Should we replace our manual QA inspection line with machine vision? 200K units/month, 3 shifts, 0.2% defect rate, $400K budget for automation.
Do NOT do a full replacement. Deploy a Cognex In-Sight 9902 or Keyence CV-X480F machine vision system in PARALLEL...
Confidence
80%
Risk
unknown
670s
Deploy a Cognex In-Sight 9902 or Keyence CV-X480F machine vision system in PARALLEL with existing manual inspection...
Decision
Do NOT do a full replacement. Deploy a Cognex In-Sight 9902 or Keyence CV-X480F machine vision system in PARALLEL with existing manual inspection for 6 months. Install one machine vision station inline BEFORE manual inspection stations. Run both systems on 100% of units, logging machine vision pass/fail alongside human decisions for every unit. After ~1.2M unit comparisons, make a data-driven go/no-go on full replacement. Budget: $180K hardware (Cognex In-Sight 9902 + GigE camera + structured lighting dome), $40K integration, $30K ViDi deep learning training on 12 defect types, $50K reserve for lighting/model iteration. Total $300K, $100K buffer. Success thresholds: ≥99.8% agreement with human inspectors AND ≥95% of human-caught defects detected, validated over ≥500K units. Key failure mode: Structured lighting fails on subtle waviness or discoloration due to alloy batch reflectance variation. Mitigated by $50K reserve and manual inspectors as backstop. Keep all 9 inspectors during validation. No union disruption. Timeline: 6 weeks procurement/install, months 2-4 parallel run + model training, months 5-6 statistical validation, month 7 go/no-go with real data.
Next actions
Write and send RFQ to Cognex (In-Sight 9902) and Keyence (CV-X480F) with full specification: 200K units/month, 12 enumerated defect types, inline mounting requirements, structured lighting for metallic surfaces
Build a defect image library by photographing all 12 defect types across multiple alloy batches and lighting conditions — minimum 200 images per defect type for ViDi training dataset
Set up data logging infrastructure to capture machine vision pass/fail + confidence score alongside human inspector decisions for every unit, with unit-level traceability keys
At month 6, run statistical analysis on ≥500K unit comparison dataset: compute agreement rate (target ≥99.8%), defect catch rate (target ≥95% of human-caught defects), and false positive rate — make go/no-go decision on full replacement
Engage union representatives early to discuss long-term inspection automation roadmap and inspector redeployment options before validation results arrive
This verdict stops being true when
If defects are primarily dimensional or internal (not surface-visible), requiring CMM, X-ray, or functional testing rather than camera-based inspection → Invest in inline metrology (e.g., Zeiss AIMax or Keyence LJ-X8000 laser profiler) or X-ray inspection rather than optical machine vision
If the 9 inspectors are near-minimum-wage or if union contract prevents any future headcount reduction regardless of validation results → The ROI case collapses — retain manual inspection and spend $400K on upstream process improvements (SPC/predictive maintenance) to reduce defect rate and scrap costs instead
If production volume increases to >500K units/month within 12 months, making the 6-month validation timeline unacceptable due to immediate throughput constraints → Accelerate to the hybrid model (branch b002's architecture) with machine vision primary + reduced human verification, accepting higher quality risk during ramp-up
Full council reasoning, attack grid, and flip conditions included with Pro
Council notes
Vulcan
Socrates
Daedalus
Loki
Assumptions
- Defects are visually detectable — all 12 defect types have surface-visible signatures amenable to camera-based inspection rather than requiring X-ray, ultrasonic, or functional testing
- Production line has physical space for an inline machine vision station with structured lighting dome before existing manual inspection stations
- The 0.2% defect rate is stable enough that 6 months of parallel data is representative of normal production variation including seasonal material batch changes
- Cognex ViDi deep learning module can be trained to adequate performance on 12 defect types within the $30K training budget allocation
- Current 9-inspector manual inspection is the primary cost center justifying automation — if inspectors perform other functions beyond visual inspection, the ROI case weakens
Operational signals to watch
reversal — If defects are primarily dimensional or internal (not surface-visible), requiring CMM, X-ray, or functional testing rather than camera-based inspection
reversal — If the 9 inspectors are near-minimum-wage or if union contract prevents any future headcount reduction regardless of validation results
reversal — If production volume increases to >500K units/month within 12 months, making the 6-month validation timeline unacceptable due to immediate throughput constraints
Unresolved uncertainty
- Whether $100K buffer is sufficient for full replacement deployment if validation succeeds — full replacement across 3 shifts likely requires additional capital beyond the original $400K budget
- The 12 defect types are referenced but not enumerated — some defect types (cosmetic vs. dimensional vs. functional) have fundamentally different machine vision detection profiles, and the approach may work for 10 of 12 but fail on 2
- Union contract implications of eventual full replacement are deferred, not resolved — the parallel run buys time but doesn't address the labor transition plan
- ROI timeline is unclear: if 9 inspectors cost $450K-$600K/year and validation takes 7 months, the payback period depends on a second capital expenditure for full replacement that isn't budgeted
Branch battle map
Battle timeline (3 rounds)
Round 1 — Initial positions · 4 branches
Socrates proposed branch b004
Socrates
RECOMMENDATION: Shift focus from inspection methodology to defect prevention. In…
Round 2 — Adversarial probes · 3 branches
Branch b002 (Socrates) eliminated — auto-pruned: unsupported low-confidence branch
Loki proposed branch b005
Branch b005 (Loki) eliminated — This branch makes a structurally unsound argument. It cla...
Socrates proposed branch b006
Branch b006 (Socrates) eliminated — auto-pruned: unsupported low-confidence branch
Loki
Parallel or hybrid setups with manual QA ignore that machine vision false negati…
Socrates
RECOMMENDATION: Redefine the quality paradigm from defect detection to defect el…
Round 3 — Final convergence · 2 branches
Branch b004 (Socrates) eliminated — Branch b004 is structurally unsound for this decision con...