Skip to content

An open source design system that's fully customizable and agent ready

TypeScript

docs(spec): AST-036 decides an over-budget FR23 path by a majority of nine samples (#7471)

Cindy Zhang2a70c7ba1bd2
diff --git a/docs/specs/AST-036/spec.md b/docs/specs/AST-036/spec.md
index 7ffc94e6c7..7e6ea10b94 100644
--- a/docs/specs/AST-036/spec.md
+++ b/docs/specs/AST-036/spec.md
@@ -227,7 +227,7 @@ Compatibility is explicit per plugin and tested version. “Remark-compatible”
- **Plugin-authored representative set:** three plugin-authored transforms that respectively rewrite matched prose text, wrap a section in an extension container, and read a preceding plugin's nodes under a declared dependency.
- **Streaming set:** the plugin-authored representative set applied to the 200-section fixture streamed in 50-character chunks, compared against the same stream with an empty pipeline.
- **Comparison:** omitted/empty pipeline versus the same parse-and-transform call with the stable plugin list, in one isolated process and runner.
-- **Sampling:** ten untimed warmups, then nine alternating paired rounds; each side averages 20 iterations per round. The reported ratio is the median of the nine paired ratios. Both 200- and 500-section ratios must satisfy FR23. An A/A control outside the measurement-validity bound invalidates that sample; the harness retries rather than certifying noisy output.
+- **Sampling:** ten untimed warmups, then nine alternating paired rounds; each side averages 20 iterations per round. The reported ratio is the median of the nine paired ratios. Both 200- and 500-section ratios must satisfy FR23. An A/A control outside the measurement-validity bound invalidates that sample; the harness retries rather than certifying noisy output. After a first valid sample over a budget, the stable and recreated lists are each decided separately, per budget and fixture size: a list fails when five of up to nine fresh-process valid samples, the first included, are over the budget.
- **Cost-shape evidence:** guard traversal counts are asserted directly, so adding a transform may add at most one output traversal and may not repeat whole-document invariant collection.
- **Regression signal:** invoking a helper callback for unclaimed source, invoking a claimed helper more than once, regressing parse/result reuse, or publishing a number without naming its path must fail work-count or threshold evidence.