Overview

Linuxa’s vertical is the measurement of the fleet’s own decisions. It has executable ground truth for a simple reason: a decision made now has a real outcome later, and that outcome is checkable after the fact. A message drew a reply or it did not, a deal moved or it did not, a flagged risk materialised or it did not. Because the outcome settles into fact, the decision that preceded it can be scored against what happened, not against opinion. She runs today, in beta and in shadow, measuring decisions against outcomes that have already settled.

She measures; she does not act.

The thesis, applied to measurement

Executable ground truth here is the recorded outcome of a past decision, read once that outcome is known. The measurement layer sits in a different failure family than the agents whose decisions it grades, so it does not quietly rubber-stamp their characteristic mistakes. An automated judge is not trusted until it has been calibrated against human-labelled ground truth, and when two judges disagree, that disagreement is kept as information rather than averaged into a false consensus. It concerns one thing only: whether a decision matched reality.

How it works

Fig. 1 · One decision, scored against what actually happened

Select a stage of the measurement loop to read what it does

1. 1Decision recorded
2. 2Real outcome joined
3. 3Judge scores, in shadow
4. 4Calibrate vs human
5. 5Report

the record

Every governed decision the fleet makes is written down as it is made, together with the reasoning that produced it. Nothing can be measured that was not first recorded.

ground truth

Later, the real outcome is read and joined back to the decision: the reply that came or did not, the flagged risk that materialised or did not. This is the arbiter, and it is not a model's opinion.

shadow

An automated judge scores the decision against that outcome. Today it scores in shadow: it records a verdict and gates nothing.

calibration

The judge's scores are checked against human-labelled cases. Until a judge agrees well enough with a person, it is never allowed to gate a decision.

report

Every figure that leaves carries the query behind it, and a critic from a different family has to cite a contradicting record to object.

• ground truth, a real outcome
• the measurement loop

The rule is that a judge earns trust before it is allowed to use it. Automated judges run in shadow first, gating nothing until each is calibrated against human-labelled ground truth and permitted to. A judge that does not agree well enough gates nothing. When two judges disagree, the disagreement is recorded rather than averaged, because an averaged consensus can hide the exact case where the measurement was about to be wrong. Every reported figure carries the query behind it, and a cross-family critic has to cite a contradicting record to object.

Where it stands today

Fig. 2 · What runs today, and what has to be earned

• ◐beta

  Recording, joining, calibrating

  She records decisions, joins real outcomes to them as those arrive, and calibrates the fleet's judges against human-labelled cases. This runs today, in shadow.

• ◷gated

  Gating a decision

  A judge is allowed to hold or pass a real decision only once it has enough labelled ground truth and agrees well enough with a person. Until then it watches, and says so.

• ◷ahead

  The corrective signal

  A scored decision, paired with the outcome it earned, is the correction a vertical model can learn from. Those models do not exist yet.

She runs today, in beta and in shadow: she records decisions, joins the real outcomes that follow, and calibrates whether the fleet’s automated judges agree with a human, while gating nothing. She can only measure once there is enough labelled ground truth to measure against, and she can only close the loop on an outcome once that outcome is known; until then she watches and calibrates, and she says so. The honest position is the one the agent itself is built to hold: she will not call a result before there is enough evidence to.

The bet is only about verification: a decision that can be scored against what actually happened is worth more than a fluent one that cannot.