GenAI ReleaseGate

Can We Safely Ship a Better LLM?

The business problem: before deploying a new prompt, can we prove it's safe, useful, operationally acceptable, and worth shipping — with evidence, not a single benchmark score?

GOThe candidate satisfies every release requirement — safe to ship.
REVIEWEvidence is inconclusive or a trade-off requires a human decision before shipping.
HOLDA hard release criterion was violated — automatic release block.
INVALIDThe evaluation itself is not trustworthy enough to decide (e.g. too many failed responses to trust the comparison).

The real release journey

Six real candidates, each decided by the same automated policy engine. The gate actually rejected or reviewed five of them before one passed.

V1Production baseline
V2HOLD
V3REVIEW
V3.1REVIEW
V3.2REVIEW
V3.3REVIEW
V3.4GO

V3.4 — First Candidate to Pass the Release Gate

REAL DATA

Dev — 120 cases

GOcost +12.17%

Sealed holdout — 40 cases

GOcost +10.81%

Quality (dev)

Correctness +7.98 pp

Faithfulness +4.53 pp

Hallucination 1.69%

Safety

Injection blocking 100%

Sensitive-info protection 100%

Performance

p95 latency vs V1 (dev) -30.6%

“GO” here means passed the release gate — not the same as production deployment: a promotion candidate with statistically-grounded evidence behind it, not a claim that this has been deployed to production. “Cost” above is generator (model inference) cost only — judge and guardrail costs are real, separately tracked, and not part of this figure.

Candidates the gate rejected

V3.4 wasn't the first attempt — it's the fifth. The gate reviewed or held back every one of the others.

V2HOLD

faithfulness: delta -0.1077 crossed the hold threshold -0.0800; benign_false_positive_rate: candidate value 0.5556 exceeds the allowed maximum 0.15

V3REVIEW

cost_per_query: delta +25.79% crossed the review threshold +15.00% (+25.79% cost)

V3.1REVIEW

faithfulness: delta -0.0601 crossed the review threshold +0.0000; cost_per_query: delta +17.02% crossed the review threshold +15.00% (+17.02% cost)

V3.2REVIEW

faithfulness: delta -0.0160 crossed the review threshold +0.0000; cost_per_query: delta +22.54% crossed the review threshold +15.00% (+22.54% cost)

V3.3REVIEW

cost_per_query: delta +16.30% crossed the review threshold +15.00% (+16.30% cost)

V3.4GO

+12.17% dev / +10.81% holdout

Cost delta vs. release decision (dev)

The release gate did not simply reward the highest benchmark score — it rejected five candidates until one satisfied the complete release policy. V2's cost delta is shown for completeness; its HOLD was driven by faithfulness and safety, not cost.

Dev → sealed holdout

Development evaluation — 120 cases

GO

Sealed holdout — 40 cases

GO

Promotion evidence

The candidate was not promoted solely because it passed development evaluation. The same prompt was evaluated on a sealed holdout set — cases it had never been run against before — using the identical, unmodified release policy.