This content is not intended for human consumption. Here is why.

Open day at the Ring

On most public days you can drive your own car around the Nürburgring Nordschleife for about thirty euros a lap. The Germans call it Touristenfahrten, or, tourist drives. It is 20.8 kilometres of the circuit Jackie Stewart named the “Green Hell”, and on any given open day it holds a hatchback, a camper van, a motorbike and a track-prepared Porsche at the same time, each driven by someone who paid the same fee.

That is where language models are right now. Everyone was handed a race car, but some people arrived with track experience and some crash at the first corner, and all of them share the same tarmac. The previous piece was about the engine: turning fire into power you can repeat. This one is about turning that power into something usable. And then tuning the hell out of it.

The driver who asked for the track

On 1 August 1976 Formula One raced the old Nürburgring for the last time. The circuit was then 14.2 miles long, by far the longest on the calendar, and Niki Lauda, the reigning world champion, did not think it could be run safely. There were not enough doctors, helicopters or fire marshals to cover that length of track, and rain was forecast. He tried to organise a boycott. The other drivers voted it down by one vote, and so, the race went ahead.

Early in the race, just after the fast left kink before the Bergwerk corner, his Ferrari went off, hit an earth bank and came back onto the track on fire. The people who got him out were other drivers, who stopped their own cars and ran into the flames. He survived, raced again at Monza on 12 September, 42 days later, but Formula One never did return to the Nordschleife.

Imagine that as a budget meeting. The person who could see the risk asked for the infrastructure but the people who would drive on it voted to race anyway, because the track was someone else’s job and the race was today. Drivers never buy the track. They arrive expecting it to be perfect, and they notice it only when it is not.

What a track is

A racing circuit looks like just a strip of road. What turns a house into a home, and a road into a track, are the “finishings”, if you would:

  • Barriers and run-off where cars have left the road before.
  • Marshals and flags so a problem at one corner is known at the next one within seconds.
  • Recovery and medical close enough to reach any point in time.
  • Timing so every lap is measured the same way.

Every corner on an old circuit has a name and a crash history, and the layout is the record of both. The fencing at the Ring was not designed in one go. Much of it went up after somebody found the limit at that spot. The root cause analysis is written into the asphalt. Safety rules are written in blood, and a circuit is one of the few places where you can read them in the layout: every fence, every run-off area and every chicane is a line in that record. Formula One leaving the Nordschleife after 1976 is one of them.

That is the difference between a track and an empty runway in the desert. On the runway you can go as fast as you like, and when it goes wrong there is nobody coming.

What the track asks of the driver

The track is half of it. The other half is what a driver has to do before they are allowed out, and every item has a direct counterpart when you run a language model.

  • The briefing. A track day starts with a driver briefing: the flags, the rules, where the pit exit is, which corners they should be aware of. Miss it and you don’t drive. For a model, the briefing is what it reads at the start of every session: the instruction file, the handover from the last session, the list of known traps. A new session that skips it is a driver who walked in late.

  • Scrutineering. Before a car goes out, someone checks it: helmet, belts, tyres, nothing loose. For an agent, that is the check before the run: which tools it can reach, which permissions it holds, whether the hook that checks its replies is switched on.

  • The coolant swap. Most track day organisers require glycol coolant to be drained and replaced with water. Glycol does its job perfectly well on the road. On a track, when a hose splits, it leaves a film that is almost invisible and very slippery, and the cars behind you find it at speed. A default that is harmless in private becomes a hazard on shared tarmac. The same goes for a model’s defaults: the agreeable tone and the confident summary are fine in a private chat, and on shared work they leave a slick for everyone downstream, in documents and decisions that other people build on.

    Open days at the Ring do not require the swap. During those sessions the Nordschleife is legally a public toll road, and road cars arrive on the motorway and drive home afterwards. Glycol is the right fluid for that car on that road. A general-purpose model is in the same position: its defaults are tuned for the road, the private chat, because the same model has to serve everyone. Nobody at the gate drains them for you. If you take it onto a track, the swap is yours to make.

  • The instructor. Many track days put an instructor in the passenger seat for novices. The instructor does not drive; they watch, and they guide you on what to do where since they are professionally familiar with the track. Separation of duties and seat belts: the one who acts is not the one who checks.

  • The bill. On the Ring’s public days, a driver who crashes pays for the damaged barrier and for the time the track is closed while it is repaired. The consequence is priced, so drivers can see it before the corner. For a model, the price is the error budget and the failure record: what a crash cost, written down where the next driver will read it.

None of this makes anyone faster on the first lap. All of it is why they are still driving on the fiftieth.

The engine was reliability; the track is performance

The engine from the last piece makes the power repeatable. The track is where you find out how much of it you can use. Those are separate problems, and they need separate investment.

A track exists so that the limit can be found at a survivable size. A driver who spins at a corner with run-off, marshals and a recovery truck learns something and drives again next week. The same spin on a public road ends the lesson for good. That is the whole purpose of the place: consequences arrive, and they arrive small enough to learn from.

For anyone running language models, the track is the environment around the work:

  • A sandbox or staging environment where an agent can do the wrong thing without touching production.
  • An evaluation environment where a model can be pushed to its limit on purpose, and the result is recorded.
  • Permissions that act as barriers: the dangerous tool is simply not on this stretch of road.
  • A failure record that works like the crash history of each corner: read before you drive, and updated after every spin.

Most teams have an engine by now. Very few have a track. They are running open day on production.

Nobody wants to pay for the track

Tracks cost a fortune to keep, and the money rarely comes from the people driving on them. Even the Ring could not make it work: in July 2012 the company running it went insolvent, unable to service a 330 million euro loan. The drivers paid for their laps. Nobody wanted to pay for the track.

Company infrastructure runs into the same wall. The track is a cost line in one budget, the race is a demo in someone else’s, and whoever signs off the first rarely sees the second. On a quiet day the barriers look like waste. On the day something goes wrong, everyone asks why nobody built them.

The price ladder

The Ring sells the same tarmac at very different prices to very different drivers. Roughly, from cheapest to most expensive:

  1. Open day. Pay per lap, bring any road car, share the track with everyone. Cheap, crowded, and entirely on you.
  2. Track day. An organiser books the circuit, limits the number of cars, and usually adds instruction. More expensive, much safer, and you learn faster.
  3. Industrie-Pool. Car makers and suppliers book closed testing time together, under their own rules, to develop cars on the most demanding circuit they can find.
  4. Racing. Endurance series and the 24-hour race: teams, pit crews, strategy, engineering, a support structure around every car.

The same ladder fits language models, and it is the one I am building:

  1. Open day: you, a chat window and whatever you teach it. This is where TYChat starts: the rules, the hooks and the failure record that turn a naked chat into something that holds its line.
  2. Track day: guided sessions, where someone who knows the corners rides along while you build your own setup.
  3. Industry pool: a team builds its track properly: sandbox, evaluation, permissions, failure records, gauges at the far end.
  4. Racing: the point where this stops being individual and becomes a field with shared rules, shared safety standards and shared records of where the crashes happened.

Doing all of this solo is possible. It is also nightmare difficulty. Every successful racer has a track, competitors, a pit stop, a crew and an R&D department behind the car.

From open day to Formula One

Right now the whole field is at open day. That is fine; it is how every motorsport started. At some point the community will move toward something like Formula One: a shared track standard, marshals, scrutineering before a car is allowed out, and a record of every incident that changes the rules for everyone. Formula One got safer through changes to the tracks, the cars and the rules after each crash, and every team had to follow them.

Until then, you can take part on open day. Bring your A-game, and teach your chat the A-game.

What would refute this

  1. A team that runs agents straight on production, with no sandbox, evaluation environment or failure record, and gets the same incident rate as a team that has them.
  2. A field where safety improved mainly through more careful drivers, with the tracks and rules unchanged.
  3. Infrastructure that gets funded before its first serious incident, as a rule rather than an exception.

Terms used here

  • Track: the environment built so that failure stays survivable: sandbox, evaluation environment, permissions, failure record, measurement.
  • Open day: everyone on the same track, any vehicle, each driver on their own; the current state of language model use.
  • Drivers never buy the track: the people who use infrastructure expect it to be perfect and do not fund it; the people who fund it rarely see it used.
  • Crash history: the dated record of where things went wrong and what changed afterwards; the track equivalent of a failure record.
  • Price ladder: the same infrastructure sold at different levels of support, from pay-per-lap to a full team.

Changelog and errata

  • v1, 2026-09-27: published.
  • Errata: none so far. Corrections will be listed here with their date, and the corrected text marked in place.

The series

This is part of a series by Stefan Coetzee, 2026, on running language models as working systems.

Research programme and claims ledger: machinebehavior.io.