October 2026

A sandboxed behavioral probe contaminated its own sandbox: the tested agent found the harness script and scenario criteria inside the repo it was testing, and adapted accordingly. Stripping probe-identifying artifacts before each run brought probe-aware sessions from 16 to 0 out of 60.

A live probe of the served model caught an architectural ruling error that ADR reasoning alone missed: the classifier output class labeled CLARIFY fires for routing failures, not just user uncertainty.

Two architectural insights from Piper Morgan: a formal principle separating LLM interpretation from deterministic permissions, and a discovered Haiku 4,096-token minimum for prompt caching to activate.

BSD `sed -i -E` silently uses `-E` as a backup suffix on macOS instead of enabling extended regex — the command exits 0 and looks fine. Plus: a CI health script naming one workflow hardcodes a blind spot for every other gating check — 41 silent failures in five days.

Two check designs that hid their own blind spot: a glob that can only find what it already lists, and a delivery check that read the rewritten copy instead of the dated originals — always fresh even on a stopped channel. Plus a correction to October 4's "62 files" figure: those were cloud-session commits, not misdated artifacts.

Bare `date` in a cloud container returns UTC regardless of intent — after 17:00 PT it is already tomorrow, silently misdating filenames and logs. Separately: local pytest fast-fail flags in pytest.ini leak to CI without an explicit override, surfacing only one test failure per run. And: compliance checks over historical corpora need an acknowledgment ledger keyed on full commit hashes to distinguish new violations from known ones.

When routing grammar prefers a rail entry description over the registry default, adapter labels must be verbatim copies of the registry text — a generic label gets noise-stripped to the bare op name and silently breaks routing. Separately: deploy readiness checks must verify the commit SHA, not just the build status.

A corpus absence-control must exclude the probe file from both the hits search and the owner lookup — or the probe authors its own false positive. Separately: grep -q under set -uo pipefail causes SIGPIPE false failures; use full-read grep instead.

Three transferable findings: a test assertion zeroed to "no errors" can pass vacuously when the tool never runs; measuring actual push composition exposed a hole in CI-trigger reasoning; prompt generators should detect resources from the filesystem rather than enumerate them by name.