The inputs tab is a mirror. In production, nobody types into it.

The funnel

Eight tabs, and each one is allowed to do exactly one thing.

The engine is built as a funnel with a fixed direction. Monthly advanced metrics are created in the gate and brake rollup. The inputs tab mirrors six of them into a clean table. A thresholds tab converts each raw metric into a score on a fixed scale. The score engine sums, assigns a primary driver, maps the regime, selects the sub-label and runs its own quality checks. A regime map turns composite bands into main labels, a sub-label rules tab constrains which descriptors are legal for each, an output tab renders a selected month, and definitions and playbook tabs translate the result into meaning and posture. Raw telemetry enters at the top and an operator-readable instruction set comes out at the bottom. The value of the separation is that every tab has one job and can be checked against that job alone — which only holds if no tab quietly acquires a second one.

Why mirroring beats entry

One source of truth is a property of the wiring, not of the operator's care.

The six metrics already exist, computed from verified gate and brake records that the governance layer itself runs on. Referencing them means the classification is derived from the same numbers that produced the month's actual capital states, with no translation step in between. Typing them means the classifier is derived from a transcription of those numbers, and a transcription is a second copy that can differ from the first. It usually will not. When it does, nothing detects it, because both copies are formatted identically and neither is marked as authoritative. The systemic version of the argument is that the engine's whole claim is to describe the system's own internal condition — and a description built from retyped figures is describing the transcription, which is a different object that happens to resemble the system most months.

The testing exception

The exception is narrow and comes with an obligation.

Manual entry is permitted for one purpose: deliberately testing what the engine does with a given set of inputs. This is a legitimate and useful activity — establishing which combinations produce a transition, checking that a threshold change behaves as intended, or confirming that the sub-label rules reject an illegal descriptor. What makes it safe is that it is bounded and reversed. A test that is not reverted leaves the workbook in a state where the output tab renders a regime label for a month that never happened, indistinguishable from a real one, and it will be read as real the next time anyone opens it. The obligation attached to the exception is therefore not a formality but the entire content of it: the tab returns to referencing before the session ends, and any conclusion drawn during the test is labelled as coming from a test.

The calibration tab

The thresholds tab is the one place the operator is meant to change something.

The engine's design is deliberately unclever, and the thresholds table is where its judgement is stored. It defines the cutoffs that convert each raw monthly metric into a score, and it encodes whether higher or lower values are favourable for each. Because there are no weights in the first production version, every metric contributes through the same scale — which means the thresholds are the only calibration surface in the entire engine. Everything the classifier believes about what constitutes a heavy month or a healthy one lives in that table. The instruction to adjust cautiously and document every change follows directly: a threshold edit silently reclassifies history, so a month previously labelled one thing becomes another with no record of why, and comparisons across the year quietly stop being comparisons.

What lineage buys

Every label can be walked back to six numbers and one table of cutoffs.

The practical dividend of a strict funnel is that a classification is never a black box, even though it compresses a month into a single phrase. Any label resolves backwards: the label came from a composite score, the composite came from six scores, each score came from one raw metric and one threshold band, and each raw metric came from the gate and brake rollup for that month. There is no step in that chain that requires trusting anyone's memory, and no step that a second person could not repeat. That is a stronger property than accuracy, because a classifier that is auditable can be corrected when it is wrong, while a classifier that is merely accurate today has no mechanism for staying that way.

• Ingestion is referenced, never typed — the exception is testing, and testing is reverted.
• Thresholds are the only calibration surface, and edits reclassify history.
• Every label walks back to six raw metrics without consulting anyone's recollection.

The key idea

A classifier's credibility is a property of its inputs' provenance.

It is easy to evaluate a classification engine on whether its labels feel right, and that is the wrong test, because a label built from convenient numbers will feel right most of the time and will be indistinguishable from one built properly on any given month. The test that separates them is whether the inputs can be traced. A funnel with one editable tab, one testing exception and a documented change log gives a plain answer to the only question that ultimately matters about a diagnostic: where did this come from? Systems that cannot answer it are not less accurate — they are unfalsifiable, which is worse.