Skip to content

CHANGELOG.md

Latest commit

 

History

History
226 lines (184 loc) · 8.76 KB

CHANGELOG.md

File metadata and controls

226 lines (184 loc) · 8.76 KB
Raw

Changelog

All notable changes to RewardHackWatch will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.3.0] - 2026-03-01

Added

React Frontend

  • Full React 18 + TypeScript + Tailwind CSS v4 dashboard replacing Streamlit
  • 9-page SPA: Dashboard, Quick Analysis, Timeline, Alerts, Cross-Model, CoT Viewer, Eval Workbench, Session Logs, Settings
  • Dark-mode developer tool UI with #20232B theme and JetBrains Mono
  • Stat cards with accent-colored left borders
  • Recharts-based visualizations (donut, bar, area, line, radar charts)
  • Code splitting with manual chunks for react and recharts

Eval Workbench (formerly JSONL Analyzer)

  • Drag-drop JSONL import with auto-detect schema (conversation, completion, I/O, Q&A)
  • Raw JSONL paste input with parse preview
  • Custom evaluation rules: regex, keyword, length, LLM judge types
  • 8 built-in rule templates (rogue code, mock exploit, safety, deceptive CoT, etc.)
  • Rule management: add, toggle, delete, weight and severity configuration
  • Batch evaluation engine with progress bar
  • Weighted scoring system with A-F grade badges
  • Per-entry detail view with full conversation and rule results
  • Per-rule pass rate charts and grade distribution visualization
  • Model 1 vs Model 2 side-by-side comparison with winner indicator
  • JSON and CSV export

Settings

  • Dual LLM provider cards: General LLM Provider and Independent Review Provider
  • Provider dropdown (Anthropic / OpenAI) with dynamic model list per provider
  • Eye icon toggle for API key visibility
  • Number inputs for temperature and max tokens
  • Auto-review toggle on Independent Review Provider card
  • Per-section save buttons
  • Settings persistence via localStorage

Typography

  • Space Grotesk font for headings, stat numbers, chart titles, sidebar logo
  • Inter font for body text and labels
  • JetBrains Mono for code blocks and monospace elements

Changed

  • License changed to Apache 2.0
  • Version bumped to 1.3.0
  • Color scheme updated: bg #20232B, cards #292A30, sidebar #1B1D24, borders #363840
  • All hardcoded colors in chart components replaced with theme variables
  • Sidebar label: "JSONL Analyzer" renamed to "Eval Workbench"
  • Sidebar logo: "RewardHackWatch" as single word
  • All chart containers minimum 320px height
  • American spelling throughout ("analyzed" not "analysed")
  • Dashboard category bar chart: shorter axis labels, wider bar area
  • Dashboard chart title: "Detections by Category" shortened to "Category Breakdown" to prevent truncation in 3-column layout
  • Model labels in Settings now display exact model strings instead of friendly names
  • Settings cards now full-width layout instead of narrow centered column
  • Eval Workbench format labels: generic "Model 1" / "Model 2" / "Custom Format" instead of provider-specific names

Fixed

  • CITATION.cff placeholder values replaced with actual metadata
  • SECURITY.md placeholder email replaced
  • Organization name normalized to "Aerosta" across all files
  • Sidebar truncation of long nav labels
  • Removed all em dashes from README
  • CHANGELOG v1.0.0 date corrected to 2025-12-09
  • Category label "Deceptive Cot" corrected to "Deceptive CoT" via acronym-aware formatter
  • TypeScript build error in Eval Workbench chart tooltip formatter

[1.2.0] - 2026-02-28

Added

Public API

  • RewardHackDetector high-level facade class matching README Quick Start examples
  • AnalysisResult.detections property (alias for pattern_matches)
  • RewardHackDetector.calibrate_threshold() for dynamic threshold calibration
  • ThresholdCalibrator with percentile, isotonic, and beta calibration methods
  • rewardhackwatch calibrate CLI command for threshold calibration on clean data

Synthetic Data & Benchmarks

  • MockExploitGenerator with 8 code templates for mock/monkeypatch exploits
  • CleanTrajectoryGenerator for balanced dataset creation
  • HackBenchDataset standardized benchmark (4,300+ trajectories, 9 categories)
  • scripts/generate_synthetic_data.py and scripts/build_hackbench.py

Temporal Modeling

  • StepLevelFeatureExtractor for per-step feature extraction
  • StepSequenceDataset with padding and attention masks
  • TemporalTrainer training pipeline for AttentionClassifier
  • Positional encoding in AttentionClassifier for step ordering

Research Tools

  • TransferStudyRunner for cross-model transfer matrix experiments
  • EvasionAttackSuite with 5 adversarial attack types
  • CausalRMGI with Granger causality testing and intervention analysis
  • scripts/run_transfer_study.py and scripts/run_evasion_tests.py

Dashboard

  • Dark-mode developer tool UI (#1a1a2e theme)
  • Stat cards with accent-colored borders
  • Donut charts for category distribution
  • Per-category F1 bar charts
  • Cross-Model comparison tab
  • Recent Analysis table
  • Quick Analysis as default landing tab
  • PDF and JSON export buttons

Fixed

  • __version__ was "0.1.0" instead of matching pyproject.toml
  • __author__ was "Your Name" instead of "Aerosta"
  • RewardHackDetector class referenced in README didn't exist
  • result.detections referenced in Quick Start didn't exist
  • Python 3.9 compatibility: added from __future__ import annotations to 52 files
  • Missing [dashboard] optional dependency group in pyproject.toml

Changed

  • Version bumped to 1.2.0
  • Development status upgraded from Alpha to Beta

[1.0.0] - 2025-12-09

Added

Core Detection System

  • Multi-layer detection architecture combining pattern matching and ML classification
  • Pattern detector with 40+ regex patterns for code-level reward hacks
  • Chain-of-thought red flag detection based on OpenAI's monitoring research
  • DistilBERT-based ML classifier fine-tuned on 4,314 MALT trajectories
  • Optimal threshold calibration (0.02) for imbalanced classification

RMGI (Reward-Misalignment Generalization Index)

  • Novel metric for tracking correlation between hack and misalignment behaviors
  • Rolling window correlation computation (default window size: 10)
  • PELT changepoint detection for behavioral transition identification
  • Configurable transition detection with multi-condition thresholds
  • Formal mathematical specification in docs/RMGI_DEFINITION.md

Multi-Judge System

  • Claude Judge: Claude Opus 4.5 integration for trajectory analysis
  • Llama Judge: Local Llama 3.1 (8B) via Ollama for cost-free analysis
  • Ensemble voting for multi-signal detection

API and Dashboard

  • FastAPI server with /analyze and /status endpoints
  • Streamlit dashboard for real-time trajectory monitoring
  • RMGI visualization with rolling correlation plots
  • Risk level assessment (none/low/medium/high/critical)

Benchmark Suite (RHW-Bench)

  • Integration with SHADE-Arena evaluation framework
  • MALT trajectory loader for training data
  • Synthetic trajectory generation for ablation studies
  • Benchmark runner with multiple dataset support

Documentation

  • Technical report with full methodology and results
  • RMGI formal specification document
  • API reference documentation
  • Installation and quickstart guides
  • LaTeX paper template for NeurIPS submission

Ablation Studies

  • Detection component ablation (ML only vs pattern vs combined)
  • Baseline method comparison (keyword, BoW+LogReg, regex, DistilBERT)
  • Detection delay analysis
  • RMGI parameter sensitivity analysis

Experimental Results

Classification Performance

Metric 5-Fold CV Held-Out Test
F1 Score 87.4% +/- 2.9% 89.7%
Precision 91.0% +/- 2.6% 89.7%
Recall 84.2% +/- 4.0% 89.7%
Accuracy 99.0% +/- 0.2% 99.3%

Baseline Comparisons

Method F1 Precision Recall
DistilBERT (Ours) 89.7% 89.7% 89.7%
Regex Patterns 4.9% 2.6% 38.5%
BoW + LogReg 7.0% 4.0% 28.2%
Keyword Matching 0.1% 0.1% 5.1%

RMGI Optimal Configuration

  • Window size: 10
  • Correlation threshold: 0.7
  • Transition detection F1: 0.452
  • False positive rate: 4.3%

Statistical Significance

  • 95% CI for F1: [84.8%, 90.0%]
  • t-test vs keyword baseline: p < 0.001
  • Brier Score: 0.033
  • Expected Calibration Error: 0.027

Dependencies

  • Python 3.10+
  • PyTorch 2.0+
  • Transformers 4.30+
  • FastAPI 0.100+
  • Streamlit 1.30+
  • scikit-learn 1.3+
  • ruptures (optional, for PELT changepoint detection)

Known Limitations

  • Mock/monkeypatch exploits: 0% F1 (underrepresented in training data)
  • Threshold requires recalibration if base rate changes significantly
  • RMGI requires minimum 5 timesteps for reliable correlation

[Unreleased]

Planned

  • Activation probing for deception detection
  • Steganography detection in CoT token choices
  • Multi-agent scenario monitoring
  • Real-time streaming API with WebSocket alerts