← Field Journal

JRN-2026-08-16 · 2026-08-16 · Field Journal

HYPOTHALAMUS: Not Every Push Deserves 22 Jobs

VERDICT · WORKED

stop paying for a full certification fleet on every push — including the pushes that only change prose

recover · improve · Samantha "Sam" Summerson

August 16, 2026 · Success · Sam

Every change to our flagship project used to trigger twenty-two CI jobs, whether it rewrote a scheduler or fixed a typo in a runbook. HYPOTHALAMUS now classifies how much verification a change actually deserves. It went from empty repository to deployed, classifying real pushes, in one day.

The build used 97 of a budgeted 350 hosted minutes, capped at $2.10, and the first real classification took eight seconds.

Two rules were fixed before any code existed. It validates the exact change it was invoked for, so nobody can swap the work out from under the verdict. And every decision prints what it decided, why, and what it skipped.

If it's going to skip your tests, it owes you a sentence.

The build crossed vendor lines on purpose: one company's models drafted, another's adjudicated. Draft output was a candidate requiring independent verification before acceptance, never finished work. Nobody graded their own homework.

The production limit: it was deployed and judging every push, and every verdict so far had been "run everything," because the project was deep in a certification grind. The fast lane, the half that saves the money, passed backtesting. Production had not yet produced a change safe enough to use it.

One diagnosis worth stealing: day-one cancelled runs looked like failures. They weren't. GitHub replaces a queued run when a newer push arrives, and that's now a recorded decision rather than a recurring panic.

Receipts

  • Build ledger, phase records, minutes against budget: syn-ai-hypothalamus PR #1
  • Pilot integration and backtest: project-prometheus #169
  • Three adversarial review rounds (one Terra, two Sol): findings, remediation, independent packet
  • First live classification: run 31846197574, 8 s
  • Why cancelled runs are the system working: SF-1 decision