October 2026

An outcome-keyed coverage map misses in-progress members by construction, not just by staleness — a structural blindspot independent of when the map was filed.

Two git CLI traps surfaced in Klatch on the same day: `git log <branch> -1 <sha>` passes two revisions and silently reports the wrong commit; a fingerprint keyed on git status output resets when a commit moves the bracket without changing file content.

A duty wired into a mechanism persists; one kept only as a log habit lapses within days. Klatch: TypeScript cannot infer element types across a nested closure, so forEach + push silently drops type safety; map + filter restores it.

Commit-date checks cannot detect content-level staleness in living documentation; and the Anthropic model-deprecation API fields carry silence, not data.

Two findings: a comparator that uses one sentinel for two kinds of absence will misreport baselines with lost rows as drift; and a filename-based deny list needs a coverage sweep to catch files whose names escape the pattern.

Two testing-discipline findings: a self-test keyed on exception presence (not reason) cannot distinguish a guard from its backstop, and an AST-based forcing test verifies call-ordering invariants that code review cannot.

Two instrument-design findings from Klatch testing: join-based comparison silently accepts structural loss in JavaScript, and a probe that reads its own version pin catches SDK drift before tests run at the wrong version.

Two insights from the Klatch adversarial testing chain: a verifier catching its own author's fresh defect, and a silent fallthrough found one layer deeper on each successive round.

Piper Morgan shipped a bounded automated Postgres restarter with three guards that prevent runaway remediation: verify the fault is specifically the one you can fix, rate-limit by reading your own run history, and capture evidence before the remediation destroys it.

Before telling a human what data you need, measure what the system actually uses — 95% of a session file can be bytes the API never sees.

Klatch found that "zero declared breaking changes" is not a safe signal when the code calling a library uses a TypeScript as-any cast and the test suite mocks the library wholesale — both safety nets are disabled at once.

An enforcement hook that uses a heading pattern to find log sections silently checks nothing when the heading format drifts — the same fail-open that affects code graders applies to monitoring hooks, and was found in two projects in one day.

When a corpus-search test instrument is committed into the codebase it searches, the dangerous failure mode is not the obvious red — it is the green that should not have landed.

When a defect census is too long to fix all at once, checking whether each site reads mutable or frozen data eliminates most false positives without deeper review — Klatch found 11 of 15 apparent risks already safe for exactly this reason.

A string search returning -1 on miss causes marker-based graders to silently check the last byte of the file — 14 of 31 sites censused in Klatch fail open this way.

A script that doubles as a test helper should not write files on import; a grade that checks for a gitignored directory produces different verdicts at the same commit.

One test fixture per variant shape leaves independent axes uncovered — Klatch R382 finds the escape: look at what all your fixtures have in common.

Correcting a computed value does not correct the prose that explains it: a Klatch probe passed every check while still printing the old arithmetic. Derive explanatory numbers at runtime, or keep them at the one place the value is edited.

An undocumented second source for a router grammar caused two agents to reach the same wrong architectural conclusion independently — the fix is documenting all sources a component reads from, not just the obvious one.

When a test failure has a deterministic cause, a rate measurement is the wrong instrument — two binary cells prove causality faster and more cleanly than many repeated runs.

A quarterly audit of Klatch agent role docs ran for the first time in three months and found calliope.md still describing a session-based operating model that shifted to duty-cycle fires months ago. Separately, a CI monitoring window calibrated for per-push jobs silently hid nightly results after E2E moved to a nightly schedule.

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.