# Pattern Translation — calibration revision

Revision: qa-2026-09-14-calibration. Pattern names are examples, not exclusive answer keys.

## Observable rules

M is the median high-low range of the last 20 shown bars; a marked forming bar is included when present. Each chart has its own 1 M ruler. The vertical divider marks the evaluation window.

Compression compares four-bar high-low envelopes: the mean over the last third must be at most 0.5 times the 80th percentile of the earlier envelopes. Expansion compares the first half mean with the last third mean, at the same 0.5 threshold, and takes precedence as Neither in this drill. These describe observed range behavior; neither predicts a future move or trade validity.

A failed break uses a level with at least one rejection and two total touches, a later push beyond 0.30 M and the last close back on the original side. Two consecutive closes more than 0.10 M beyond the line establish acceptance for that visit, except a return gap over 0.5 M. Actual positions and stops remain unknown.

Swing extremes within 0.15 M are clustered. For compression, a qualifying flat edge needs at least two rejections at the outer edge of the last two thirds. All levels meeting the mechanism rule are accepted within 0.5 M click tolerance. If none qualify, No level is the answer. The highlighted reference is one example, not the only permissible line.

## Three tasks

- Translate: classify Compression, Failed break, Both or Neither and point at any qualifying level.
- Matching panels: the prompt states a mechanism and level side. Every panel satisfying both is accepted. Textbook labels can overlap.
- Trigger: the line is an explicitly supplied external reference. A completed close more than 0.10 M through it in the stated direction is Yes; a forming bar alone beyond it is Can't tell; otherwise No.

Unresolved visual evidence can also be recorded as an uncertain read, with a reason. This receives no correctness verdict and is excluded from confidence scoring; coverage is reported alongside accuracy. A single correct answer does not prove that low confidence was unjustified. Future data cannot change the grade.