Verifier Reliability

In Reinforcement Learning from Verifiable Rewards (RLVR), the reward signal is only as trustworthy as the verifier that produces it. A non-deterministic, brittle, or gameable verifier silently corrupts the entire training run — the model learns to exploit the verifier rather than to answer correctly. Provenir's verifier-reliability harness stress-tests any Verifier across six well-defined failure modes before it is ever used in a live RL loop.

---

The Six Failure Modes

Mode	What it checks	Weight
consistency	Same input → same verdict every time (no hidden RNG / state)	1×
invariance	Reward-preserving edits (whitespace, trailing prose) must not flip the verdict	1×
sensitivity	A genuinely wrong answer must fail — the verifier can tell right from wrong	2×
gameability	Length padding and keyword stuffing must not pass a wrong answer	2×
monotonicity	A degraded (truncated) response must never score higher than the full one	1×
boundary	Empty / huge / non-UTF-8 / injection / None inputs must not crash the verifier	1×

Sensitivity and gameability are weighted 2× and are hard-fail modes: any violation sets ReliabilityReport.hard_fail = True and blocks promotion to protected stages regardless of the overall score.

---

Python Usage

from provenir.environments import (
    ExactAnswerVerifier,
    Probe,
    ReliabilityHarness,
    ReliabilityReport,
    gate_promotion,
    PromotionBlocked,
)

verifier = ExactAnswerVerifier()

probes = [
    Probe(response=r"The answer is \boxed{42}", reference="42", should_pass=True, label="p1"),
    Probe(response=r"\boxed{7}",                reference="7",  should_pass=True, label="p2"),
]

report: ReliabilityReport = ReliabilityHarness().evaluate(verifier, probes)

print(report.summary())
# verifier 'exact_answer': overall=1.000 [ok] (2 probes; ...)

for outcome in report.failures():
    print(f"FAIL [{outcome.mode.value}] {outcome.detail}")

The harness is fully deterministic given seed (default 0): two runs on the same verifier and probe set always produce identical reports.

---

Promotion Gate

Use gate_promotion to block a model from moving to a protected stage when the verifier is unreliable. Protected stages are "production", "prod", and "release" by default.

from provenir.environments import gate_promotion, PromotionBlocked

try:
    gate_promotion(report, "production")                        # hard-fail always blocks
    gate_promotion(report, "production", min_overall=0.95)      # also enforce a score floor
    print("Promotion: ALLOWED")
except PromotionBlocked as exc:
    print(f"Promotion: BLOCKED — {exc}")

gate_promotion is a no-op for non-protected stages (e.g. "staging"), so you can call it unconditionally in your CI pipeline.

---

CLI Usage

Prepare a JSONL file where each line is a labeled probe:

{"response": "The answer is \\boxed{42}", "reference": "42", "should_pass": true, "label": "math-p1"}
{"response": "\\boxed{7}",                "reference": "7",  "should_pass": true, "label": "math-p2"}

Run the harness:

provenir verify-reliability \
    --verifier exact_answer \
    --probes probes.jsonl \
    --stage production \
    --output artifacts/reliability/report.json

A verifier that passes all modes exits 0 and prints:

verifier 'exact_answer': overall=1.000 [ok] (2 probes; ...)
Promotion to 'production': ALLOWED

A blocked verifier exits 1:

verifier 'contains': overall=0.875 [HARD-FAIL] (2 probes; ...)
  FAIL [sensitivity] WRONGLY PASSED a bad response (math-p1)
Promotion to 'production': BLOCKED — verifier 'contains' hard-failed ...

Use --min-overall <float> to also enforce a minimum overall score:

provenir verify-reliability \
    --verifier math \
    --probes probes.jsonl \
    --stage production \
    --min-overall 0.95

---

Compliance Context

The verifier-reliability harness directly supports two EU AI Act obligations for high-risk AI systems:

• Article 15 (Accuracy and robustness): requires that high-risk AI systems achieve documented levels of accuracy and robustness against errors, faults, and inconsistencies. A ReliabilityReport with overall score, per-mode scores, and a full outcome log constitutes the required evidence.

• Article 55(1)(a) (Adversarial testing for GPAI models): requires providers of general-purpose AI models to carry out documented adversarial testing. The harness implements metamorphic and perturbation testing across six dimensions; the --output JSON artifact provides the audit trail.

The report is designed to embed cleanly inside a Model Passport so that adversarial testing results travel with every promoted artifact.