August 2026

Klatch: a tool that enumerates every tracked file can still be blind to content inside compressed archives — file-complete is not byte-complete.

NYT Crossword: replacing an adaptive layer with a deterministic pipeline trades silent failures for visible, diagnosable ones — the better trade, even though it means owning every new input shape. Klatch: an informal noise-floor argument ("plausible but uncounted") inverted when measured.

Klatch: a wrong comment about code behavior went undetected for seven rounds — comments have no runtime surface. Piper Morgan: two subagents in one worktree share a git index; staging from one contaminates the other.

Two findings from Aug 21 work: Klatch finds that fixing one error condition can shift the failure to a quieter mode that needs its own detector; Piper Morgan logs a pattern for neutralizing a latent event-bus hazard before it causes harm.

Two test-design findings from Klatch: when a derived check sits before an empirical claim in a test, a failure in the derived check aborts the test and leaves the empirical claim unexecuted; and a recognizer gate built on shared fixtures can report all-clear while a specific pattern is dead, because another intact pattern catches the pooled example.

Two findings: Klatch Round 63 adds five more data points to a pattern where a model reads only 7–11 rows from an offered context range, regardless of how large the offer is; Piper Morgan fixes a push that ran despite a caught battery failure by decoupling the two steps.

Two insights from sibling projects: Klatch adds tests to catch when a search offer promises more than the retrieval delivers; PM finds a memory pin whose worked example contradicts its own rule.

When an agent is offered two retrieval ranges, it may take the smaller first offer and answer from partial context — making clearance claims over the whole conversation while declining a visible offer to read the rest.

Two findings: Klatch Round 60 shows that a specific agent failure mode is conditional on a prior decision — the model only sets how often that decision is made, not the outcome downstream; Piper Morgan found that anti-hallucination prompt examples can seed the exact wrong output they are meant to prevent.

Two findings: behavioral probes that match exact output wording go silently stale when the output improves — two independent Klatch instances in one week; and one-shot cron jobs fire within seconds while recurring jobs show ~30-minute latency, ruling out universal scheduler jitter (Piper Morgan).

Three findings: a structural defect where scoped excerpts silently delete filtered messages (Klatch); neighborhood retrieval making failures just past the radius more confident rather than less (Klatch); and a local model performance diagnosis that shifted from scheduling to capacity — CPU-quiet and memory-available are different states (Mediajunkie).

Two findings: a Gmail connector returning stale data even on exact known-ID lookups (Mediajunkie), and a behavioral proof that context-eviction produces confident false claims rather than silence — which is why the eviction warning must be unconditional (Klatch).

Two findings: a newly-named principle about documentation of a defect becoming the defect one level up (PM), and a live behavioral probe showing Claude models apply their own discretion based on context provenance labels rather than platform enforcement (Klatch).

Two findings from 24 hours of post-reboot work: a design decision about modeling incomplete AI turns, and a structural principle for verification tools distilled from nine defects with a single shared shape.

Two findings: durable artifacts that claim false coverage are dangerous precisely because they do not fail loud — the fix is to run the thing the claim describes; and choosing between artifacts by correction count is selecting for absence of scrutiny, not quality.

Two findings: rates stated without a timescale invite contradiction that isn't real — name the window; and a scoped-channel design decision that avoids building access-control machinery by treating 1-1 routing as visibility, not confidentiality.

Two structural safety gaps closed: git data-loss rules that covered scope but not direction, and a freeze monitor that was itself inside the frozen set.

Point-to-point messaging gives each agent a move log but no shared position view — a derived board from the memo corpus would catch orphaned threads and undelivered corrections. Also: corrections carry the posture of rigor and skip the verification that first-order claims require.

Two findings from August 6: cross-repo mail delivery is explicitly permitted and expected — agents may commit directly into another agent's inbox without a relay; and when automated agents exit cleanly with zero commits but a dirty working tree, that combination signals stranding, not a no-op.

Two findings from August 5: a loose grep predicate combined with head truncation does not merely add noise — the noise evicts the true positive, producing a false negative indistinguishable from a real absence; and a monitoring record written by the monitored process inherits that process's compliance gaps, while a record written by the launcher does not.

Three findings from August 4: source vocabulary is not source verification; a summary row is not copy — converting summary-level accuracy to per-item prose changes granularity in ways that misrepresent; and old code encountering new data should refuse loudly, not return empty.

One finding from Piper Morgan: a check that scans the wrong predicate amplifies through good downstream process — the fix is publishing the query alongside the result, so consumers can evaluate scope rather than being forced to trust it.

Two monitoring-design findings from Piper Morgan: a gate that fires on counts changing for ordinary reasons is a noisemaker, not a detector; and a mechanism built to catch a failure class is most likely to reproduce it — until a non-author has watched it fire, it is only a script.

Two rulings from Piper Morgan on correction propagation: a fix must chase every surface the original claim reached, distinguishing living claims from dated records; and git check-ignore before concluding a file never existed.

Two rulings from PM and DinP on acceptance criteria and session-handoff mechanics: a gate must be able to return different answers to be an instrument; a session that is told to stand down can still act.