Imported from shmindmaster/assay (skills/scrap-variance-rollup/SKILL.md). Install upstream with npx skills add shmindmaster/assay --skill scrap-variance-rollup. Copyright stays with the author.

Scrap variance rollup

Turn CASTLINE scrap rows into the weekly rollup.

Inputs

Only mcp__kestrel-foundry-data__get_scrap_records (8 weeks max). Weeks end on Sundays: 2026-06-14 through 2026-08-02.

Math

1. Group rows by cell + cause_code. Sum cost_usd (and weight_lbs when asked).
2. Current week: 2026-08-02. Baseline: mean of the trailing 4 weeks — 2026-07-05, 2026-07-12, 2026-07-19, 2026-07-26.
3. variance_pct = (week - avg) / avg * 100, rounded to one decimal.
4. Top mover: the largest positive variance_pct whose current-week cost is above $5,000. Below $5k a big percentage is noise — say so instead of inflating it.
5. Cause text comes verbatim from references/02-cause-codes.md (e.g. P02 = "Gating/risering design inadequate for section thickness").

Rules

• No data from anywhere but the tool. No estimates, no smoothing, no forecasts.
• Known pattern: Molding P02 rises across the last 4 weeks of this dataset. Report the computed numbers, not the expectation.
• Negative variance is an improvement; label it as such, don't hide it.