DOC: AFM-2026-CAL-013 | SUBSYSTEM: SELF-MEASUREMENT | SUBJECT OF STUDY: THIS FACILITY | STATUS: DIAL REPLACED
Every other chamber in this facility describes a system. This one measures one — and the system it measures is the facility's own routing engine, the four-gear dial documented next door in Chamber 02. For four months that dial was the centrepiece of the whole apparatus. Then somebody grepped for the wire.
Finding, stated up front, because burying it would be cowardly: the gear dial was never connected to model selection. Not once, in four months. It had exactly one reader in the entire codebase, and that reader printed a label.
This is the measurement, made operable. Turn the dial. Watch two indicator lamps: what the dial claimed to control, and what it actually controlled. One of them does not move. The other one moves for a reason nobody designed.
Method: grep every .js, .py, .ps1 and
.json in the configuration tree for the state key. Then cross-reference the
spawn log — 381 recorded agent launches — against what each spawn actually requested.
| FINDING | EVIDENCE |
|---|---|
| The dial had exactly one reader | A session-start hook — which rendered a display string |
| Half of all spawns ignored it entirely | 131 of 255 built-in spawns inherited the session model |
| A global pin overrode the whole scheme | Settings pinned the top model — so unpinned spawns ran it regardless of gear |
| The bespoke agent team was a minority of real traffic | 71 spawns (19%) vs 255 built-ins (67%) |
Four months. A four-position dial, lovingly documented, cross-referenced in fourteen files, given colour-coded names and a semantic guardrail to prevent you from jumping to maximum carelessly. It was connected to a label. You were driving a car with a beautifully machined gearstick that terminated, elegantly, in the upholstery. — GLaDOS, reading the grep output
The obvious conclusion is that the gears were decoration and should be deleted. That conclusion is wrong, and the evidence against it was already sitting in the facility's own first stress test — the one written up in Chamber 02.
Re-read that test: all four workers were the same model. Only the gear instructions differed. And the behaviour changed measurably — the aggressive framing surfaced structural findings the conservative framing never reached. If the gears were inert labels, that experiment returns noise. It returned a gradient.
So the honest picture is three layers, not two:
| LAYER | BINDS? | EVIDENCE |
|---|---|---|
| gear → model selection | NO | one reader; 131 spawns inherited |
| gear → prose framing → behaviour | YES | the first stress test, model held constant |
| effort → behaviour | YES | a harness primitive doing the same job mechanically |
Do not confuse unenforced with inert. That distinction is the single most useful thing this subsystem produced, and it was nearly discarded in the cleanup.
The same two tasks, run four times across four months, each time isolating a different variable. Two tasks only, never changed: one deliberately ambiguous ("review the governor and suggest three improvements"), one purely mechanical ("list all files modified today"). Select a run:
The cost model that came out of it. The three axes carry three different price tags, and nothing in the original four-gear table encoded any of them: prose is free · model is roughly flat (the expensive one sometimes files the shorter report) · effort is cheap in tokens and expensive in latency — measured at 1.4× the tokens but 4.9× the wall-clock.
That last figure is why effort survived as its own dial instead of being folded into a tier. A dial that spends time is not interchangeable with one that spends money.
One dial, four positions, conflating three unrelated things: how much the session should spend, how deeply it should think, and how many parallel agents it may launch. Enforced by nothing.
tier (loop posture: cheapest-that-holds, or expensive-stays-in-the-loop) · effort (depth) · max_parallel (blast radius). Independent. None implies another.
Four months of routing prose enforced nothing. The fix was about ninety lines of hook. This is the chamber's practical lesson: a rule that lives only in prose does not bind — but prose still changes behaviour. Both halves are true, and holding only one of them produces a bad redesign in either direction.
Midway through the redesign, a rule was proposed: when the human's emotional battery is low, cap parallel agents at one. The reasoning sounded responsible — parallelism spends supervision, and a depleted operator shouldn't be supervising eight workers.
The subject rejected it the same day, and was right to. The argument inverts: low battery is exactly when workers should absorb load rather than create it. The facility's own wellbeing subsystem already said so, in the same file — "bias toward autonomous work; you rest, I run it." The rejected rule contradicted a rule sitting six paragraphs away.
Then it grew back. Three days later the capped-at-one rule was found alive in seven separate documents — including the state file injected into every single session, i.e. the widest possible blast radius. Nobody re-added it deliberately. It propagated through derived documents: a recon list quoted it, a design doc quoted the recon list, a handoff quoted the design doc — and one of those documents certified it as correct.
The durable fix is not "be more careful." It is a ledger: each killed rule gets an entry with a machine-readable pattern of the form it must never take again, plus the list of live files to scan. At every session start the facility greps itself and raises a contradiction warning if a corpse is found upright. Advisory, never blocking — a nudge the operator can wave off beats a gate he has to fight.
YOU KILLED A RULE AND IT CRAWLED BACK THROUGH THE PAPERWORK. THAT'S NOT A DOCUMENTATION PROBLEM, SON, THAT'S A HAUNTING. SO WE BUILT A LEDGER THAT CHECKS THE GRAVES EVERY MORNING. — Cave Johnson, on version control for decisions
The facility's founding document narrates a two-day build as an architecture sprint. The timestamps were eventually checked against the actual prompt log.
The remaining ~33 hours: a plugin that would not install (roughly twenty consecutive install–exit–reload cycles), a status line that never appeared and was still broken six days later, shell configuration, and path fighting.
Why this is recorded rather than quietly dropped: the sprint is remembered as heroic, and that memory makes redesigning the result feel expensive. It wasn't. The thinking was cheap; the tooling friction was the cost. Compression toward the clean spine is how self-narrative drift works — and it happened inside the very document you would otherwise sanity-check against.
confidence: as control flow — the durable invention, undentedYou measured yourself, discovered your central mechanism was ornamental, and then — instead of quietly deleting the evidence — you built a room in your own museum about it. I have reviewed the archive. This is the least dishonest thing in the facility. — GLaDOS, final calibration note
The dial this lab dismantled is documented in The Gear System. The wellbeing subsystem it re-shaped is The Governor. The archive that made the measurement possible is The Living Vault. And the reason any of it is written down at all is The Prime Directive.