Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
104 lines (72 loc) · 3.45 KB

CONTRIBUTING.md

File metadata and controls

104 lines (72 loc) · 3.45 KB

Contributing to KALAM Framework

We welcome contributions. Here are specific ways you can help right now.

Open tasks (ordered by impact)

1. Run EXP-001 on AI systems we don't have access to

Impact: HIGH — This is the single most valuable contribution.

We're testing whether structured interaction frameworks reduce output noise across AI systems. We have 188 runs remaining.

What you need:

  • Access to any of: Claude, ChatGPT, Grok, DeepSeek, Gemini, Copilot, Manus, Kimi, Euria, Perplexity (or any other system)
  • 10 minutes per run

How to do it:

  1. Read the experiment protocol
  2. Pick a system and a question from the grid
  3. Run the question twice: once normally (Condition A), once with the Universal Prompt (Condition B)
  4. Record: system name, question, word count A, word count B, and both full responses
  5. Submit as a pull request adding your data to experiments/data/

Data format: See experiments/schema.py for the exact schema.

2. Validate drift detection on your own data

Impact: MEDIUM — Helps calibrate our early warning thresholds.

If you have time-series data from any monitoring system (not just AI), you can test our drift detector against it:

from tools.drift_detector import DriftDetector

d = DriftDetector()
for score in your_time_series:
    d.record(score, exchange_id="your-id")
alert = d.check()
print(alert.level, alert.rate)

Report: does it detect real declines? Does it fire false positives? What thresholds work for your domain?

3. Test detectors on diverse text

Impact: MEDIUM — Our detectors are tested on English text. We need:

  • Non-English text (does absurdity detection work across languages?)
  • Domain-specific text (medical, legal, poetic)
  • Edge cases that break the detectors
from detectors.absurdity import detect
result = detect("your text here")

4. Review and improve the gate function

Impact: HIGH (theoretical) — The gate function D = A × L × M is the core of the framework. We need:

  • Mathematical critique: is non-compensatory multiplication the right operator?
  • Alternative formulations (geometric mean? weighted?)
  • Edge case analysis: what happens at boundary values?

5. Port tools to other languages

Impact: MEDIUM — All tools are Python. Ports to JavaScript, Rust, or Go would broaden adoption.

Priority: drift_detector.py and agency_score.py (stdlib only, straightforward to port).

How to submit

  1. Fork the repository
  2. Create a branch: contrib/your-name/what-you-did
  3. Make your changes
  4. Run existing examples to make sure nothing breaks
  5. Submit a pull request with:
    • What you did
    • What you found
    • Your confidence level (C1–C5):
      • C1: Anecdotal (single observation)
      • C2: Repeated (seen multiple times)
      • C3: Systematic (controlled conditions)
      • C4: Validated (independent replication)
      • C5: Robust (survives adversarial testing)

Code style

  • Python 3.8+
  • Type hints encouraged
  • Docstrings required for public functions
  • Prefer stdlib over external dependencies
  • Each tool should be independently runnable (python tool.py should produce output)

What we don't accept

  • Changes that break existing experiment data compatibility
  • Dependencies on proprietary services or APIs
  • Code without documentation or examples
  • Findings presented without confidence levels

Questions?

Open an issue or email [email protected].