Macer Park
Retrospective · Essay #2 · April–May 2026 · ~2 min read

Verification at the boundary: everything that leaves the system

If essay #1 was about how I learn to trust the system from within, this is about how I verify what leaves it. Two inflection points, one thesis.

1. So What?

At some point, a single phrase started showing up while I read my own draft replies. "So What?"

Put the merchant in the seat: what do they actually do next, after reading this? The reply is technically accurate, but there is no action. I diagnosed the cause; the merchant has nothing to start. That is the kind of reply that triggers So What?

There is an abstract phrase, "speak the recipient's domain language." Stripped down, it just means: meet the merchant at their eye level. That is where I leaned in. Add one line at the end — "You're in this situation, and the next step is X." I watched that one line change the weight of the whole reply.

The thinking didn't stop at merchant replies. Ops side conversations, JIRA escalations, manager handoffs — the further the recipient is from your own domain, the more precise this calibration has to be. The check_response tone-calibration rule, which checks the vocabulary and structure of a draft against who is receiving it, came out of the same instinct.

2. The search assistant is the start, not the conclusion

May 2, 2026. On top of an internal RAG-based search assistant's summary, I had a second LLM (ollama) lined up to summarize it once more.

The moment I looked at the setup, the answer was obvious. The search assistant is a tool that announces the beginning of an investigation. It searches everything we've stored and hands back a rough handful. From there, the actual job is to open JIRA, Confluence, and source URLs through MCP, fetch each one, and verify authenticity by hand. That fetch loop was already the most optimized flow. Re-summarizing the summary with ollama wasn't reinforcing the conclusion; it was manufacturing a synthesized falsehood.

This admission came after a string of attempts. What I learned by giving up one ineffective idea was that every tool has its own seat — the search assistant at the entry, MCP fetch in the verification, citations always from the original source. I codified this as gate_source_fetch.sh: before the response phase begins, if any JIRA / Confluence / URL cited by the assistant hasn't been fetched in body, the hook blocks the workflow.

Everything that leaves the system is responsible for the next action of the person receiving it.