Blog
Field notes
Why stale models hallucinate packages and write last-year's syntax — and how a pre-flight knowledge check fixes both.
-
Grounding the agent: how Hipper's pre-flight check actually works
Hipper separates the reasoning engine from the knowledge index. Three steps — ground, guard, inject — run at the agent's planning phase, in sub-second time and a couple hundred tokens. Here's the architecture.
-
The modernization tax: why a stale model burns tokens on last year's syntax
Every model is frozen at its training cutoff. When a library ships a breaking change after that date, the model keeps writing the old API — and you pay for it in a retry loop. Here's the tax, and how to skip it.
-
Slopsquatting: when your agent invents a package and an attacker is waiting
AI models hallucinate plausible-but-fake package names. Attackers pre-register them. Here's how a hallucination becomes a supply-chain attack — and how to stop it before the install runs.