V-AC Axis Split

Overview

A verification-AC group must not commingle checks that pass by invariant rule with checks that pass by observed rate. The two axes answer different questions; mixing them yields false confidence at the quality-review step.

The Pattern

When a single group contains both deterministic entries (for example, «schema matches XSD», «HTTP status is 200») and statistical entries (for example, «p99 latency below 500 ms», «error rate below 1 % over one hour»), split into two separate groups before finalising the wishlist.

Why Split

Deterministic and statistical validity rely on different evidence chains. A rate-based criterion grouped with rule-based checks masks that the deterministic axis always passes while the statistical axis is the real uncertainty source.

How to Apply

1. Tag each entry as deterministic (unambiguous rule) or statistical (threshold over a window).
2. If both axes coexist, split into a deterministic group and a statistical group.
3. A statistical entry must cite the measurement window, sample size, and confidence interval.
4. A deterministic entry must be tied to a spec-runner assertion or an equivalent monotonic check.
5. When an entry is split, preserve or reassign every Covers:, Verifies: and Evidence: binding so the automatic spec graph stays complete.

Pattern 2: Gate-Activation Dry-Run

When a task ships a new validator gate or flips one from advisory to fail-hard, /dr-plan runs the gate in dry-run over its --scope and lists every failing file. Each is either rewritten in-task or removed from the scope by a documented waiver — never deferred, because the gate would otherwise reject the very archive that activated it.

Pattern 3: Threshold Gates

A threshold gate (abstention, confidence cutoff, rate limiter) gets two entries: a deterministic one proving it reads the right field and applies the right constant (unit test), and a statistical one proving the resulting rate lands in a documented band (eval or shadow script over a representative sample).