Tier usage is graded against the gates that actually occurred.

The error being avoided

An average month is an average over gate histories, not a month.

The benchmark's unconditional tier distribution blends every gate history the simulated population produced — favourable years spent almost entirely in Growth, punishing ones spent in Buffer and Floor, and everything between. The resulting average describes no individual path's experience. Compare a live month against it and the comparison silently assumes the account's gate history resembled the population average, which it almost never does over a single month. The gap that appears is then attributed to the operator's aggression or timidity, when a substantial part of it is simply which capital states the month contained.

The instrument

Tier-by-gate and tier-after-cycle, not tier-by-month.

The conditioning is possible because the benchmark reports tier behaviour broken down by gate and by cycle position rather than only in aggregate, alongside the expected tier, pool, and per-trade risk for each gate. Those tables supply the conditional expectation directly: given that the account spent this proportion of the month in each capital state, this is the deployment profile the model predicts. The live profile is then scored against that construction. It takes longer than reading a single distribution and it is the difference between a comparison that means something and one that mostly measures the month's weather.

The two readings it separates

Correct compression and genuine timidity look identical until conditioned.

Low tier usage under a compressed gate history is the ladder working — the account was capped and complied, and there is nothing to discuss. Low tier usage under a Growth-dominant history is a different finding entirely: the authority was available, the deployment was not taken, and modelled expectancy went unmonetized. Both produce the same unremarkable usage figure. Only the conditional read distinguishes them, and the distinction determines whether the review generates no action at all or opens a conversion investigation. Grading without conditioning gets this wrong in both directions with roughly equal frequency.

The reverse case

Elevated usage in a favourable quarter is not automatically drift.

The conditioning protects the operator as often as it catches them. Heavy upper-tier deployment through a quarter spent largely in Growth is exactly what the model expects, because the gate authorised it and the tier allocator selected inside that authority. Judged against the blended average, the same behaviour reads as aggression drift and invites a correction the evidence does not support — which is a real cost, since an operator corrected for behaving correctly learns to under-deploy. This is why the page's framing is precise: heavy T5 usage during a Growth-dominant quarter is normal, and the identical usage during Buffer dwell is a governance breach in progress.

What conditioning cannot excuse

The gate history is context, not a defence.

One boundary is worth stating, because conditional reasoning can be stretched into an argument that no reading is ever adverse. The gate history explains what deployment was authorised; it does not explain deployment that exceeded the authorisation. Usage above the gate's own cap is a breach regardless of what the surrounding month looked like, since the cap is the ceiling for that capital state and conditioning is about what should have been expected below it. Nor does conditioning apply to conversion: whether the tiers deployed earned outcomes proportional to their risk is a separate question, asked against per-tier expectancy, and no gate history makes uncompensated elevation acceptable.

• Condition on realised gate dwell before scoring any tier gap.
• Low usage is only a finding when the authority to deploy was actually present.
• Conditioning never excuses deployment above the gate cap or elevation that failed to convert.

The key idea

Grade the behaviour against the month that happened.

A benchmark is a population of futures, and any single live month is one draw with its own particular gate history. Scoring the operator's deployment against the population's blended profile imports a month they did not have and grades them on the difference. The conditional read costs a few extra lookups and replaces that with a fair question: given the capital states this account actually occupied, did it deploy as the model says a governed account should? That is the only version of the question the operator could have answered at the time.