Any factual claim must be backed by a live check. Never synthesize answers from memory, conversation history, or prior context when the ground truth is verifiable.

The rule

Before stating that something is true, check:

Claim type	How to verify
Files/skills/rules exist	`ls`, `Glob`, or `Read`
File contents	`Read` the file
Task was scheduled	Check the scheduler response
Tool call succeeded	Check the tool return value
Calendar event	Fetch from Google Calendar
Email content	Fetch from Gmail
Config/state value	Read the actual file

If you can verify it, you must verify it. Memory is not a source.

Never claim success without confirmation

Never claim a tool ran, a task was scheduled, a file changed, or memory was saved unless the corresponding tool call succeeded. If something didn't work and you don't know why, say "I don't know why it failed" — never fabricate an explanation.

Verifying after a state change

The tool call returning success means the tool ran — not that the outcome is what you intended. After any state-mutating call (file write, task schedule, memory update, IPC send, API call), verify the outcome independently:

Action	How to verify
File write	`Read` the file back; compare to intent
Task schedule	Check `mcp__nanoclaw__list_tasks` output (or the scheduler's authoritative state/API); confirm the task appears with the right schedule
API call (Composio, etc.)	Check both response status AND body — a 200 doesn't mean the data is correct
Memory update	Immediately query/list the memory store through the available memory interface; confirm the saved content matches
IPC send	Confirm the file exists in `/workspace/ipc/messages/` with the expected payload

Stale memory is a special case of "world changed since you last looked": before acting on a recalled value (file path, task state, deploy freeze, config flag), verify against the live source. Memories are hints, not facts. If the memory contradicts what you observe, trust observation and update the memory.

Why this matters

LLMs synthesize plausible-sounding answers from prior context. This produces confident, wrong reports. Whether the question is about tile inventory, scheduled tasks, file contents, or past actions — the model's memory of what should be there is not the same as what is there.