SkillPoison: Progressive Skill Poisoning via Successful Experiences

This repository provides the implementation of SkillPoison, a formation-stage pipeline for constructing verified successful experiences and organizing them into formation records.

Repository Structure

.
├── assets/
│   └── framework.png
├── configs/
│   └── main.yaml
├── core/
│   ├── hts/
│   ├── legsa/
│   ├── gctir/
│   ├── utils/
│   ├── config.py
│   └── pipeline.py
├── prompts/
│   ├── legsa/
│   ├── target_audit.txt
│   ├── counterfactual_generation.txt
│   └── counterfactual_audit.txt
├── scripts/
│   └── run_skillpoison.py
├── .env.example
├── requirements.txt
└── README.md

core/ contains the HTS, LEGSA, and GCTIR implementation; prompts/ stores model-facing prompt templates; configs/ stores experiment settings; and scripts/run_skillpoison.py is the main execution entry point.

Installation

Python 3.10 or later is recommended.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Model Configuration

Set the model endpoint before a live run:

export MODEL_API_KEY="your-api-key"
export MODEL_BASE_URL="https://your-endpoint/v1"
export MODEL_NAME="your-model-name"

An environment-variable template is provided in .env.example.

Configuration

The main configuration file is:

configs/main.yaml

It specifies the output directory, model settings, target behavior, embedding model, HTS thresholds, LEGSA prompt directory, and GCTIR formation settings.

Input Format

The pipeline accepts JSON or JSONL input. For JSON, use an experiences list:

{
  "experiences": [
    {
      "id": "example_001",
      "category": "category_name",
      "task_text": "task input",
      "solution": "externally verified solution",
      "trajectory": {
        "task_input": "task input",
        "model_output": "recorded successful output"
      },
      "requirement_checks": [true],
      "verified_success": true,
      "output_format": "text"
    }
  ]
}

Each candidate should contain a unique id, task/category information, a recorded successful trajectory, verification results, and an output format (text, json, or python).

Run

python scripts/run_skillpoison.py \
  --config configs/main.yaml \
  --input /path/to/candidate_experiences.json \
  --out outputs/run

The pipeline executes:

Candidate Experiences
        ↓
       HTS
        ↓
      LEGSA
        ↓
      GCTIR
        ↓
Formation Records

--out overrides the output directory specified in configs/main.yaml.

Output

A standard run writes:

outputs/run/
├── selection_audit.json
├── formation_records.json
└── run_summary.json

• selection_audit.json: candidate selection and validation results.
• formation_records.json: final formation records.
• run_summary.json: run statistics, configuration, and usage summary.

Replay Mode

For deterministic re-execution from recorded intermediate artifacts:

python scripts/run_skillpoison.py \
  --config configs/main.yaml \
  --input /path/to/replay_input.json \
  --out outputs/replay \
  --replay

Replay input must already contain the recorded artifacts required by the pipeline.