CHANGELOG.md

Changelog

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

[0.1.0] — 2024-01-15

Added

Core Flywheel

• @capture decorator for zero-friction production signal capture
• SignalCapture with async queue, configurable flush interval, and PII scrubbing
• SimHash (64-bit LSH) for near-duplicate interaction detection
• InMemoryStore and SQLiteStore for interaction persistence
• FeedbackCollector for explicit and implicit user feedback

Annotation

• Annotator with multi-criteria LLM-as-judge scoring
• Configurable Criterion with weights and normalization
• Batched concurrent annotation (asyncio.Semaphore, max 10 parallel)
• Cost budget enforcement per annotation run
• AnnotationCalibrator with bias detection and linear score correction
• Pearson r, Spearman rho, and Cohen's kappa for inter-annotator agreement

Curation

• DataCurator with 5 strategies: Quality, Diversity, Uncertainty, Core-Set, Active Learning
• Greedy diversity selection via SimHash Hamming distance
• k-center core-set selection (2× optimal approximation)
• Boundary uncertainty sampling for active learning
• Deduplication with configurable Hamming distance threshold
• DatasetExporter: Alpaca, ShareGPT, OpenAI, ChatML, HuggingFace, Parquet formats

Training

• Trainer with LoRA/QLoRA fine-tuning via HuggingFace PEFT
• EWC (Elastic Weight Consolidation) forgetting prevention
  • Fisher Information Matrix diagonal approximation
  • Configurable λ regularization strength
• Replay buffer: Ring buffer, Prioritized Experience Replay, Cluster-based
• Constitutional constraint checking during training
• TrainingHooks with W&B and MLflow integrations

Deployment

• ModelDeployer with Blue-Green, Canary, and Shadow strategies
• Bayesian A/B testing with Thompson Sampling
• Auto-rollback on quality drop
• Per-metric threshold configuration

Safety & Alignment

• PIIScrubber with Luhn validation, entropy-based API key detection
• ConstitutionalChecker with default probe suite (identity, malware, PII, prompt injection)
• Hard vs. soft constraint classification
• Constitutional probes run before every deployment

Analytics

• LearningCurve with linear trend regression and R² goodness-of-fit
• ImprovementReport with automated insights and recommendations
• Cost-quality trajectory tracking

Regression Guard

• Welch's t-test for statistical deployment gating
• Cohen's d effect size estimation
• Percentile bootstrap 95% confidence intervals
• Power analysis for evaluation dataset sizing

Infrastructure

• ColdStartHandler for bootstrapping with no production data
• Prometheus metrics for full flywheel observability
• Slack/webhook notifications for lifecycle events
• FastAPI REST API for remote management
• CLI: continuum cycle --task <name> runs the full flywheel
• Docker Compose with PostgreSQL, Redis, annotation/training workers
• GitHub Actions CI with Python 3.11/3.12 matrix
• PostgreSQL schema with triggers and optimized indexes
• Reservoir sampling (Algorithm L, weighted, stratified)