One Place Decides Which Model Runs

Accumulation is the default and it is silent. A tool arrives for 1 job, stays for 5 more, and 18 months later nobody can name what renders the hero shot or why. The fix is 1 file that answers 1 question: what runs for what. Ours lists 5 categories, and every row carries the model, the price, the axis scores and the date of the last probe. That is 40 characters of decision per row. The alternative is what most stacks actually run: 11 subscriptions, 3 of them abandoned, and a render bill nobody can attribute to a decision. We built 200 wood-gasification machines across the UK and Europe on the same discipline, 1 drawing, 1 owner and no undocumented substitution on the shop floor. A pipeline that cannot name its own components is not a pipeline. It is a habit with a login.

The table is only worth having if something checks it, and a press release is not a check. The loop is 3 steps and it runs at 08:00 every morning: 1 scan for releases, price moves and news across the 5 categories, 1 comparison against the table that prints deltas rather than a report, and 1 routing change, made only when a live probe or render beats the incumbent on that category's weighted axes. Deltas only is what keeps it readable. A morning that surfaces 2 changes gets read. A morning that surfaces 40 lines of status gets filed. The scan is a cron rather than a meeting, the log is 1 page per day, and the whole loop reads in 10 minutes. 3 steps, 1 file, 0 meetings.

No routing decision moves on an announcement. That is rule 1, and it is written at the top of the file because it is the rule that gets broken first. A vendor benchmark is a claim about a distribution, and the only thing that counts here is a probe or a render that beats the incumbent on the axes that matter for the category. For video that is 3 axes: motion coherence, identity hold and price per second. For coding it is the pass rate on 30 of our own tasks rather than a leaderboard. The probe costs cents, and it is cheaper than 1 wasted 40-shot render. I signed a EUR 20 million personal guarantee against a plan that assumed every future moment would arrive on schedule, and the company folded on Christmas Day 2008. What survived it is 1 habit. Measure the instrument, then trust the number. A model nobody has probed on our own work is 1 line in a press release. Not a decision.

Rule 2 exists because every price figure has a shelf life measured in weeks. Balance and price are re-verified on every run and never carried forward, because a number that was true in March is a guess in June. Each row records 3 price columns: the list price, the last verified balance, and the date of that reading. When a balance falls under the floor for a lane, the lane is marked dry and the run falls back to the next scored entry rather than discovering the problem mid-render. A stale balance shows up as a 402, and a 402 at 60 per cent of a render is 20 minutes gone. That 1 convention has saved more afternoons than any model upgrade we have bought. 3 columns, 1 floor per lane, 0 surprises. A price you did not re-read this month is a memory rather than a fact.

A fleet that accumulates is a bill. A fleet that is chosen is a system.

The table is auditable a month later because nothing in it is unsourced. Every external fact, whether a release date, a context window or a price move, carries its own source line beside the entry, so the question is never who said that but where it is written. 1 convention, applied 100 per cent of the time. That is what lets the file survive a handover, and it is the same standard the public writing runs on: every claim carries a source and a verdict. When 2 entries disagree, the newer source line wins and the older one is struck rather than kept. A table with 1 unsourced row is a rumour with extra columns. 12 departments and 1 operating standard was the version of this at a EUR 75 million industrial group, and it worked for 1 reason. The document was the authority rather than the person who happened to be in the room.

The finished mechanism is 1 table, 5 categories, 2 rules and 1 daily log, and it fits on 1 screen. That is deliberate. A routing system that needs a meeting to change is a system that stops changing, and the value is not the sophistication of the scoring. It is that the decision has 1 home and the check has 1 clock. 16 years of field work, from rural Nigeria to post-conflict Serbia to Kazan to Fukushima, taught the same shape in 4 industries: the instrument that gets used is the 1 that takes 10 minutes to read. The probability that an undocumented stack drifts into a model nobody chose is not small. 210 energy systems were deployed on the same shape, 1 written order, 1 owner per step and no substitution without a line. 1 place decides which model runs. That is the whole system.