Search
Posts
The Kernel Must Not Import Its Agents
We extracted a stateful watcher agent from a governance kernel and enforced a one-way agent-to-kernel dependency edge — the kernel must not import its agents.
ReadMaking a Fire-and-Forget Writer Safe Under Failure
Five properties an unattended capture hook needs before you turn it on: idempotent intake, durable outbox, backpressure, atomic receipts, consent gate.
ReadLiveness Without Health Is Theater
A heartbeat that fires on every run proves a job ran, never that it succeeded — the fix is two markers plus meta-monitoring the watchers.
ReadAdversarial Review: The Six Lenses That Halted a Rollout
A six-lens adversarial review checked a team knowledge system against live state, broke three shipped assumptions, and gated 18 risks to halt the rollout.
ReadFix the Dependabot Pile-Up: Policy Over Patches
15 dependabot PRs piling up? Stop merging individually. Group minor/patch updates, auto-merge when green, keep security updates isolated.
ReadNoise-Robust LLM-Judge Evals: Don't Sign a Coin Flip
An un-seeded LLM judge is a coin flip even at temperature 0. How I made signed eval verdicts replayable instead of attesting noise.
ReadEvery Safety Gate Has a Failure Direction
Safety gates fail in opposite directions: one crashed fail-closed, another failed open. Both from a swallowed error. A shell-script control-failure case study.
ReadThe Relevance Score That Broke Our Cite-or-Refuse Gate
A retrieval test asserted every relevance score sits in [0,1] — it failed on a live -1.83 silently corrupting a cite-or-refuse RAG safety gate.
ReadThe Moat Is the Trust Layer: Turning a Local-RAG App into a BYOK Document-Intelligence Platform
Naive local-RAG app became Intent NEXUS: BYOK document-intelligence platform. The moat is enforced trust; adversarial review caught the evals lying.
ReadShipping gpt-5.4 as One Config Line
Deploy gpt-5.4 to production as a one-line config with eval-gated rollout. Key finding: report word-ceilings came from the model, not the prompt.
Read