# Production-Readiness Checklist for Computer Vision

Eighteen checks across six layers of the stack, plus the eight-statement release
gate. Source: VisionRoute — "What 'production-ready computer vision' actually
requires" (published 2026-08-31). Free to copy and adapt.

Scope: narrow, still-image decisions in software products. Regulated and
safety-relevant domains (medical, biometric, safety-critical) need substantially
more than this checklist.

## Output contract & thresholds (1–3)

- [ ] **1. Every response is validated against a schema.** Malformed or off-schema
  output takes a defined failure path — it never crashes the flow and never
  silently passes as a prediction.
- [ ] **2. Confidence thresholds come from your evaluation data.** Set per failure
  direction, from the measured tradeoff on representative images — not from a
  library default.
- [ ] **3. Below-threshold behavior is specified and implemented.** Abstain, queue
  for review, or apply a safe default — chosen per decision, in writing.

## Review & abstention (4–6)

- [ ] **4. The human review queue has an owner and an SLA.** A queue nobody drains
  is a silent outage. The product defines what happens while an item waits.
- [ ] **5. Reviewer decisions are captured as labeled data.** Every override is a
  free training and regression example.
- [ ] **6. The abstention rate is budgeted and monitored.** A model that quietly
  routes 30% of traffic to humans has changed your unit economics without
  telling you.

## Versioning & change (7–9)

- [ ] **7. Every decision records its full configuration.** Model or API version,
  prompt, thresholds, preprocessing — enough to reproduce any past decision.
- [ ] **8. No change ships without a regression run.** Model swap, prompt edit,
  threshold tweak, provider update — each runs the evaluation suite first, and
  the diff is read by slice.
- [ ] **9. Provider change policy is explicit.** Pin model versions where the
  provider allows it; where it does not, monitor for behavior shifts you did
  not initiate.

## Monitoring & drift (10–12)

- [ ] **10. Input drift has signals.** Source mix, image quality, class balance —
  the distribution you validated against is the assumption production erodes
  first.
- [ ] **11. Outcome quality is tracked per slice in production.** Direct labels
  where you have review coverage; proxies (overturn rate, disputes) where you
  do not.
- [ ] **12. Alerts have thresholds and a named receiver.** A dashboard nobody is
  paged on is documentation, not monitoring.

## Economics & latency (13–15)

- [ ] **13. Latency is measured at production volume.** p50 and p95, on your
  serving path, from your region, under real concurrency.
- [ ] **14. Per-decision cost is computed end to end.** Including retries,
  preprocessing, storage, and the human-review minutes each abstention buys.
- [ ] **15. Economics are re-checked at volume milestones.** The stack that made
  sense at 5,000 images/day can be the wrong stack at 50,000 — calendar it.

## Rollback & incident (16–18)

- [ ] **16. The previous version is deployable quickly.** Model, prompt, and
  thresholds roll back together as one unit — a model rollback under new
  thresholds is a third, untested configuration.
- [ ] **17. A kill switch exists.** One action routes all traffic to human review
  or a safe default. Rehearsed, not theoretical.
- [ ] **18. The incident playbook names the first three checks.** Did the input
  distribution change? Did the provider change? Did we change a threshold?
  Most vision incidents are one of these three.

## The release gate — all eight true, or not yet

1. Acceptance criteria are met on the current regression suite, by slice, on the
   exact build being shipped.
2. Every component version in the decision path — model, prompt, thresholds,
   preprocessing — is recorded and reproducible.
3. The low-confidence path has been exercised end to end in staging, including
   the human queue and its notifications.
4. A named person owns each of: the review queue, drift alerts, and regression
   reruns.
5. p95 latency and per-decision cost at expected volume are measured on your
   serving path, not projected.
6. Rollback to the previous full configuration has been rehearsed at least once.
7. Every failure mode with a named business cost has monitoring that would
   actually catch it.
8. The team can answer "when do we rerun the suite?" with both a calendar
   interval and a list of trigger events.

Passing the gate means the stack is operable today — not certified forever.
Models, providers, and your input distribution drift; the gate is re-earned at
every material change.
