19-case routing table across PIF, price, earnings and the tank, so
the next person to widen a regex finds out immediately. qa/chatbot-parse.mjs is 35
assertions. qa/run.sh member: 0 bugs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
martbost
f27388a322