How Brutor learns

Every judgment Brutor shows — health, verdict, workload class, the coloured edge on a Mission Control card — is learned from the run ledger, not configured. This page explains that learning process and the exact vocabulary the cards use. (The same content is available in the product: Mission Control → How Brutor learns.)

Judgments are comparisons against learned baselines

Every governed call an AI System makes lands in its run ledger. From those runs Brutor learns the system’s baselines: how many actions a task takes, what it costs, which tools it touches, how runs end. Health, drift and the Big-T workload class are all comparisons against those baselines — which is why a brand-new system cannot be judged on day one. There is nothing honest to compare it to yet.

The learning window

While a baseline is still accumulating evidence, the system reads Learning. A learning system can still earn an Assured verdict — nothing it did violated its contract — but its card edge stays blue, not green: assured so far is not the same as proven.

The same principle powers the workload-class chip: below 20 finished runs no class is shown at all. A blank you can trust beats a number you can’t.

The card colour ladder

The left edge of every Mission Control card is the worst of its health and its verdict. Groups inherit the worst edge in their subtree — never an average.

Edge	Meaning
Red	Violations, or the system is suspended. Something the contract forbids happened — this outranks everything.
Orange	Drifting, stalled, or silent. Silent matters most: liveness is the only signal that fires when a system stops calling — every other metric needs traffic to exist.
Yellow	Degraded health, or assured with exceptions — the verdict holds, but with named caveats you should read.
Blue	Learning. The baselines have not seen enough runs to vouch for the system yet. It can be Assured and still blue: watched, not yet proven.
Green	Healthy and assured. Both must hold — a green edge is earned twice, never averaged into existence.
Gray	No judgment exists at all: never computed, or a pseudo-card (unattributed / observed traffic). Absence of evidence is not health.

The two chips

Each card carries two separate judgments, deliberately not merged:

• Health describes behaviour: healthy, learning, degraded, drifting, stalled, silent, suspended.
• The verdict describes conformance to the contract: assured, assured with exceptions, violations, insufficient evidence.

A system can behave normally while violating its contract, and vice versa — collapsing the two into one traffic light would hide exactly the cases assurance exists to catch. See Health & the Assurance Report for how the health score is computed.

Judgments never follow the window selector

Health, verdicts and workload class are always computed on a fixed 30-day window, whatever time range the dashboard shows. The window pills scale measurements — spend, runs, tokens, data transfer — never judgments. There is no view in which a struggling system looks fine.

The honesty rules, in one place

1. Absence of evidence is never green — no data reads as unknown, not healthy.
2. Groups show the worst state inside them, never an average — nine healthy systems and one in violation reads “Violations”.
3. A judgment Brutor cannot stand behind is not shown — below the evidence floor the card shows a blank, not a guess.
4. A silent system is a finding, not a quiet dashboard — liveness fires precisely when everything else has nothing to measure.
5. Judgments never follow the window selector — see above.