Impact Quantification Rubric (Locked)
Every recommendation cites a specific quantified impact. Hedging is banned by HC#3.
Estimation Methods
D1 Incremental Refresh
- Sample the candidate key: MIN, MAX, and the "recent-N-days" count via
get-summary-data. - Estimate
new-rows-per-refresh-intervalfrom the recent-day sample. - Impact: refresh-time reduction ≈
1 - (new-rows / total-rows). - Reporting: "Estimated refresh-time reduction ~72%: 8.3M new rows per day out of 300M total."
D2 Aggregation
- Identify the intersection dimensions used on every sheet.
- Sample
COUNTDacross those dimensions to estimate post-aggregation row count. - Impact: row-reduction ≈
(post-aggregation-count / pre-aggregation-count). - Reporting: "Estimated row reduction ~89%: 300M rows → ~33M rows after aggregating on {Product Category, Region, Order Date}."
D3 Hidden-Field Pruning
- Count total fields vs. unused fields.
- Impact: file-size reduction ≈
(unused / total)percentage. - Reporting: "Estimated file-size reduction ~38%: 47 of 124 fields never placed on any sheet."
- Caveat: not exact because some fields (strings, high-cardinality dimensions) are wider than others. Reporting includes the caveat.
D4 Filter-at-Extract
- Sample
get-summary-datacount with and without the candidate filter. - Impact: row-reduction ≈
(1 - filtered-count / total-count). - Reporting: "Estimated row reduction ~62%: 187M rows drop to 71M after applying [Order Date] >= 2024-01-01 at extract time."
What Counts as Quantified
- Any percentage (row reduction, file-size reduction, refresh-time reduction).
- Any absolute count (fields pruned, rows removed).
- Any ratio comparing before and after.
What Does NOT Count
- "Should improve performance."
- "Likely faster."
- "Could be significant."
- "Depends on your data."
Sample-Size Caveats
- Estimates for D4 (filter-at-extract) depend on the sampled data. If the sample is representative, the estimate is close. On extract-configured sources with existing filters, the sample already reflects those filters.
- Report the caveat if the sample size is smaller than a reasonable threshold (< 10,000 rows for large tables).
Refuse to Ship Without Numbers
If a finding cannot be quantified (e.g., because the tool surface fails or the sample returns zero rows), drop the finding rather than shipping it with hedged language.