01 / 10
Retrieval
It already knows what you tried last time.
Before any answer, the incident becomes a search: its title and description, plus equipment, process, tool and product context. One pass reads all three layers at once, weighted by trust — and the result says which layer answered.
1A deviation arrives as a question, not a search box.its title and description carry the query
2Context facets from your master data lift matching candidates.soft boosts, never hard filters
3One search reads all three layers, weighted by trust.your own resolved history carries the highest weight
4The ranking names which layer answered.and where each match came from travels with it
5Your own cases carry the highest trust weightweb sources rank last, and say so
6A weak match is served flagged, never padded ina zero result is shown as zero, not filled with near-misses
Knowledge retrievalDraw St. 3 edge split · illustrative example
Searching knowledge base…
The incident, as asked
Edge split at draw station 3
STP-02 · die D-118 · HSLA 340 · coil C-88213 · 1.6 mm
Context facets from your master data
Technology · StampingEquipment · die D-118Process · STP-02Product · P-1160
Context facets read from your master data lift matching candidates. The ranking is deterministic and explainable — your own experience weighted highest, with a per-row reason for every result — never a prompt’s whim.
Three layers · one search, weighted by trustweight descends 01 → 03
01Your own resolved casesyour data only
Your own resolved incidents and knowledge-base entries.
weighted highest2 matches
02A curated industry foundationshared, curated
Standards-based solutions curated for the industry.
weighted below your own cases1 match
03External sources, ranked lastlowest weight
Public references, curated for quality.
weighted lowest0 matches
Similar past cases
Ranked across your own resolved cases, a curated industry foundation, and external sources — your own experience weighted highest.
DEV-0455Edge split after die regrind on press 2Candidate
91%
Companylikely ≥ 56% · N=11
DEV-0319Blank edge burr drift · incoming coil lotsCandidate
78%
Companypoint 85% · 90% CI 71–95% · N=24
KB-2207Draw die clearance control noteCandidate
74%
IndustryRelated guidance · illustrative
A weak match is served flagged, never padded in silently
Weak matches are never mixed in unlabelled. If nothing strong is found, the nearest matches are served openly flagged as weak, with their real scores — a zero result is shown as zero, not filled to look fuller.
5
6
02 / 10
Incidents
Edge split at draw station 3.
The same failure has happened twice before. Retrieval ranks what your plant already resolved, and every match carries its recorded applications and any detected recurrences.
1The incident opens with its impact already quantified.downtime, cost, units, recurrences
2Beside it, the prior cases: the action taken, and what happened after.
3The marker is the lower bound — the number decisions usenot the 85% point estimate
4N=0 shows an empty track, not a zerono applications yet, so nothing to score
Incident detail + retrievalIncidents · product screen · illustrative data
Incidents/Edge split at draw station 3
Edge split at draw station 3
CriticalOpenReported 29 Jul 2026, 06:12Has 2 recurrences
Downtime
145 min
unplanned
Estimated cost
€14.3k
lost output + repair
Units lost
3,140
affected units
Recurrences
2
linked cases
Similar past cases
Ranked across your own resolved cases, a curated industry foundation, and external sources — your own experience weighted highest.
DEV-0455Edge split after die regrind on press 2CandidateCompanyComputed
Relevance
91%
Occurred
14 Mar
Source
Press shop · STP-01
Matched
Draw radius cracking
Problem
Edge splits at the draw radius after a scheduled die regrind.
Root cause
Die clearance left above drawing tolerance after regrind.
Corrective action (D5)
Re-grind to drawing and re-qualify the first 30 blanks.
Evidence
Clearance measurement record, regrind log, 30-blank re-qualification run.
Effectiveness
likely ≥ 56%
≥56% BOUND
0%50%100%
point 77% · 90% CI 56–93% · N=11
conservative bound · 95% credible lower bound
2 recurrence(s) detected in 11 recorded applications.
DEV-0319Blank edge burr drift · incoming coil lotsCandidateCompanyComputed
Relevance
78%
Occurred
03 Oct
Source
Blanking line · BL-04
Matched
Coil property drift
Problem
Burr height on blanked edges widened across two incoming coil lots.
Root cause
Incoming coil at the low end of the elongation band.
Corrective action (D5)
Quarantine the lot and widen incoming inspection sampling.
Effectiveness
likely ≥ 71%
≥71% BOUND
0%50%100%
point 85% · 90% CI 71–95% · N=24
conservative bound · 95% credible lower bound
KB-2207Draw die clearance control noteCandidateIndustry
Relevance
74%
Occurred
—
Source
Curated industry reference
Matched
Die maintenance control strategy
Problem
Die clearance drifts undetected between regrind cycles.
Root cause
No clearance verification step between regrinds.
Corrective action (D5)
Add clearance verification to the regrind release check.
Effectiveness
No application data yet
0%50%100%
This fix has not been applied and tracked yet, so effectiveness cannot be estimated.
3
4
03 / 10
Root cause confirmed
Six branches, four candidates.
Six branches, four candidates, one confirmed cause. Two branches survive the evidence; the rest stay on the record rather than disappearing.
1Six 6M branches, no history required.
2Guided questions, asked one at a time.
3Four hypotheses — AI, historical, engineer, provisional.
4One cause confirmed, challenged against your own history.
5Two branches survive into the confirmed causematerial and method, traced
6Provisional differs in kind, not in shadehatched, and excluded from the floor
Ishikawa 6M & hypotheses8D session, Ishikawa selected at D4 · analysis output, not a screen · illustrative data
Ishikawa · 6M
Guided investigation questionsno history required
asked one at a time · up to 8 per session · earlier in the session
Has die clearance at station 3 been measured since the last regrind?
What changed in coil supplier / lot vs. the last good run?
Is blank-holder pressure logged per shift on STP-02?
H1 · AIMaterial issue
Low-elongation coil lot C-88213 reduced formability at the draw radius.
72% confidenceA50=19% vs 24% nominal
H2 · HistoricalMissing standard
No clearance re-check specified after a die regrind on D-118.
68% confidence↗ Similar RCA (score=0.87)
H3 · EngineerStandard not followed
Blank-holder pressure set 8% below the control plan value.
54% confidenceshift log, STP-02
H4 · Provisional — AI unavailableHuman dependency
In-die inspection relies on manual checklist completion — no system verification before release.
38% (provisional — not an AI assessment)
excluded from the evidence floor
Root cause confirmed
Excess die clearance at station 3 (post-regrind) combined with a low-elongation coil lot → edge split.
5
6
04 / 10
RCA workspace
Every step, unlocked by evidence.
Steps open when their work exists. The board orders the analysis; the gate downstream keeps the report from outrunning the evidence.
1Three of ten phases done — the board marks only what it can verify.
2Step 6 opens the moment a hypothesis exists.
3The triage record stays attached — marginal Gauge R&R included.
4Locked until a hypothesis existsa data dependency in the workspace, not a role check
5Every locked step shows the same refusal“Locked — complete previous steps first”
RCA workspaceRCA Sessions · stylized · illustrative data
Progress3 of 10 phases complete
1 · Triage & Containment▾
Incident data and initial containment actions
2 · Evidence Capture▾
Photos, measurements, and on-site observations
3 · Is / Is-Not AnalysisDone
Structured comparison to isolate the failure
4 · Change Point TimelineDone
Changes in 72h before the incident
5 · HypothesesDone
AI-generated and engineer-added root cause candidates
4 candidates · 1 provisional
6 · Evidence per Hypothesis
Locked — complete previous steps first
Targeted evidence collection for each hypothesis
7 · Root Cause Declaration
Locked — complete previous steps first
8 · Actions & Control Plan
Locked — complete previous steps first
9 · Closure Record
Locked — complete previous steps first
10 · Lessons Learned Review
Locked — complete previous steps first
1 · Triage & Containmentopen
Title
Edge split at draw station 3
Equipment
STP-02 · die D-118
Occurred at
29 Jul 2026, 06:12
Shift
Night · C
Downtime
145 min
Affected units
3,140
Lot
HSLA 340 · coil C-88213
Containment
Station 3 held · 100% edge inspection
Measurement System Analysis
MSA ValidatedValidated
Detection MethodVisual inspection
Gauge R&R24.0% (marginal)
4
5
05 / 10
Quality gate
The gate you cannot talk past.
Three outcomes of the same review. The button count is the argument — the evidence floor’s blocked state has no user-facing override.
1One score, one gate at 84 of 120.
2A · Normal — two ways forward.
3B · Evidence floor — one, and nothing else.
4C · Knowledge grounding — two, one of them “proceed anyway”.
5One button. No user-facing override.the middle panel is the product working
Quality gateRCA Sessions · product screen · illustrative data
RCA quality score96/120
0gate 84120
A · Normal2 actions
Validated root cause backed by direct evidence. Knowledge items retrieved and engaged with.
Evidence chain
Clearance measurement · post-regrindCoil certificate · A50=19%Prior case · Similar RCA (score=0.87)
Approve & generate reportRequest revision
B · Evidence floor1 action
A report cannot be generated until at least one validated root cause is backed by direct or circumstantial evidence (ADR-V3-04). Return to analysis to add evidence.
Excluded from the floor
H4 · Provisional — AI unavailable
Return to analysis
C · Knowledge grounding2 actions
Match-grade knowledge items were retrieved for this incident but the analysis shows no engagement with them. Review the retrieved knowledge before generating the report, or proceed if this analysis intentionally departs from prior history.
Retrieved, not cited
DEV-0455 · relevance 0.91 · match grade ≥ 0.82KB-2207 · relevance 0.84
Review retrieved knowledgeProceed anyway
Degraded AI output in this sessiondisclosed, not hidden
The AI service was unavailable for part of this analysis. Template content is labeled “Provisional — AI unavailable” and provisional verdicts are excluded from the quality-gate evidence floor.
5
06 / 10
Process capability
Every number comes from an engine.
Two parameters, one engine. One is measured well enough to judge. The other is not, and says so instead of printing a number.
1Twenty-four readings of bolt torque at station 4.
2One point crosses +3σ — the only rule that fires.
3The decision number is the interval bound, not the estimate.Ppk lower bound 0.92 against a 1.33 target
4With four readings the same engine refuses.
5Control limits amber · spec limits redboth dashed, never the same colour
6n=4 draws no limits at allnot established is not zero
Statistical evidenceQuality · product screen · illustrative series, recomputed with the shipped engine
CapableNot capableNot establishedNo spec limit
SPC capability evidence
Computed capability and control-chart signals over the pre-incident window. The decision number is the Ppk 95% CI lower bound (conservative).
Ppk lower bound
0.92
95% CI lower bound — the decision number · point estimate 1.33 · target 1.33
Not capablebelow the 1.33 capability target (in-spec but not capable)
I-chart · bolt torque Nm · station 4 · last 24 units
Nelson run-rules tripped
Rule 1 · 1 point
Limit source
Process-anchoredCapability not established (n=4)
Blank-holder pressure · STP-02
Control-chart baseline not established (n<20)
n = 4 — not enough data. The engine refuses to report a capability index instead of guessing one.
not 0.00 · not a dash · not a tick
The thesis
The interval decides, not the point estimate — the point lands exactly on 1.33 and the lower bound 0.92 still refuses.
5
6
07 / 10
Process capability · Ask SPC
Ask the statistics a question.
Ask in plain language, read the answer in two layers. The prose is phrasing; the block beneath it is the engine, and it is the authority.
1One question, in plain language.
2It becomes a deterministic engine call over twelve parameters.
3One answer, two refusals — rendered identically.
4The engine call sits above its own answerprose phrases it, the engine authors it
5The refusal gets the same two layersequal weight to an answer
Ask SPCAll active parameters · most recent 30 readings each · illustrative conversation
You
Which parameters are trending toward a limit?
SPC assistant
Engine call — find_parameters_trending_toward_limitdeterministic
One of the twelve parameters is projected to cross a limit within the horizon. Two others cannot be answered honestly yet — here is why for each.
REF-04 · zone-3 peak tempHeading for UCL
Theil-Sen slope +0.31 °C per reading · projected crossing inside the pre-stated horizon
PRESS-02 · seal forceNo reliable trend
Slope interval spans zero — the engine declines to project
MILL-11 · bore ØInsufficient history
n below the reliability floor — no answer offered
e.g. Which parameters are trending toward a limit?AskThinking…
The engine callexpanded
find_parameters_trending_toward_limit
scan · 12 parameters · theil_sen slope + 90% CI
get_trend_projection ×2
per-parameter drill-down · returns the named refusals
The assistant chooses which shipped tool to run; every statistic comes from the deterministic engine.
If its prose states a figure the engine did not return, the prose is withheld and only the engine output is shown.
The refusal has equal weight
Two of the three verdicts are refusals, rendered exactly like the answer — same card, same engine line. Computed, never guessed.
4
5
08 / 10
Quality
You stop asking. It watches on a schedule.
Monitoring stays off until you turn it on, one workspace at a time. A routine is a scope, a cadence and a set of checks; once enabled it runs on its cadence and records an outcome every time — including the quiet ones.
1A routine starts with a scope — a line, a parameter, a set.
2Five checks run together.
3On a cadence you set — every shift.
4The last run found two things.one of them a parameter that stopped reporting
5The detector most dashboards missa sensor that quietly stopped reporting
6Quiet is a recorded outcome, not silencedifferent from No runs yet
MonitoringQuality · product screen · illustrative data · off by default, enabled per workspace
Active
New monitoring routine5 checks
Routine name
SMT reflow · morning watch
What should it watch?
A parameterAn equipment itemA production lineA parameter set
Scope reference
SMT-01
Line name, e.g. Assembly Line 3.
What to check
Baseline breaks
Alert when readings break the control rules measured against a locked baseline.
Small drifts
Catch small, sustained shifts away from target before they show up as obvious spikes.
Trend to limit
Warn when a steady trend is heading toward a spec limit.
Capability
Flag when a parameter's spread no longer fits comfortably inside its spec window.
Stalled data
Alert when a parameter silently stops receiving data.
How often
HourlyPer shift (8h)DailyWeekly
Notify me when a run finds something
On/ OffCreate routine
SMT reflow · morning watchActive
Scope
line: SMT-01
Cadence
Per shift (8h)
Next run
28 Apr 2026, 05:52
QuietFindingsErrorNo runs yet
Last run · 2 findings21:47 · 12 parameters resolved
zone-3 peak temp — baseline break · zone-1 soak — Expected ≥1 accepted measurement within the last 480 min, got 0 (last measured: 26 Apr, 23:04)
Run nowView runsPause
Vial neck monitor
Checks
5 checks
Last run
Not run yet
State
PausedNo runs yetan emptier state than Quiet
5
6
09 / 10
Signals
The circuit.
Four stations, and the loop closes on the next investigation. A signal proposes work, a reviewer opens it, the work becomes knowledge — and the knowledge changes what the next investigation retrieves. The engineer never re-derives the statistics by hand.
1A signal crosses the trigger threshold.
2The investigation inherits the chart’s evidence.
3The fix is measured, not assumed.
4It becomes knowledge with a statistical fingerprint.retrieved by the next investigation, never auto-applied
5The return arc is the heaviest lineknowledge changes what the next investigation retrieves
6Cited, never applied — a person still decides
SPC ↔ incidents ↔ knowledgesystem behaviour · not a screen · illustrative data
Reuse decision at ①
Confirm resolves as Incident created — opening the investigation or Linked to the existing open incident — opening the investigation.
Reuse decision at ④
Cited (never applied) when a future incident shows a similar signature — evidence, not automation.
History
Proposals a reviewer has already confirmed or dismissed.
ConfirmedView incident →
Nelson run-rules fired · Rule 2
Injection temperature · IM-07
Dismissedrecorded
Capability below target · No run-rule trip (capability signal)
recorded, not deleted · 24 Jul
Why the ring closes
Slide 1 retrieved these entries. This is where they came from.
5
6
10 / 10
Refusals · stated, never guessed
The refusals.
Eight states the product refuses to guess, each naming its own threshold — and beneath them, the one refusal with no user-facing override at all.
1Eight states where the product declines to answer.
2Each names what a lesser tool would have printed instead.and beneath them, the refusal with no user-facing override
3A button-shaped void, on purposethe absence is the design
Refusal statesproduct copy · not a screen · surfaces, labels and thresholds are the shipped product; incidents, part numbers and measurements are illustrative
Capability not established (n=3)
InsteadCpk 1.00, printed with full confidence from three readings.
Control-chart baseline not established (n<20)
InsteadControl limits drawn from a dozen points, flagging noise as signal.
No application data yet
This fix has not been applied and tracked yet, so effectiveness cannot be estimated.
InsteadA green “Effective” the moment the action is saved.
No reliable trend — the slope is not distinguishable from flat.
InsteadA projected crossing date, extrapolated from flat noise.
Zero variation in the recent history — this chart is not informative right now.
InsteadA flawless capability score from a sensor that never moved.
σ0
not assessable — zero variance observed across readings
InsteadA tidy 0.00, hiding the divide-by-zero.
Provisional — AI unavailable
72% (provisional — not an AI assessment)
excluded from the evidence floor
Instead72% shown as real AI confidence, with no caveat.
n
Insufficient (n<5)
InsteadA number anyway, with the sample size left off the screen.
Evidence floor not met
There is no button here — not disabled, not hidden behind a plan. A report cannot be generated until at least one validated root cause is backed by direct or circumstantial evidence (ADR-V3-04). Return to analysis to add evidence.
The surfaces, labels and thresholds in this tour are the shipped product. The incidents, part numbers and measurements are illustrative.
Disclosure
Ranked over the 50 most recently measured of 214 parameters
Disclosure
Knowledge base status unavailable
We could not check the knowledge base right now, so similar-case coverage is unknown. This is a temporary degraded state, not an empty corpus.
3
End of the tour
See it on your own data.
The surfaces, labels and thresholds in this tour are the shipped product — including every refusal. The incidents and measurements are illustrative. Bring yours and we will run them through the same engine.
No preparation needed — a recent deviation and its measurements are enough.
Tap anywhere to advance.
Each tap reveals the next point and lights the part of the frame it describes. Nothing plays on its own — you set the pace.
← →step back and forwardhover a block to inspect it
This note goes away once you start