Search
Posts
Coverage Said 69%, Mutation Testing Said 25%
A repo at 69% line coverage scored 24.88% on mutation testing—and the rules engine that touches user email scored 0.00%. Coverage said fine; Stryker didn't.
ReadWhen LLM Output Lies Instead of Crashing
An LLM-output parsing bug silently understated a cost report via case-sensitivity. One-line normalizer fix plus 5 defensive hardening steps.
ReadThe LLM Should Never Do the Math
A Claude Code skill that hunts Databricks cost leaks and reports confirmed dollars from the customer's own billing tables — never LLM estimates.
ReadA Denial You Can Audit Beats a Silent Drop
The governed second brain shipped: one plugin, two modes, and a governor that hands back a receipt when it refuses a write instead of dropping it silently.
ReadA Reply Your Bot Loses to a Crash Is One Your User Never Got
Making every bot reply path — streaming and file-upload — crash-durable. The unglamorous reliability work that separates a demo bot from one people depend on.
ReadN Bots in One Slack Channel Loop Forever: Three Gates
Put N AI agents in one Slack channel and, left ungated, they answer each other forever. The three moves that make a shared multi-agent channel safe.
ReadNo-Toggle Test: Make 'Not Built Yet' an Enforced Invariant
Shipping single-tenant first is sound—but 'we haven't built multi-tenancy yet' rots as a comment. A static no-toggle test makes it an enforced invariant.
ReadVerified Identity for AI Agents: Off, Warn, Enforce
Give AI agents an Ed25519 verified identity, then roll out an off/warn/enforce daemon gate that refuses to act on an unverified one — like an unsigned commit.
ReadThe Merge Is the Trust Boundary: Re-Derive as Untrusted
In a multi-writer knowledge store the merge is the trust boundary: re-derive the union as untrusted, re-govern from scratch, and anchor it with a signed DAG.
ReadReject PII at the Source: A Disclosure Gate at Intake
The cheapest place to enforce a privacy rule is the front door — a fail-closed disclosure gate that rejects comp data and PII at candidate intake.
Read