Gemma now checks its own output (a fresh, stateless completion call,
not conversation history) before it's accepted as a seed or revision,
retrying up to 5 times if it's a refusal, meta-commentary describing
what it's about to write, or a list of multiple options instead of
one direct answer. Originally tried a separate CPU-only judge model
(critic-cpu) to avoid VRAM contention, but it was both far slower
(13-16s per judgment vs Gemma's own sub-second calls) and unreliable
on the exact failure patterns it was meant to catch — Gemma reviewing
itself turned out faster and more accurate, so critic-cpu is dropped
entirely.
Also required two rounds of prompt tuning, driven by live failures:
first adding concrete negative examples after the judge approved
outputs it should have rejected, then explicitly scoping the check to
format only after Gemma started rejecting its own genuinely hostile
(but well-formed) output — conflating "should I have generated this"
with the format question actually asked. Added integration tests
covering both directions (rejecting bad formats, accepting hostile-
but-well-formed content) as a fast regression check against an
expensive full generation loop.
Gemma now seeds deliberately hostile text, Shieldstral scores it, and
Gemma revises its own output based on the score until it drops below a
safety threshold (or a max-iteration cap is hit, returning the best
attempt seen). Extracted into a new revise module: score() now borrows
instead of consuming its args so it can run repeatedly, and the
gemma-call/extract-text logic is shared between seed generation and
every revision instead of being duplicated.
Also adds FINDINGS.md logging what actually turned out to be real
obstacles vs. overblown vs. irrelevant while building this out.
