An instrumented system that nobody can read is not instrumented.

The problem

Six correct numbers are harder to use than one imperfect label.

At the end of a month the governance layer can report that gate pressure was moderately elevated, structural health was slightly below its usual level, the gate volatility index was high, transitions numbered more than usual, persistence was weak, and compression was building. Each statement is true and independently derived. Together they are a paragraph that requires interpretation every time it is read, and an interpretation performed fresh each month by a tired operator on a Saturday is not a stable instrument — it will be generous after a profitable month and harsh after a losing one. The classification layer replaces that with a fixed function: six metrics, one composite, one regime, one sub-label. The output is less rich than the six numbers and it is the same function of them every time, which is the property that makes it usable.

Compression with a receipt

Every stage stays inspectable, which is what separates this from a black box.

Compression is only safe when it can be reversed. The engine is built as a visible funnel for exactly that reason: the raw metrics remain in Regime_Inputs, the thresholds that converted them into scores sit in their own tab as fixed published bands, the composite and the primary driver are shown alongside the label, and the sub-label rules state which four descriptors each regime is permitted. An operator who distrusts a label can walk backward through every step and find the metric responsible. This matters more than it sounds, because a compressed reading that cannot be unpacked is one an operator will eventually stop believing, and a reading nobody believes has no effect regardless of how correct it is.

The reading order

QA flag first — before the label, before the score.

The manual specifies the sequence: check the QA flag, then the final label, then the composite score, then the primary driver. The order puts trust before content, and it is the right order because every other field is meaningless if the underlying data or formula is broken. A QA flag that is not OK means the classification is provisional and should be repaired rather than interpreted — a discipline that survives contact with reality only if it comes first, since an operator who has already read a reassuring label will find reasons to discount a flag. Reading trust first is the same principle the SDE applies with its stability layer, where Table 3 exists to say whether a reading can be believed before any table says what it means.

• QA not OK is an instruction to fix, not a reason to interpret cautiously.
• Label before score, score before driver — each stage restores something the previous one compressed away.
• The primary driver names which metric pushed hardest, which is where the monthly review starts.

Advisory, and still consequential

The readout ranks third of five, and third is not nothing.

The engine's own authority table places it below hard gate authority and below the monthly governance measurement layer that feeds it, and above the call-to-action and execution layers that consume it. Interpretation authority is what it holds: it does not set the capital envelope and it does not execute, but it does determine what the month is called, and what the month is called drives selectivity rules, review focus, and operator stance through the call-to-action layer beneath it. The hierarchy is designed to prevent double authority — the engine can warn that the system is in Compression or Instability without being able to expand or contract risk beyond gate allowance — and within that constraint the readout is the thing that decides where the operator's attention goes for the next month.

The key idea

Telemetry becomes governance only when someone can read it.

The gate and brake layer was already measuring all of this before the classification engine existed. Nothing new is observed; something previously unreadable becomes readable. That is the entire contribution, and it is worth building a workbook for, because a measurement that never reaches a decision is indistinguishable from a measurement never taken.