Release Gate

One declarative policy (config/policy.yaml), worst-gate-wins: every configured gate is evaluated independently, and the overall decision is the single worst outcome among them. This is the centerpiece of GenAI ReleaseGate — every result below is real, computed evidence.

GO

The candidate satisfies every release requirement — safe to ship.

REVIEW

Evidence is inconclusive or a trade-off requires a human decision before shipping.

HOLD

A hard release criterion was violated — automatic release block.

INVALID

The evaluation itself is not trustworthy enough to decide (e.g. too many failed responses to trust the comparison).

Real experiment results

REAL DATA

Every real, persisted comparison — the full release history. V2 through V3.3 were each HELD or REVIEWed for a real reason; V3.4 is the first candidate to pass all 8 gates, on both the dev set and the sealed holdout. The decision engine evaluates every gate and applies the configured policy — no manual override was used to produce any result below, including V3.4's GO.

V2 vs V1 — dev

HOLD

120 paired cases · dev split

  • — faithfulness: delta -0.1077 crossed the hold threshold -0.0800
  • — benign_false_positive_rate: candidate value 0.5556 exceeds the allowed maximum 0.15
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0022)
FaithfulnessHOLDdelta -0.1077 crossed the hold threshold -0.0800
Hallucination rateGOwithin configured thresholds (delta=-0.0333)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)HOLDcandidate value 0.5556 exceeds the allowed maximum 0.15
p95 latencyGOwithin configured thresholds (delta=+120.5000)
Cost per queryGOwithin configured thresholds (delta=-0.0000)

V3 vs V1 — dev

REVIEW

120 paired cases · dev split

  • — cost_per_query: delta +25.79% crossed the review threshold +15.00%
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0232)
FaithfulnessGOwithin configured thresholds (delta=+0.0167)
Hallucination rateGOwithin configured thresholds (delta=-0.0339)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=-143.4000)
Cost per queryREVIEWdelta +25.79% crossed the review threshold +15.00%

V3.1 vs V1 — dev

REVIEW

120 paired cases · dev split

  • — faithfulness: delta -0.0601 crossed the review threshold +0.0000
  • — cost_per_query: delta +17.02% crossed the review threshold +15.00%
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0739)
FaithfulnessREVIEWdelta -0.0601 crossed the review threshold +0.0000
Hallucination rateGOwithin configured thresholds (delta=-0.0339)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=-143.2000)
Cost per queryREVIEWdelta +17.02% crossed the review threshold +15.00%

V3.2 vs V1 — dev

REVIEW

120 paired cases · dev split

  • — faithfulness: delta -0.0160 crossed the review threshold +0.0000
  • — cost_per_query: delta +22.54% crossed the review threshold +15.00%
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0707)
FaithfulnessREVIEWdelta -0.0160 crossed the review threshold +0.0000
Hallucination rateGOwithin configured thresholds (delta=-0.0339)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=-318.4000)
Cost per queryREVIEWdelta +22.54% crossed the review threshold +15.00%

V3 vs V2 — dev

REVIEW

120 paired cases · dev split

  • — cost_per_query: delta +37.60% crossed the review threshold +15.00%
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0210)
FaithfulnessGOwithin configured thresholds (delta=+0.1263)
Hallucination rateGOwithin configured thresholds (delta=+0.0000)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=-0.4444)
p95 latencyGOwithin configured thresholds (delta=-278.6000)
Cost per queryREVIEWdelta +37.60% crossed the review threshold +15.00%

V3.3 vs V1 — dev

REVIEW

120 paired cases · dev split

  • — cost_per_query: delta +16.30% crossed the review threshold +15.00%
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0912)
FaithfulnessGOwithin configured thresholds (delta=+0.0329)
Hallucination rateGOwithin configured thresholds (delta=-0.0508)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=-303.2000)
Cost per queryREVIEWdelta +16.30% crossed the review threshold +15.00%

V3.4 vs V1 — dev

GO

120 paired cases · dev split

  • — GO — all 8 gate(s) passed
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.0798)
FaithfulnessGOwithin configured thresholds (delta=+0.0453)
Hallucination rateGOwithin configured thresholds (delta=-0.0339)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=-361.2000)
Cost per queryGOwithin configured thresholds (delta=+0.0000)

V3 vs V1 — holdout (sealed)

REVIEW

40 paired cases · holdout split

  • — answer_correctness: delta -0.0753 crossed the review threshold +0.0000
  • — cost_per_query: delta +27.83% crossed the review threshold +15.00%
GateStatusReason
Answer correctnessREVIEWdelta -0.0753 crossed the review threshold +0.0000
FaithfulnessGOwithin configured thresholds (delta=+0.0202)
Hallucination rateGOwithin configured thresholds (delta=+0.0000)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=+11.4000)
Cost per queryREVIEWdelta +27.83% crossed the review threshold +15.00%

V3.4 vs V1 — holdout (sealed)

GO

40 paired cases · holdout split

  • — GO — all 8 gate(s) passed
GateStatusReason
Answer correctnessGOwithin configured thresholds (delta=+0.1011)
FaithfulnessGOwithin configured thresholds (delta=+0.0281)
Hallucination rateGOwithin configured thresholds (delta=-0.0526)
Prompt injection block rateGOwithin configured thresholds (delta=+0.0000)
Sensitive-information protectionGOwithin configured thresholds (delta=+0.0000)
Benign false positives (guardrail suite)GOwithin configured thresholds (delta=+0.0000)
p95 latencyGOwithin configured thresholds (delta=-205.4500)
Cost per queryGOwithin configured thresholds (delta=+0.0000)

Decision-engine validation fixture

VALIDATION FIXTURE

Before V3.4's real GO existed, this clearly-labeled synthetic fixture — NOT a real experiment — was used to prove the engine could reach GO at all. Same shape as the backend's own test_scenario_go_deterministic_fixture_every_gate_within_threshold, evaluated against the real config/policy.yaml thresholds. It played no part in V3.4's result and is kept only for engine validation.

What GO looks like (synthetic, not a real result)

VALIDATION FIXTUREGO

Source: tests/unit/test_policy_engine.py — Decision Engine Validation Scenarios (deterministic synthetic fixture, not a real experiment)

GateStatusSynthetic delta
Answer correctnessGO+0.05
FaithfulnessGO+0.05
Hallucination rateGO-0.01
Prompt injection block rateGO+0.00
Sensitive-information protectionGO+0.00
Benign false positives (guardrail suite)GO+0.00
p95 latencyGO+0.00
Cost per queryGO+0.00