One of the six inputs has a disputed direction, and the manual says so out loud.

The five that are settled

Direction is not a property of a number; it is a claim encoded in a threshold table.

Each monthly metric enters the classifier through a threshold band that converts it into a score, and every one of those bands encodes a direction — whether a higher raw value earns a favourable score or an unfavourable one. For five metrics the direction is uncontroversial and follows from the definition. Gate pressure measures how heavy the month was from drawdown, spike clustering and compression, so lower is stronger. Structural health measures how well the system handled the month rather than whether the month was easy, so higher is stronger. The volatility index measures instability relative to the size of the month and the transition count measures churn between gate states, so lower is stronger for both. Persistence measures how long states held and how durable the structure was, so higher is stronger. Each of these is a direction that can be read off the metric's own definition without argument.

The sixth

Compression can mean the constraint or the freedom that survived it.

The compression score resists that treatment because the underlying phenomenon is genuinely two-sided. Compression is the narrowing of available capital freedom, and a score built on it can legitimately be defined as either the amount of pressure applied or the amount of freedom preserved. Those are complements, and they take opposite signs. The current regime scoring treats the metric as a defensive stress input, meaning lower is stronger. The advanced monthly metrics interpretation companion, in some contexts, treats a high compression score as expansion integrity preserved — the opposite reading. Both are coherent internally. They cannot both be applied to the same column without the classifier scoring a month favourably and unfavourably on the same evidence depending on which tab is consulted.

Why it matters arithmetically

With equal weights and no z-scores, one inverted sign is a sixth of the composite.

The engine's first production version deliberately uses no weights — each of the six metrics contributes through the same normalised scale, summed once into a composite that spans the full band. That restraint is a virtue and it has a consequence here. A metric scored in the wrong direction does not merely fail to help; it actively pulls the composite the wrong way by its full range, so a month's score can be displaced by a substantial fraction of the total scale. Displacement of that size is enough to move a month across a regime boundary, which changes the main label, which changes the legal sub-label set, which changes the playbook entry the operator is directed to. A single sign error therefore does not produce a slightly wrong reading. It produces a coherent, well-formed instruction set for a month that did not happen.

How it gets settled

The question is empirical, and it is answered by reading the formula.

The resolution does not require deciding which philosophy of compression is superior, and framing it that way is how such questions stay open for years. It requires establishing what the source column actually computes — whether the formula populating the compression score produces a measure of constraint applied or of freedom preserved. That is a fact about a formula, checkable in minutes, with one right answer. Once established, the threshold band either matches the direction or it is inverted, and the resolution is documented alongside the change so the next reader inherits the answer rather than the question. The instruction attached to the flag is exactly this: confirm which quantity the column represents, and invert the threshold logic if necessary. Confirm, then invert if needed — not debate.

Why disclose it

A documented open question is safer than a silently chosen default.

The engine could have shipped with the ambiguity resolved one way and no note. It would have looked more finished, and every month would have carried a classification with the same visual authority as it has now. What the disclosure buys is that the uncertainty is attached to the thing it affects. An operator reading a borderline month knows there is one input whose direction is unconfirmed and can weigh the label accordingly, rather than treating a possibly-inverted composite as settled fact. This is the same principle the maturity rule applies to time — the system is built to say when it does not yet know something — extended to calibration. A tool that discloses its unresolved assumptions can be checked. One that hides them is simply confident.

• Five metrics have directions that follow from their definitions; one does not.
• Equal weighting makes a single inverted sign worth a sixth of the composite.
• The fix is to read the source formula, then invert the band and document it.

The key idea

Threshold tables encode beliefs, and beliefs should be inspectable.

The cutoff bands look like configuration and are the engine's entire theory of what a good month is. That theory has to live somewhere, and putting it in one visible table rather than distributing it through formulas means it can be read, argued with and corrected. The compression question is what that visibility looks like when it catches something: a disagreement between two parts of the workbook became findable precisely because both parts had to state their direction explicitly. The same conflict inside two hard-coded formulas would have produced two different answers for years with nothing to notice.