CHANGELOG.md

Changelog

All notable changes to OBLITERATUS are documented here. Format follows Keep a Changelog.

[0.1.2] - 2026-03-03

Fixed

• Fixed spaces.GPU AttributeError crash on HuggingFace Spaces — fallback now catches both ImportError and AttributeError so the Space gracefully degrades to CPU mode when ZeroGPU is unavailable
• Added missing hardware: zero-a10g to HF Space metadata (hf-spaces/README.md) — required for the spaces package to expose the @spaces.GPU decorator

Improved

• Added mypy type checking to CI pipeline (continue-on-error while baseline is established)
• Added mypy to dev dependencies
• Version bump to 0.1.2 across pyproject.toml and __init__.py

[0.1.1] - 2026-03-01

Fixed

• Fixed all broken imports (missing function exports in telemetry, evaluation, analysis modules)
• Resolved all ruff lint errors across the codebase
• Corrected GitHub org name in all documentation and configuration files
• Updated test count in README to match actual collectible tests
• Softened overclaim language in documentation and paper

Improved

• Added test coverage reporting (pytest-cov) to CI pipeline
• Added USER directive and HEALTHCHECK to Dockerfile for security best practices
• Synchronized requirements.txt with pyproject.toml dependencies
• Removed duplicate THEORY_JOURNAL.md from docs
• Hyperlinked all arXiv references in README
• Added Pliny the Prompter attribution

[0.1.0] - 2026-02-27

Added

• 15 analysis modules for mechanistic interpretability of refusal mechanisms
• Analysis-informed pipeline (informed method) — closed-loop feedback from analysis to abliteration
• Ouroboros compensation — automatic detection and compensation for self-repair after excision
• Steering vectors — reversible inference-time guardrail removal (Turner et al. / Rimsky et al.)
• Community contribution system — --contribute flag and obliteratus aggregate for crowdsourced results
• 47 curated model presets across 5 compute tiers (CPU to multi-GPU)
• 10 study presets for reproducible ablation experiments
• 4 ablation strategies: layer removal, head pruning, FFN ablation, embedding ablation
• 4 abliteration methods: basic, advanced, aggressive, informed
• Web dashboard (docs/index.html) with config builder, model browser, results visualizer
• Gradio playground (app.py) — one-click obliteration + chat in the browser
• Colab notebook for zero-install usage
• Evaluation suite: refusal rate, perplexity, coherence, KL divergence, CKA, effective rank
• lm-eval-harness integration for standardized benchmarking
• Reproducibility framework with deterministic seeds and full metadata logging
• Telemetry (opt-in only, anonymized, allowlisted fields)
• 823 tests across 28 test files (incl. CLI dispatch, shared fixtures)
• Research paper (paper/main.tex) with geometric theory of refusal removal
• Dual license: AGPL-3.0 + commercial

Security

• trust_remote_code defaults to False — users must explicitly opt in
• All temporary paths use tempfile.gettempdir() for cross-platform safety
• Telemetry never collects model names, prompt content, file paths, or PII