llm-api-probe

name	llm-api-probe
description	Debug reasoning of hosted LLMs (GPT-4, Claude, Gemini) through API-based probes — logprob analysis, consistency testing, chain-of-thought extraction, counterfactual probing, and calibration checks. No model weights required.

LLM API Probe Skill

Probe the reasoning of hosted language models without access to their weights. Uses logprob analysis, prompt perturbation, and structured probing to understand how and why a model produces its answers.

Like weight-level debugging but for models behind an API wall — GPT-4, Claude, Gemini, and any OpenAI-compatible endpoint.

Prerequisites

• Python 3.8+
• An API key for the target provider:
  • OpenAI: export OPENAI_API_KEY="sk-..."
  • Anthropic: export ANTHROPIC_API_KEY="sk-ant-..."
  • Google: export GOOGLE_API_KEY="..."
• pip install openai (or the appropriate provider SDK)

Quick Start

# 1. Start the API probe server
export OPENAI_API_KEY="sk-..."
python src/NeuralDebug/llm/llm_debug_session.py serve --api openai --api-model gpt-4 -p 5681

# 2. Run logprob analysis
python src/NeuralDebug/llm/llm_debug_session.py cmd -p 5681 logprobs "The capital of France is"

# 3. Test consistency
python src/NeuralDebug/llm/llm_debug_session.py cmd -p 5681 consistency "What is 2+2?" --samples 5

# 4. Extract chain-of-thought
python src/NeuralDebug/llm/llm_debug_session.py cmd -p 5681 cot "If x+3=7, what is x?"

# 5. Clean up
python src/NeuralDebug/llm/llm_debug_session.py cmd -p 5681 quit

Commands

Command	Alias	What it does
logprobs <prompt>	lp	Analyze per-token confidence, entropy, and top alternatives
perturb <prompt>	pert	Swap/remove prompt parts and compare outputs
cot <prompt>		Force chain-of-thought reasoning and return structured steps
consistency <prompt> [--samples N]	cons	Run the same question N times and measure agreement
counterfactual <prompt> --entity E --replacements R1,R2	cf	Replace entities and test how the answer changes
calibrate --test-file <path>	cal	Check stated confidence vs actual accuracy on a test set

Server Flags

Flag	Short	Description	Default
--api <provider>		API provider: openai, anthropic, google	(required)
--api-model <name>		Model name (e.g., gpt-4, claude-3-opus)	(required)
--port <N>	-p	TCP port for the probe server	5681
--temperature <T>		Default temperature for API calls	0.0
--max-tokens <N>		Max tokens per API response	256

Example Workflow

# Investigating why GPT-4 sometimes gets a math problem wrong

export OPENAI_API_KEY="sk-..."
python src/NeuralDebug/llm/llm_debug_session.py serve --api openai --api-model gpt-4 -p 5681

CMD="python src/NeuralDebug/llm/llm_debug_session.py cmd -p 5681"

# 1. Check confidence — is the model uncertain?
$CMD logprobs "What is 37 × 43?"
# → "1591" with logprob -0.41, entropy 1.8 — the model is uncertain

# 2. Test reliability — how often does it get it right?
$CMD consistency "What is 37 × 43?" --samples 10
# → "1591" 6/10 (correct), "1581" 3/10, "1601" 1/10

# 3. Force reasoning — does step-by-step help?
$CMD cot "What is 37 × 43?"
# → Step 1: 37 × 40 = 1480
#   Step 2: 37 × 3 = 111
#   Step 3: 1480 + 111 = 1591 ✓

# 4. Test generalization — is this multiplication-specific?
$CMD counterfactual "What is 37 × 43?" \
    --entity "43" --replacements "17,89,51"
# → 37×17="629" ✓, 37×89="3293" ✓, 37×51="1887" ✓

$CMD quit

Response Format

Every command returns JSON following the NeuralDebug protocol:

{
  "status": "ok",
  "command": "logprobs",
  "result": {
    "prompt": "The capital of France is",
    "tokens": [
      {
        "token": "Paris",
        "logprob": -0.05,
        "entropy": 0.21,
        "top_alternatives": [
          {"token": "the", "logprob": -3.8},
          {"token": "Lyon", "logprob": -6.2}
        ]
      }
    ],
    "summary": {
      "mean_entropy": 0.21,
      "high_entropy_tokens": []
    }
  }
}

How It Works

The API Probe uses black-box techniques — no weights, no activations, just prompt-in / response-out:

Technique	Mechanism
Logprob Analysis	Reads the logprobs field from the API response
Prompt Perturbation	Sends modified prompts and diffs the outputs
Chain-of-Thought	Prepends "Let's think step by step" and parses structure
Consistency Testing	Sends the same prompt N times at temperature > 0
Counterfactual	Replaces named entities and compares answers
Calibration	Asks "how confident are you?" and compares to ground truth

Architecture

src/NeuralDebug/llm/
├── api_probe.py              # APIProbe class — all 6 probe techniques
├── llm_debug_session.py      # CLI entry point (--api flag routes here)
└── ...

The APIProbe class accepts any async call function matching async (prompt: str, **kwargs) -> dict, making it provider-agnostic.