Decision D5 · 5 of 21
Value ledger method per KPI
Switchback where actions alternate, adjusted baseline otherwise, power analysis first
RECOMMENDED
- Status
- RECOMMENDED — awaiting Vinayak
- Options
- IPMVP-style adjusted baseline; ITS; DiD with untreated lines; synthetic control or CausalImpact; switchback
- Trade-offs
- Switchback gives the cleanest causal estimate but needs actions that can alternate; ITS is easy but weak under drift
- Recommendation
- Per action class. Switchback where an action can alternate (set-point trims, head trims, sequencing rules). Adjusted baseline with normalisers (output, mix, shift, ambient) otherwise. Before any test, a power analysis states the smallest effect the test can detect in the planned window; if that effect is larger than the expected effect, the record stays
modeled. - Why
- Switchback gives the cleanest causal estimate when the plant allows alternation; adjusted baseline is what L5 packs already do. The power analysis stops a test from promising what it cannot show.
- Evidence
- L5 verification packs (
cusum_decrease, normalisers); E34 (EVOP on running plants); E11 (benefit decay without upkeep). - Reversible
- Yes, per KPI and per action class.
- Blocks
- ValueRecord contract, any customer value statement.
Page history: last 2 changes
- docs(research): retire stale research to archive/research-2026-10 with a register
ab84821 - docs(technical): split decision board into DECISIONS.md
b4db9d4