TL;DR
Most AI teams are at level 1 or 2 and don't know it. The five stages: (1) Vibe testing, eyeballed in a notebook. (2) Manual eval, spreadsheet of cases. (3) Automated eval, CI runs eval gates. (4) Continuous eval, production drift + shadow deploys. (5) Adversarial + governance, chaos, compliance evidence, regulator-ready audit trails. Skipping levels doesn't work; each unlocks the next.
If you ask "where is our AI quality program?" most teams answer with examples, "we have a Promptfoo setup, " "we run an eval before deploy", without a framework that places those examples on a maturity arc.
This is that framework. Five stages, each with concrete checkpoints. Use it to assess where you are and to make the case for what investment is needed to move up.
Level 1: Vibe testing
Characteristic: someone runs the model in a notebook, eyeballs the output, ships when it "looks good."
Indicators:
- No persistent eval set, the test is whatever was in the notebook that day
- "Quality" is a discussion, not a number
- Quality decisions are made by whoever has the loudest opinion
- Regressions discovered by users, not engineers
- No version control on prompts
This is where every team starts. It's appropriate for a v0 prototype. It's catastrophic for a shipped product. The first signal you've outgrown it: the same regression keeps shipping, because nobody can prove it shipped.
Level 2: Manual eval
Characteristic: a curated eval set of test cases lives in a spreadsheet or YAML file. Before each release, someone runs the model against the set and reviews outputs by hand.
Indicators:
- 50-500 cases curated and version-controlled
- Pre-release manual eval is part of the deploy checklist
- Quality conversations have a number ("score went from 4.2 to 4.0")
- Prompts in version control with diffs reviewed
- Still no continuous coverage in CI
This is the level most teams are at when they think they're more mature. The eval set is real but it runs on someone's laptop on Friday afternoons. The gap to level 3 is automation, not coverage.
Level 3: Automated eval
Characteristic: every PR runs an eval suite in CI. A failing eval gate blocks merge.
Indicators:
- Eval set in 1k-5k cases, segmented by category
- CI runs an eval-gate stage; merge blocks on threshold violation
- LLM-as-judge for scoring at scale, with human spot-checks
- Cost-aware test tiering (fast on PR, full on merge)
- Per-segment metrics, not just aggregate
- Quality dashboard the team actually looks at
Level 3 is the threshold for "we ship AI features responsibly." Most production AI teams should be here. Below 3, the rate of customer-visible regressions is unacceptable for any product of consequence.
Common pitfall: shipping a level-3 eval suite that nobody trusts. If the suite produces too many false positives, engineers learn to ignore failures. The fix is investment in calibrating the LLM judge against human raters until inter-rater agreement clears 0.7.
Level 4: Continuous eval
Characteristic: quality is monitored continuously in production. Drift, distribution shift, output quality, and user feedback all feed back into the system.
Indicators:
- Production trace logging for every inference, sampled at 1-10%
- Online evaluators score a fraction of live outputs
- Drift monitors page on distribution shift
- Shadow deploys for model upgrades, with auto-promotion gates
- Cost tracking on every release; cost regressions block deploy
- User feedback (thumbs up/down, edits) routed back into the eval set
Level 4 is where the eval set becomes a living artifact, not a one-time deliverable. The cycle: production identifies a failure → it gets added to the eval set → next release is gated against it. This is what "AI quality engineering" actually looks like in practice.
Investment cost: typically 1-2 quality engineers per ML team, plus tooling (LangSmith / Phoenix / Langfuse / Helicone or equivalent). Worth it from the moment your product is mission-critical.
Level 5: Adversarial + governance
Characteristic: the program goes beyond functional quality into safety, security, compliance, and audit.
Indicators:
- Red-team / adversarial eval suite with regular runs (weekly+)
- Compliance evidence trail, for regulated industries, every model decision is auditable
- Bias and fairness evaluations on every release
- Privacy testing, PII leakage, memorization checks
- Security testing, prompt injection, jailbreak resistance
- Governance: model cards, eval cards, change-control records
- External audits or certifications (SOC2, ISO 42001, EU AI Act readiness)
Level 5 is mandatory for regulated industries (finance, healthcare, government, EU operators). It's a competitive moat for everyone else, when an enterprise customer asks "how do you test for bias?" the L5 answer wins the deal.
Where do most teams sit?
Distribution we see across alt.qa customers in 2026:
| Level | % of teams |
|---|---|
| 1, Vibe testing | ~20% |
| 2, Manual eval | ~35% |
| 3, Automated eval | ~25% |
| 4, Continuous eval | ~15% |
| 5, Adversarial + governance | ~5% |
The painful truth: most teams who self-describe as "level 4" are actually at level 3. The difference is whether quality monitoring runs continuously in production with feedback loops, vs. just running in CI.
How to move up a level
The transitions that work:
- 1 → 2: Curate a 50-case eval set in a YAML file. Run it on every release. Total investment: 2 days.
- 2 → 3: Wire the eval set into CI as a merge gate. Add an LLM judge for scale. Total investment: 1-2 weeks.
- 3 → 4: Add production logging, an online evaluator, drift monitors, and the feedback loop from prod into the eval set. Total investment: a quarter, with a dedicated engineer.
- 4 → 5: Add red-team eval, compliance trail, fairness/safety/security eval. Total investment: a quarter to half-year, with dedicated quality and compliance leads.
The mistake is trying to skip. Going from 2 to 4 directly produces tooling no one understands and an eval set no one trusts. Each transition compounds, automated eval (3) is what makes continuous eval (4) trustable; continuous eval (4) is what makes adversarial governance (5) practical.
The single highest-leverage move
If you're at level 1 or 2 and you can do exactly one thing this quarter, make it the 2→3 transition: get an eval gate into CI. It will surface every prompt regression, force the team to articulate quality bars, and unlock everything that follows. Without it, no amount of monitoring or governance will produce trustworthy AI.