Maintenance wants the backlog re-sorted by raw dollars: 'Confidence is just the AI hedging.' Top two items: a seal failure at $65k/yr, confidence 0.30 — two logged events, no sensor on the seal. Slow cycles on the case packer at $38k/yr, confidence 0.90 — months of dense cycle data.
Q1 What's the right sort, and why?
A
Raw dollars: $65k is $65k — chase the seal failure and let confidence sort itself out. Two logged events and no instrumentation make the $65k closer to a guess than a measurement. That's optimizing the loudest loss — exactly what EV ranking exists to prevent.
B
Expected value: $38k × 0.90 ≈ $34k beats $65k × 0.30 ≈ $19.5k — the case packer goes first, and confidence measures verifiability, not hedging. Confidence reflects sample size, sensor coverage, and pattern stability. A $65k figure resting on two events may never verify — EV puts scarce Kaizen capacity where savings are real and provable.
C
Split the team across both so neither number is ignored. Kaizen capacity is scarce — halving it on a low-confidence item buys less than finishing the high-EV one and instrumenting the seal so its next ranking is trustworthy.
Covered in Module 01
Line 5 loses about 40 minutes a shift to changeovers running past plan, and between changeovers its cycle time runs 8% over ideal.
Q2 Which Big Loss factors are bleeding, and what data feeds each?
A
Setup/Adjustment (availability), fed by changeover reason-code logs — and Reduced Speed (performance), fed by cycle-time delta versus ideal. Overrunning changeovers land in the Setup/Adjustment availability factor; running 8% over ideal cycle is the Reduced Speed performance factor. Each traces to its own data source.
B
Startup and Production Rejects — lost minutes are lost output either way. No parts are being rejected in this picture. Folding time losses into quality factors breaks the taxonomy the ranking depends on.
C
Breakdowns and Minor Stops — the line keeps stopping and slowing. Neither. Planned changeovers overrunning aren't breakdowns, and a steady 8% speed gap isn't a minor-stop pattern. Misbucketing aims the Kaizen at the wrong fix.
Covered in Module 01
An improvement card reads: 'The palletizer feels sluggish lately — someone should look at it.'
Q3 What turns this into a backlog item KaizenFlow can rank?
A
Enter it as-is with a conservative dollar guess so it isn't forgotten. A guessed number sorted against measured ones corrupts the whole ranking. The backlog stays trustworthy only if nothing on it is an opinion.
B
Route it to maintenance — feelings about equipment are their department. Routing doesn't answer whether a loss exists. The cycle-time data does, in minutes, before anyone's time is committed.
C
Check the palletizer's cycle-time delta versus ideal. If the data shows a real Reduced Speed loss, it enters with dollars and confidence; if not, the card dies. Every backlog item traces to a measured factor. The feeling is a prompt to look at the measurement — the measurement is what gets ranked.
Covered in Module 01
Rejects on Lines 1 and 2 spike on the same days, and each line's reject series correlates strongly with the other's. A teammate concludes Line 1's scrap is somehow causing Line 2's.
Q4 What's the sharper hypothesis to test first?
A
Line 1 causes Line 2 — the correlation is strong and consistent. There's no plausible mechanism by which one line's scrap makes another line scrap. Strength without mechanism is exactly the trap.
B
Whichever line spiked first each day is the cause of the other. Time order can't separate two effects of a common driver — one will always lead by minutes for boring reasons. Test the shared inputs instead.
C
A common driver hitting both lines — the same material lot, ambient conditions, or a shared upstream process — moves both series at once. Two effects of one cause correlate with each other. The lurking-common-driver check exists precisely for this shape: test the shared inputs before inventing a line-to-line mechanism.
Covered in Module 02
The top correlate for a weld-defect spike is plant compressed-air pressure, r = 0.78. Your weld engineer says pressure swings would plausibly change torch cooling — and proposes a 30-minute test: run the torch at controlled low pressure and inspect the welds.
Q5 What does the mechanism-plus-test combination buy you that the 0.78 alone doesn't?
A
It converts a statistical lead into a testable causal claim — if low pressure reproduces the defect, you've confirmed the branch before spending on a fix. Correlation points; the mechanism explains why; the small confirmation test is the corroborating second signal. That's the full confirm-before-you-act sequence.
B
It proves causation outright, so the countermeasure can skip verification. A mechanism makes the claim plausible and the test makes it confirmed — but the countermeasure still gets its own before/after. Nothing skips verification.
C
Nothing — at 0.78 the correlation was already actionable. A 0.78 can still be a confounded proxy. Acting on it without mechanism or confirmation is how dehumidifiers get bought for resin problems.
Covered in Module 02
One week into the after-window for your guide-rail fix, scheduling moves a heavier product mix onto the line. Minor-stop minutes tick up, and the experiment suddenly looks like a failure.
Q6 What's the right handling?
A
Do nothing — after means after, and the numbers are the numbers. Ignoring a known confounder lets the mix change masquerade as the fix failing. The comparison must be like-with-like or it isn't a comparison.
B
Log the mix change against the window, and extend the after-period until it covers mix comparable to baseline — attribution depends on comparable windows. The change boundary is still clean; the windows no longer match. Documenting the confounder and restoring comparability keeps the experiment honest in both directions — no false failure, no false win.
C
Stop the experiment and book the first clean week as the result. One week cherry-picked before the mix shift is a single-good-day baseline in reverse. A finance-grade claim can't be built on the convenient slice.
Covered in Module 03
You pre-stated that the SMED change should cut setup time by at least 8 minutes. The verified result: 2 minutes. Production wants it announced as a win — 'improvement is improvement.'
Q7 What's the honest call?
A
Quietly restate the expected effect as 2 minutes so the record shows a hit. That's bookkeeping against your own experiment design. The ledger's value rests on bars that don't move after the fact.
B
It's a miss against the pre-stated bar. Treat it as one — dig into why the effect came in small before deciding whether to iterate or move on. The bar was set before go-live precisely so a too-small result couldn't be rationalized afterward. Two minutes may be noise, partial adoption, or a wrong mechanism — the miss is information.
C
Announce the win — 2 minutes times every changeover still adds up. Re-grading the result after seeing it is exactly the post-hoc rationalization the pre-statement exists to block. Next time, no bar will mean anything.
Covered in Module 03
A downtime fix shows a gross delta of $54k. The after-window happened to run a lighter product mix that historically stops less on its own. The ledger's normalized delta is $37k, confidence weight 0.80.
Q8 What figure does finance sign, and why is it smaller than $54k?
A
$54k — the stops really did drop; the rest is statistical pessimism. Part of that drop belongs to the mix, not the fix. Claiming it invites the exact challenge normalization exists to pre-empt.
B
$37k — normalization is required, but the confidence weighting is optional color. The confidence weight is part of the sign-off figure by design. Skipping it hands finance a number the ledger itself doesn't stand behind.
C
About $30k. Normalization strips the improvement the easier mix would have delivered anyway; the 0.80 discounts for how cleanly the change was isolated. $37k × 0.80 ≈ $30k. The sign-off figure is deliberately conservative: conditions-adjusted, then certainty-weighted. That's the number that survives challenge.
Covered in Module 04
Finance pushes back on a $22k ledger entry: 'A new supervisor changed the scheduling in that window. How do I know this isn't their doing?'
Q9 What makes the claim defensible?
A
The entry's attribution note and normalization detail — it should already name the scheduling overlap and show what was adjusted. If it can't, the claim isn't ready to present. A defensible entry answers the shared-credit question before it's asked: what was normalized, what was discounted, and what else changed in the window. That's the standard the ledger is built to.
B
Verified means verified — finance shouldn't relitigate the ledger. Finance's challenge is the test the number exists to pass. A claim that can't survive the shared-credit question was never finance-ready.
C
The sensors measured the delta directly, so the number speaks for itself. Sensors measure the change in the metric, not who caused it. Attribution is exactly the part raw measurement can't settle.
Covered in Module 04
Your capper fix verified at $31k/yr on Line 1. Line 6 runs the same capper model, but its losses are dominated by changeovers, not minor stops. Line 8 runs a different capper with a near-identical minor-stop signature to Line 1's.
Q10 Where does the rollout go first?
A
Line 8 — the loss signature matches what the fix actually attacks. Pilot there with its own baseline and ledger entry before booking anything. Replication targets the loss profile, not the nameplate. Line 6's same-model capper isn't bleeding the loss this fix stops — and every site verifies fresh, because effect size varies.
B
Line 6 — same machine model means the fix transfers directly. The fix cures minor stops; Line 6's problem is changeovers. Same hardware with a different loss signature is where copy-paste rollouts go to die.
C
Both simultaneously, booking $31k per line up front. Effect size varies by site, and one of these sites doesn't even have the disease. Savings transfer as a hypothesis — the ledger books them one verification at a time.
Covered in Module 05
The AI proposes a revised bearing-lubrication schedule: $45k/yr at confidence 0.90. The data spans three weeks from a single crew — but your maintenance lead can explain the mechanism cleanly, and it matches the bearing vendor's wear data.
Q11 Mechanism is solid, window is thin. What's the move?
A
Reject it — any recommendation on under a month of data is spurious by definition. Thin data is a flag, not a verdict. With a solid mechanism and independent corroboration, the right response is to widen the evidence, not bin the lead.
B
Implement plant-wide — 0.90 confidence plus a mechanism clears every check in the book. The stress-test asks whether the pattern survives different crews, mixes, and windows. Three weeks of one crew hasn't answered that, whatever the score says.
C
Treat it as a promising hypothesis: extend the window across crews and mixes — or run a bounded pilot on one line — before committing the full spend. A named mechanism plus vendor corroboration earns the recommendation a test, not blind adoption. High confidence on three single-crew weeks is still thin — verify it survives other conditions first.
Covered in Module 06
To 'remove bottlenecks,' a sister plant auto-implements any AI recommendation above 0.85 confidence — no specialist review. Two quarters in, they've spent on three recommendations that turned out to be coincidences in the data.
Q12 What's the design flaw?
A
The AI needed more training data before any recommendations were used. More data helps, but the failure mode is structural: spend was gated on a score instead of on human sign-off. That flaw survives any amount of data.
B
The threshold was just too low — raising it to 0.95 fixes the process. A spurious correlation on a lucky window can score arbitrarily high. No threshold substitutes for the mechanism-and-confound check.
C
The gate removed the human. Confidence scores measure pattern strength in the data the model saw — only a specialist can check mechanism, data thinness, and confounds before dollars move. The AI ranks and proposes; the specialist verifies and decides. Search enough signals and some clear 0.85 by chance — that's exactly the judgment a score threshold can't apply.
Covered in Module 06