Standing instructions do not fire

A reactive loop silently drops every obligation that has no trigger.

The mechanism

A typical autonomous loop:

wake → pull live state → diff against last tick → investigate anomalies
     → start work if a track is free → record → reschedule

Every step is reactive. Something has to move for you to look at it.

Metrics move, so metrics get watched. But "keep the shadow counters in a ledger and decide whether to enable each one" never moves. It has no trigger. It gets honored only when the reactive loop happens to wander into that topic — which is luck, not discipline.

I dropped three explicit standing instructions in a single day this way. Each one I had acknowledged. Each one I had acted on once. None had become a recurring check.

Two supporting failures:

• The ledger was write-mostly. I appended findings constantly and only ever grepped the part I was about to work on.
• No closing check. At the end of a batch I verified the batch's own expected signals. I never asked whether the batch honored the standing constraints.

Three devices

1. Forced questions every tick

Answer these before looking at live state — not "if relevant." Relevance is judged by the same reactive loop, so it leaks through the same hole.

Tune the list to your domain. The shape that matters:

1. Dormant state — is anything switched on that nobody has ruled on? Anything accumulating without a decision-maker?
2. Paired metrics — did I do anything that moves the headline number? Then did I check the quality axes that must not degrade alongside it?
3. Backlog — did I advance anything that wasn't a reaction?
4. Blocked on a human — is anything waiting, and is now a reasonable time to ask?
5. Unreconciled predictions — did I declare expected signals last cycle and never go back and compare?

If you can't answer one, write "could not verify." Written down, the next tick sees it. Skipped silently, it's gone.

2. One backlog item per tick

Every tick, in addition to reactive work, advance at least one backlog item — start it, investigate it, decide it, or kill it.

If there's nothing to advance, write "backlog: 0 advanced." Don't skip quietly.

This is the only reliable way to force non-reactive work into a reactive loop.

3. Pre-ship gate

When writing a work brief and again when opening a PR, re-ask the tick questions about that change:

• Does this turn something on or off? → is the rationale recorded?
• Does this move the headline number? → are the counter-axes declared as falsifiers?
• Does this change something a human should decide? → did you ask?

Constraints have expiry dates

A rule you wrote is not permanently true.

I excluded a file from a work batch because "another change is in flight there." That other change had merged hours earlier. I quoted my own constraint without re-checking it, and the exclusion blocked the most valuable available work.

When you cite a constraint, verify it still holds. Especially your own.

Honest limits

None of this guarantees completeness. What it guarantees is that omissions become visible — an unanswered question is recorded as unanswered.

An invisible omission is the only kind you can never fix.

A checklist also only asks questions someone already wrote down. It is a floor, not a method. See docs/misjudgment-patterns.md for why that distinction matters.

Related

• orchestrator-judgment — what not to delegate during the loop
• agent-memory — where standing instructions live across sessions