From No Data to Hardened: trusting a baseline exactly as much as it deserves.

The problem with young baselines

Percentile fences wobble violently at small N.

The five-zone classification stands on percentile thresholds, and percentiles are exactly the statistic small samples corrupt worst. With a handful of ATR observations, P80 is essentially whichever reading happened to be second-largest — one quiet week or one news spike relocates every fence, and with them every zone placement the key will issue. A live reading classified 'Extreme' against fifteen observations may be perfectly ordinary against two hundred. The failure is silent: the lookup still resolves, the zone still displays, the coefficient spectrum still computes. Nothing about a thin baseline looks thin at the point of use — which is precisely why the reliability grade has to be attached to the output rather than left for the operator to remember.

The ladder

Five grades, count-driven, attached to every read.

The Sample Status logic maps each key's observation count to a grade: No Data (the key has never been fed), Insufficient (readings exist but classification would be noise), Provisional (early structure — usable with explicit caution), Developing (the fences are settling), and Hardened (the baseline has earned production trust). The grade is computed live by the same COUNTIFS machinery that powers the baselines, so it can never drift out of date, and it displays beside the zone and ratio wherever the key's outputs appear. The operational translation is graduated, not binary: a Hardened key's zone read is a fact to act on; a Provisional key's read is context to weigh; an Insufficient key's read is a reminder to go collect samples before pretending to have a baseline.

Operating under thin status

What to actually do when the key hasn't earned trust yet.

Thin status doesn't halt trading — it reorders the evidence. Under Provisional or worse, the static classification and the VIP's structural view carry more weight than the key's own percentiles, the All-Sessions baseline (usually deeper than any single session's) becomes the better comparison population, and coefficient selection leans on the ladder's default tiers rather than fine-grained zone distinctions the fences can't yet support. Meanwhile the repair is mechanical: every session traded feeds the sample log, every logged reading moves the key up its ladder, and the status upgrade arrives on its own schedule. The one prohibited move is the tempting one — reading a Provisional key's 'Extreme' with Hardened confidence because the number happened to support what you wanted to do.

• Session-specific keys harden slower than All-Sessions keys by construction — three sessions split one sample three ways.
• A new symbol enters through the maintenance rule: Instrument Master first, baselines validated before live use.
• Status is per-key, not per-symbol: the same pair can be Hardened at 5M London and Insufficient at 15M Asia.

The pattern repeating

The same honesty gate as everywhere else in the stack.

The sample-status ladder is the volatility layer's copy of a discipline MARS enforces at every altitude: the EV lab's minimum sample gate treating thin-lens outputs as provisional, the attribution layer's patience with quota-limited variants, the benchmark's insistence on documented assumptions before comparison. In each case the mechanism is identical — outputs carry their own evidentiary grade, and the grade governs how much decision weight they may bear. Volatility baselines just make the need unusually vivid, because the numbers look so authoritative: a percentile is a percentile whether it stands on twelve observations or two hundred, and only the status column knows the difference.

The key idea

Trust is accumulated, displayed, and never assumed.

The matrix could have shipped with baselines that simply worked from day one, pretending authority it hadn't earned. Instead every key starts at No Data and climbs — and every reading it ever issues admits where the climb currently stands. That's the honest shape for a tool whose entire value is replacing impressions with placements: the placement is only better than the impression when the population behind it is real, and the status ladder is the running proof of exactly how real it is.