NerfProbe
Scientifically-grounded LLM degradation detection for developers.
Installation
pip install nerfprobe
Quick Start
CLI
# Run core probes on a model
nerfprobe run gpt-5.2 --tier core
# Run specific probes
nerfprobe run gpt-5.2 --probe math --probe style --probe code
# Use different provider
nerfprobe run claude-opus-4.5 --provider anthropic
# Custom endpoint (vLLM, Ollama, local)
nerfprobe run my-model --base-url http://localhost:8000/v1
# Output formats
nerfprobe run gpt-5.2 --format json > results.json
nerfprobe run gpt-5.2 --format markdown
Model Registry
# List known models (10 SOTA as of Dec 2025)
nerfprobe list-models
# Research unknown model
nerfprobe research qwen3:8b --provider alibaba
# -> Outputs prompt to paste into any LLM
# Parse research response
nerfprobe research qwen3:8b --provider alibaba --parse '{"context_window": 32768}'
Python API
import asyncio
from nerfprobe import run_probes, OpenAIGateway
async def main():
gateway = OpenAIGateway(api_key="...")
# Run core tier
results = await run_probes("gpt-5.2", gateway, tier="core")
for r in results:
print(r.summary())
# math_probe: PASS (1.00) in 234ms
# style_probe: PASS (0.87) in 189ms
# timing_probe: PASS (1.00) in 156ms
# code_probe: PASS (1.00) in 312ms
await gateway.close()
asyncio.run(main())
Probes
| Tier | Probes | Description |
|---|---|---|
| core | math, style, timing, code | Essential degradation signals |
| advanced | fingerprint, context, routing, repetition, constraint, logic, cot | Research-backed detection |
| optional | calibration, zeroprint, multilingual | Requires logprobs or multi-call |
| all | All 14 probes | Comprehensive testing |
Gateways
| Gateway | Providers |
|---|---|
OpenAIGateway |
OpenAI, OpenRouter, vLLM, Ollama, Together, Fireworks |
AnthropicGateway |
Claude models |
GoogleGateway |
Gemini models |
BedrockGateway |
AWS Bedrock (Claude, Titan) |
DashScopeGateway |
Alibaba Qwen models |
ZhipuGateway |
GLM models |
OllamaGateway |
Local Ollama models |
Environment Variables
# API keys (or use --api-key flag)
export OPENAI_API_KEY="..."
export ANTHROPIC_API_KEY="..."
export GOOGLE_API_KEY="..."
export OPENROUTER_API_KEY="..."
Research Basis
All probes are grounded in peer-reviewed research:
- MathProbe: 2504.04823 - Quantization Hurts Reasoning
- StyleProbe: 2403.06408 - Perturbation Lens
- TimingProbe: 2502.20589 - LLMs Have Rhythm
- CodeProbe: 2512.08213 - Package Hallucinations
- FingerprintProbe: 2407.15847 - LLMmap
- ContextProbe: 2512.12008 - KV Cache Compression
- RoutingProbe: 2406.18665 - RouteLLM
Dependencies
nerfprobe-core- Probe implementationshttpx- HTTP clienttyper- CLI frameworkrich- Terminal output
License
Apache-2.0
Release files for nerfprobe 0.2.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| nerfprobe-0.2.0.tar.gz | 49.1 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| nerfprobe-0.2.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 70.5 kB
Release files / nerfprobe-0.2.0.tar.gz
| Download URL | nerfprobe-0.2.0.tar.gz |
|---|---|
| Size | 49.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
34b0d38d5d100ed887913dfc2428fc955d129d446aae208ef44394dc7df52e85
|
|
BLAKE2b-256 checksum How to use checksums |
461dc8849de5dabd3b9138b75eb2a5bb76e93ab33ea8f8814cd282ca1276cdf9
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Dec 23, 2025.
Transparency logRelease files / nerfprobe-0.2.0-py3-none-any.whl
| Download URL | nerfprobe-0.2.0-py3-none-any.whl |
|---|---|
| Size | 21.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
7f86f449f95ab72ecbdaad9f6b3735452a4ab411b883e87417d7ed45a4512585
|
|
BLAKE2b-256 checksum How to use checksums |
80e248ec6cd5c00c2be19ce2ff5a1b69d665857b1ad888f77c4e2c492cbb41c3
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Dec 23, 2025.
Transparency log