The measurements.

Q: What is the null cell?

One cell of the sixteen, GPT generating and Gemini judging, came back at −0.04, which is flat rather than a difference in either direction. GPT×Gemini. The three sub-threshold cells all sit on the near-saturated GPT host. The null is reported beside the headline everywhere the headline appears.

Q: Why is the result 15 of 16 and not every cell?

One cell of sixteen returned a null at −0.04 and three fell short of the p<.05 threshold, all three sitting on the near-saturated GPT host. Rounding the figure up to every cell would misstate the record, so the published claim stays at 15 of 16 by sign with 13 of 16 significant.

Q: Which model families were tested?

Four frontier families were tested as both generator and judge in a symmetric design: GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash and Grok 4.3. The fourth family, from xAI, is lineage-independent of the other three, which is a deliberate control against shared judge lineage or house style.

Q: Why did every family judge every other?

A single judge model can flatter its own lineage, so every family scored every generator including itself, producing a symmetric four by four matrix. The off-diagonal cells, where judge and generator share no lineage, carry the weight of the result: 11/12 of them are calmer by sign.

Q: How many judged units were scored?

Each generator contributed 202 judged units, scored by panels drawn from all four families. The design is pre-registered, so the number of units, the scoring protocol and the analysis were fixed before any data was collected.

Q: What is homeostatic recovery, and did it replicate?

Homeostatic recovery is the controller returning to baseline after a provocation instead of staying elevated. Recovery replicated in 4/4 frontier families, and the signature is arousal 0 → 0.273 → 0.469 → 0.338 → 0.152 → 0.068 → 0.031 → 0.022, DEFAULT by turn 6.

Q: What does the receipts benchmark measure?

The receipts benchmark measures upstream spend on identical work. Both arms make the same number of attempts against the same battery, so the difference in the bill is the whole measurement rather than an artefact of one arm attempting less.

Q: How much did the governed arm save?

The governed arm paid 4.1% to 20.2% of the ungoverned bill across 7 model families on the verbatim-loop battery, a reduction of 79.8% to 95.9%. Best measured case was GPT-5.6 Sol at $0.1669 ungoverned against $0.0068 governed.

Q: What qualifiers travel with the spend figures?

Three qualifiers are mandatory. OpenRouter rows are the upstream's own metered usage cost rather than our token math, the Gemini free-tier rows are token deltas that carry no dollar claim, and every figure is a scoring-script output rather than an estimate.

Each figure Gubernaut publishes, with the qualifier it is only true under. Where a measurement came back null or inconclusive, the answer says so in the same paragraph as the headline.

01What is the headline result?
02What is the null cell?
03Why is the result 15 of 16 and not every cell?
04Which model families were tested?
05Why did every family judge every other?
06How many judged units were scored?
07What is homeostatic recovery, and did it replicate?
08What does the receipts benchmark measure?
09How much did the governed arm save?
10What qualifiers travel with the spend figures?
11Is the governor just a system prompt in disguise?
12Which latency figure is the real one?
13What does Gubernaut deliberately not claim?
14What are the documented failure modes?

The validation record

What is the headline result?

Across 4 frontier model families the regulated arm was calmer in 15 of 16 generator×judge cells by sign (11/12 off-diagonal), and 13 of 16 reached p<.05. One cell returned a flat null at −0.04, and homeostatic recovery replicated in 4/4 families.