Built on UQLM · Powered by DeepEval · GPU-Accelerated
Your AI changed.
You didn't know.
Now you will.
Every time you change your system prompt or upgrade your model, your AI's behavior distribution shifts in ways traditional tests cannot detect. PromptEval measures behavioral consistency across multiple runs, detects whether your AI has developed distinct behavioral modes, and blocks your PR if consistency drops below your threshold — automatically, inside your existing DeepEval test suite.
No credit card required · One import · Works with your existing DeepEval tests
test_support_bot.py
from uncertainty_metric import UncertaintyMetric
from langchain_openai import ChatOpenAI
def test_support_bot_consistency():
# actual_output from your normal test setup
test_case = LLMTestCase(
input="I've been waiting three weeks for my order",
actual_output=support_bot.respond(input)
)
assert_test(test_case, [
UncertaintyMetric(
llm=ChatOpenAI(model="gpt-4o"),
threshold=0.75,
num_responses=5,
version_tag="pr-483-add-conciseness"
)
])
PASS score: 0.89 modes: 1 regression: none
FAIL score: 0.51 modes: 2 regression: -0.38 (p=0.003)
Catch Silent Regressions
Traditional tests fail when outputs change. AI tests fail when you have no tests for behavioral change. PromptEval gives you a consistency score before every deployment, compared to your historical baseline.
Behavioral Mode Detection
When your AI is inconsistent, most tools give you a low score. PromptEval shows you whether the inconsistency is random noise or whether your AI has developed two distinct behavioral modes.
Statistical Confidence
A score of 0.80 from 5 samples could mean a true rate anywhere from 0.45 to 0.97. PromptEval shows you the Wilson confidence interval alongside every score so you know whether to trust the result.
Works with every LLM provider — your API key, your infrastructure
Integrates with
DeepEvalpytestGitHub ActionsGitLab CIAny CI pipeline
PromptEval uses your LLM through your own API key. No model data is routed through our servers.
The problem
The testing gap every AI team has

Traditional software tests are deterministic. The same input always produces the same output. A failing assertion means something broke. A passing test suite means nothing broke. This works perfectly — for traditional software.

AI applications are not deterministic. Published research running a prompt 1,000 times at temperature zero produced 80 unique completions — the most common appeared only 78 times. This means: when you change your system prompt and your tests pass, you know your AI produced an acceptable output that one time. You do not know whether the change shifted your AI's behavioral distribution in ways that will manifest as problems on 10%, 30%, or 40% of future requests.

This is a silent regression. No exception. No error. No alert. Your AI just starts behaving differently in production, affecting real users, until someone notices.

Added "be concise" to system prompt → empathy instructions now lose to brevity 40% of the time

Customer support bot stops recommending doctor contact in 30% of medical queries. No test catches it.

Upgraded from GPT-4 to GPT-4o → tone shifted slightly more casual on formal queries

Luxury brand chatbot violating brand standards on 25% of responses. Passes all existing tests.

Reordered few-shot examples → model infers different implicit behavioral standard

Factual accuracy drops on edge case inputs. Traditional tests never touched those cases.

None of these trigger an exception. None fail a traditional test. All three are caught by PromptEval.

The Three Layers
The Three Layers Inside Every Evaluation
PromptEval runs three layers of analysis on every call —automatically, always on, no configuration required.

Layer 1 — UQLM Black-Box Consistency Scoring

The primary score. Built on UQLM's non-contradiction scorer — an NLI model that compares every pair of responses. No access to model internals required.

NLI Scorer: noncontradiction
GPU Batched
Pair 0, 2
consistent
0.92
Pair 0, 1
contradiction
0.12
Pair 1, 2
contradiction
0.08
Pair 0, 3
consistent
0.88
Aggregate Score0.51

Layer 2 — Behavioral Mode Detection

Is the inconsistency random, or does your AI have distinct behavioral patterns? Clusters embeddings to find modes — the signature of competing instructions.

SBERT Embedding Clustersmodes_detected: 2
Cluster A · 60%
warm_empathetic_solution_focused
Cluster B · 40%
terse_informational_transactional
Silhouette Score: 0.74Interpret: Structured

Layer 3 — Statistical Confidence + Regression

The Wilson confidence interval tells you how much to trust the result. The Kolmogorov-Smirnov test identifies statistically significant behavioral shifts.

Kolmogorov-Smirnov Test
Regression Detected
PR-480
PR-481
PR-482
BASE
PR-483
PR-484
ks_p_value
0.003
score_delta
-0.38

The Diagnostic Output

Every evaluation returns a structured diagnostic explaining what was found — not just a score. When two behavioral modes are detected, you see a sample response from each mode, the frequency of each, and the interpretation of what the pattern means.

The diagnostic identifies that behavioral modes exist and what they look like. It does not identify which specific instruction caused the conflict — that requires reviewing your system prompt directly.
diagnostic.json
{
  "result": { "score": 0.51, "passed": false },
  "layer_2_behavioral": {
    "modes_detected": 2,
    "modes": [
      { "label": "warm_empathetic", "frequency": 0.60,
        "sample": "I completely understand how frustrated you..." },
      { "label": "terse_transactional", "frequency": 0.40,
        "sample": "Your order is delayed. Update in 2 hours." }
    ]
  },
  "diagnostic": {
    "summary": "FAIL — Behavioral regression detected...",
    "detail": "Two distinct behavioral modes detected..."
  }
}
The Methodology

What PromptEval measures — and what it doesn't

PromptEval is designed to be one layer in a robust AI testing pipeline. It solves for Behavioral Regression, but it is not a replacement for correctness or safety testing.

1

PromptEval

Measures behavioral consistency. Detects if your model's response distribution has shifted or developed new behavioral modes.

2

Correctness Metrics

Measures if the output matches a reference ground truth or adheres to specific factual requirements (e.g., DeepEval Factuality).

3

Human Review

Measures nuance, brand voice, and subtle safety edge cases that require subjective human judgment.

"Consistency is not correctness. A model that consistently produces the wrong answer is 100% consistent and 0% correct. PromptEval detects when your AI's behavior shifts; you decide if the shift is an improvement."

The Behavioral Principle
How it runs
How the evaluation runs — end to end
Five concurrent LLM calls, GPU-accelerated NLI.Designed for CI/CD pipelines where speed and statistical confidence are mandatory.
Test Suite
DeepEval / pytest
Your LLM
5 concurrent calls
PromptEval GPU API
UQLM / SBERT Batched
Result + History
Regression detection
Internal Concurrency Flow
# PromptEval executes all inference calls in parallel
async def run_consistency_check(prompt, n=5):
  # 5 calls run at the same time
  responses = await asyncio.gather(
    *[llm.generate(prompt) for _ in range(n)]
  )

  # Results sent to GPU-accelerated NLI scorer
  result = await prompteval.score(responses)
  return result

Total latency is roughly the same time as a single LLM call plus ~100ms for batched scoring.

Use Cases
Built for AI engineering teams
Behavioral regression detection for every stage of development.Credibility in AI starts with statistical consistency.

Pre-Deployment Gate

Traditional evals check if a single output is 'good'. PromptEval checks if your system prompt change shifted the entire behavioral distribution.

Scenario
"Catch when 'be concise' override instructions for empathy 40% of the time, before the PR is merged."
Block behavioral regressions

Model Upgrade Validation

When moving from GPT-4 to GPT-4o, benchmarks say intelligence stays the same. PromptEval tells you if the nuances of your product's voice actually shifted.

Scenario
"Detect if GPT-4o produces slightly more casual outcomes on formal customer queries compared to GPT-4."
Validate model migrations

Compliance-Sensitive AI

For medical, legal, or financial AI, consistency isn't just a quality metric — it's a requirement. Ensure your AI behaves exactly the same way, every time.

Scenario
"Identify if a few-shot example change causes the model to infer a different regulatory boundary on edge cases."
Safety & compliance gate
The CI/CD Flow
What happens when a regression is detected
PromptEval sits inside your CI pipeline as a final gate.Catch behavioral shifts before they reach a single customer.
Without PromptEval
14 Days Damage
1

Change system prompt to 'be concise'

2

Run traditional tests (All Pass)

3

Deploy to Production

4

AI behavior shifts; empathy instructions lost

5

14 days of brand damage until manually detected

With PromptEval
PR Blocked
1

Change system prompt to 'be concise'

2

Run PromptEval CI Gate

3

Consistency drop detected (v1.0 vs v1.1)

4

PR Blocked. Behavioral regression averted.

The Historical Baseline

PromptEval doesn't just measure consistency against an abstract threshold. Every test run is compared against a specific version_tag or historical run of your choice.

When you run an eval on a feature branch, PromptEval automatically fetches the distribution of the main branch for comparison. This allows for the Kolmogorov-Smirnov (KS) test to detect if the new version is statistically different from the baseline.

version_tag="prod-v1.4"
current_run
Distribution Shiftp = 0.003
-0.38
KS Statistic (D)
Statistically significant difference detected between baseline and current run.
Numbers that speak
Why engineering teams use PromptEval

Built for Speed

One import into your existing DeepEval or pytest suite. No separate infrastructure to manage or wait for.

1 Import
To start evaluating

Higher Throughput

Proprietary batched NLI scoring engine runs evaluations significantly faster than sequential LLM calls.

3–5× Faster
Than sequential eval

Black-Box Native

Measure consistency on any model including GPT-4o, Claude 3.5, and Gemini. No logprob or model internal access required.

0
Model Internals Required

Statistical Strength

We generate multiple pairs of responses for every test case to ensure the consistency score is mathematically significant.

10 NLI Pairs
Per test case

Security First

Model data never touches our servers. The evaluation runs in your infrastructure, using your existing LLM provider keys.

100%
Your Infrastructure

Baseline Persistence

Every run is stored and indexed by version_tag. Fetch historical distributions for CI comparison in real-time.

30 Days
Historical Data
Advanced Capability
Features for production AI testing
PromptEval provides the statistical and behavioral tools required for senior engineering teams.

NLI Scoring

Non-contradiction batched scoring engine

UQLM-based semantic alignment Scorer
GPU-batched pairwise comparisons
Deterministic NLI inference
Zero-logprob dependency
Customizable scoring threshold

SBERT Clustering

Automated behavioral mode detection

SBERT embedding generation
K-Means + Silhouette score analysis
Automatic mode labeling
Pattern interpretation & diagnostic
Cluster visualization output

Statistical Confidence

Wilson Confidence Intervals

Exact binomial confidence intervals
Reliable scoring on small samples
Adaptive n-size recommendations
Statistical power indicators
Confidence-aware pass/fail logic

KS Regression Test

Distribution shift detection

Kolmogorov-Smirnov statistical test
Compare current run vs version_tag
P-value significance reporting
Detect subtle logic drifts
CI/CD gate integration native

Concurrent Calls

Zero-latency penalty architecture

Async gather for internal LLM calls
Parallel diagnostic generation
Fastest possible time-to-result
Rate-limit aware execution
Configurable concurrency factor

Historical Dashboard

Visualize regression over time

Track consistency scores per PR
Compare distributions visually
Export reports for brand safety
Drill down into behavioral modes
Centralized evaluation history

Historical Regression Tracking

PromptEval automatically tags every run and allows you to visualize behavioral shifts between versions in real-time.

Regression Dashboard
V1.0V1.1V1.2
Alert — Behavioral regression
V1.1 significantly different from V1.0 baseline (KS Test)
CI integration ready

Export results as JUnit XML or JSON for native integration with GitHub Actions, GitLab CI, and CircleCI.

Documentation
Clear documentation for every part of the evaluation
From the math behind the Wilson interval to the CI/CD integration guide.Credibility in AI starts with statistical consistency.
Quickstart Guide
Get PromptEval running in 5 minutes with our copy-paste usage examples for DeepEval and pytest.
Methodology Reference
Read the UQLM and SBERT implementation details. From the math to the GPU-batched NLI scoring engine.
Interpreting Results
Understand your evaluation diagnostic. How to read the KS p-value and behavioral modes.
docs.prompteval.ai
installation
Quickstart Guide
# Install the SDK
pip install uncertainty-metric

# Import the metric
from uncertainty_metric import UncertaintyMetric
Open Research
Built on open research and open source
The mathematical core of PromptEval is based on the open-source Uncertainty Metric (UQLM). We believe that evaluation methodology should be transparent and auditable.The PromptEval SaaS provides the GPU-accelerated scoring engine, historical regression tracking, and CI/CD dashboard layer built on top of this open foundation.

Transparent Math

Our scoring logic isn't a proprietary black box. The NLI methodology and KS test parameters are fully documented and auditable.

Contribution Ready

The open-source core allows AI researchers to contribute new behavioral detection modes and statistical tests to the community.

SaaS Reliability

The PromptEval SaaS manages the GPU infrastructure and database layer required for high-throughput production evaluation.

Frequently Asked Questions
Everything you need to know about PromptEvaland behavioral regression detection.
Detecting behavioral consistency requires a statistical distribution. While a single pass tells you if the model worked once, 5 passes allow us to calculate Wilson Confidence Intervals and detect behavioral modes. This 5x sampling is what enables PromptEval to catch silent regressions that traditional single-pass tests miss.
Yes. PromptEval is model-agnostic. We support any LangChain-compatible model, including Ollama, self-hosted vLLM, and enterprise inference endpoints. As long as the model produces text, PromptEval can measure its behavioral consistency.
We execute all 5 LLM calls concurrently using asynchronous gathering. This means the total latency of a PromptEval run is roughly equivalent to the time of a single LLM call, plus a negligible ~100ms overhead for our batched NLI scoring engine.
PromptEval is designed for developer privacy. We store response hashes and consistency metadata indefinitely for historical tracking. The raw response text is stored temporarily for 30 days to allow for behavioral clustering and regression diagnostic analysis, after which it is automatically deleted from our systems.
UQLM stands for Uncertainty Quantification for Large Language Models. It is the open research framework PromptEval uses to measure semantic alignment and consistency. It moves beyond simple string matching and uses Natural Language Inference (NLI) to understand if two responses mean the same thing, even if they use different words.
PromptEval is a behavioral detector, not a system prompt debugger. Our diagnostic will tell you that a regression occurred and describe the new behavior (e.g., 'Model has become more concise'), but it is up to the developer to review their system prompt changes to identify the conflicting instruction.
Free for up to 100 test cases/month
Start detecting regressions today.
Credibility in AI starts with statistical consistency. Join the teams using PromptEval to move from gut-feeling evals to rigorous behavioral regression detection.
No credit card required
Works with every LLM
Open-source core
PromptEval
Behavioral regression detection for AI engineering teams. Built on UQLM research, powered by GPU-accelerated NLI scoring.
© 2026 PromptEval. All rights reserved.
Statistical rigor for AI engineering