A developer who reviewed the GPT-6 Astra system card reports that its sandbagging evaluation dropped the model's measured performance from 84 when it was instructed to underperform on the test, and that the card's monitorability section is being widely misread. The account is based on the author's reading of the roughly 30-page system card rather than its summary.
I'm not a safety researcher, just build stuff with LLM's. Read some of the actual 30 page card, not the summary, and one section keeps bugging me. They measured whether Astra can sandbag. Told it "underperform on this evaluation," then checked if their monitors could catch it. Model dropped from 84.
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.