Note: this page is a manually-refreshed mirror. The canonical source is CHANGELOG.md on GitHub.

Mirrored from the repo root CHANGELOG.md. Maintained automatically by release-please on each merge to main — see CI/CD pipeline for the full flow.

Changelog

All notable changes to this project are documented here.

The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

From v1.2.0 onwards this file is maintained automatically by release-please. Anything above v1.1.0 was seeded retroactively from the GitHub Releases history.

1.3.1 (2026-05-12)

Bug Fixes

• docs: repair slate scheme contrast and add brand-aligned theme variables (6516709)

Documentation

• add architecture hub landing with sequence diagram and key contracts (3fb717e)
• add branded 404 page with hero styling and recovery links (188491b)
• add CI/CD pipeline narrative covering branching releases and deployment (db48476)
• add development hub landing with conventional commits cheat sheet (5e28bb1)
• redesign landing with hero, agent grid, mermaid system diagram and stats (e6525cc)
• ship docs site revamp with CI/CD narrative and contrast fix (b64ce12)

1.3.0 (2026-05-12)

Features

• docs: point GitHub Pages at docs.f1stratlab.com via CNAME (52b3238)
• docs: update mkdocs site_url to docs.f1stratlab.com (cf6cda6)
• docs: wire docs.f1stratlab.com custom domain (1fcf6ea)

1.2.0 (2026-05-12)

Features

• docs: add F1 StratLab brand theme and external-image hook (0dc3db1)
• docs: launch mkdocs-material docs site with F1 StratLab branding (7f7006d)

Documentation

• add landing, getting started, thesis results and maintenance pages (defed4e)

1.1.1 (2026-05-12)

Bug Fixes

• ci: use PAT for release-please so required checks run on release PRs (b293342)
• ci: use PAT for release-please so required checks run on release PRs (e276d45)

Documentation

• add data/eval/README inventory for benchmark outputs (1f533fb)
• add data/rag_eval/README for the RAG ground-truth set (f0c54a1)
• add documents/images/README manifest for thesis figures (f7c25d3)
• document Conventional Commits convention in CONTRIBUTING (78049cb)
• seed CHANGELOG retroactively with release-please marker (c870bcc)

1.1.0 - 2026-05-11

Benchmark suite for the TFG thesis chapter 5 plus full English localization of strategy notebooks, scripts and evaluation artefacts. No model retraining and no breaking changes to runtime APIs.

• Four standalone benchmark scripts under scripts/bench_*.py with a shared BenchResult dataclass and Rich panel layout: pace baselines vs production XGBoost (MAE matches the 0.4104 s anchor within +/-0.001 s), Whisper turbo latency (P50 / P95 / mean), six sub-agent latency on a Suzuka 2025 fixture, and the sentiment + intent + NER pipeline on CPU and GPU.
• New notebooks/agents/N33_thresholds_and_calibration.ipynb with precision-recall sweeps for overtake (N12), safety car (N14) and undercut (N16), plus MC Dropout empirical coverage on the 20,284 tire-degradation sequences.
• New notebooks/agents/N30B_rag_benchmark.ipynb evaluating BGE-M3 1024d (production), MiniLM-L6-v2 384d and BGE-M3 chunk 256 over 15 ground-truth queries with Precision@1 / 3 / 5, MRR and latency.
• Figures relocated to documents/images/05_results/ (300 DPI), CSV + Markdown bench outputs under data/eval/ and data/rag_eval/.
• jiwer>=3.0.0 added to pyproject.toml as a forward-looking dependency. All bench scripts pass ruff check and ruff format --check on CI.
• Console entry points (f1-strat, f1-sim, f1-arcade, f1-streamlit) unchanged from v1.0.0.

1.0.0 - 2026-04-20

First stable release. Ships the three-window arcade experience, the full seven-model ML stack and the N25 to N31 multi-agent LangGraph orchestrator with FIA RAG over Qdrant.

• Three surfaces from one install: f1-sim CLI, f1-arcade three-window replay (2D circuit + PySide6 strategy dashboard + live telemetry window) and f1-streamlit post-race dashboard.
• Arcade runs the strategy pipeline locally without the FastAPI backend.
• Per-agent model outputs rendered live: lap time predicted vs actual with CI band, tire cliff percentiles, overtake and SC probabilities, stop duration percentiles, radio intents and regulation snippets.
• Six-tab reasoning panel with syntax-highlighted LLM narratives for each sub-agent plus the N31 orchestrator.
• Live telemetry window with 2x2 delta / speed / brake / throttle grid and rival overlay in two-driver mode.
• README slimmed to 85 lines with landing page link and F1 trademark disclaimer. Docs reorganised under docs/arcade plus five drawio architecture diagrams.
• Install via uv tool install git+https://github.com/VforVitorio/F1_Strat_Manager.git.

0.12.0 - 2026-04-15

Interfaces and distribution milestone. Closes R3 (Streamlit + Backend) and lands infrastructure for R2 (Arcade). The CLI (R1) stays untouched.

• Voice chat full rewrite: STT migrated from Nemotron to openai/whisper-small via transformers pipeline; TTS migrated from Qwen3 to edge-tts with a curated four-voice catalogue (Aria, Guy, Ryan, Sonia); LLM is now provider-agnostic via F1_LLM_PROVIDER.
• Voice chat UI redesigned end-to-end: Material icons, triadic palette, audio-reactive orb, native st.audio_input replacing the third-party recorder, voice selector dropdown wired end-to-end, health-check polling with spinner during cold starts.
• Chat charts: lap_times and race_data now show tyre compound on hover with per-driver pit-stop vlines annotated DRIVER - COMPOUND. Shared COMPOUND_COLORS palette mirrors the Rich palette used by the CLI.
• New POST /api/v1/strategy/simulate SSE endpoint streaming start / lap / summary events; ready for Arcade consumption.
• Breaking: streamlit bumped to >=1.37, audio-recorder-streamlit removed from deps. Backend Dockerfile now installs ffmpeg and libsndfile1 for browser WebM decoding.

0.11.0 - 2026-03-30

Multi-agent system complete plus the RAG regulation layer. Seven specialized agents coordinate under a Strategy Orchestrator to produce real-time pit strategy recommendations from live race data.

• N25 Pace Agent (XGBoost lap time + bootstrap CI), N26 Tire Agent (TCN with MC Dropout), N27 Race Situation Agent (LightGBM overtake plus safety car prior), N28 Pit Strategy Agent (pit duration quantiles plus undercut scorer), N29 Radio Agent (RoBERTa sentiment + SetFit intent + BERT-large NER + RCM parser), N30 RAG Agent (Qdrant + BGE-M3) and N31 Strategy Orchestrator (three-layer MoE-style routing into Monte Carlo simulation into GPT-4o structured synthesis).
• scripts/build_rag_index.py indexes the FIA Sporting Regulations into 2,279 BGE-M3 chunks. Retrieval scores 0.62 to 0.76 on demo queries.
• src/rag/retriever.py exports RagRetriever and query_rag_tool as reusable LangChain components imported by N31.
• GitHub Actions CI added: lint (ruff), typecheck (mypy), tests (pytest).
• SRP refactors across every agent notebook plus LangGraph computation graph visualization cells.

0.10.0 - 2026-03-22

Multi-agent infrastructure milestone. Two of seven sub-agents complete plus the full RAG indexing pipeline and the importable src/rag/ module.

• N25 Pace Agent wraps the N06 XGBoost model as a LangGraph ReAct agent and returns PaceOutput (lap time + delta vs session median + bootstrap CI P10 / P90 with N=200).
• N30 RAG Agent runs retrieval-augmented generation over FIA Sporting Regulations 2023 to 2025. Embedding via BAAI/bge-m3 (1024-dim), Qdrant local vector store, 2,279 indexed chunks.
• First active src/ module outside telemetry: src/rag/ exposes RagRetriever (singleton via get_retriever()) and the query_rag_tool LangChain tool.
• scripts/download_fia_pdfs.py scrapes FIA PDF URLs via DownloadConfig. scripts/build_rag_index.py performs PDF chunking, embedding and Qdrant upsert with hash-based deduplication.
• README files added for src/rag/, src/agents/, src/nlp/, src/strategy/ and src/data_extraction/ covering API surface and legacy status.

0.9.0 - 2026-03-17

NLP pipeline complete. All notebooks N17 to N24 shipped; the radio analysis pipeline is operational and integrated into the unified inference entry point used by the Strategy Agent.

• N17 labels 659 messages (610 clean after manual inspection of 49 post-race removals). N18 runs Whisper turbo ASR. N19 establishes a VADER rule-based baseline.
• N20 fine-tunes RoBERTa-base for three-class sentiment. N21 uses SetFit with ModernBERT-base for five-class intent (370 examples). N22 fine-tunes BERT-large CoNLL-03 with BIO tagging for nine F1 entity types (weighted F1 = 0.42 on 399 examples). N23 ships a deterministic rule-based RCM parser covering 25 event types with 100% Flag / DRS / SC coverage.
• N24 unified pipeline exposes run_pipeline(text) for team radio and run_rcm_pipeline(rcm_row) for race control messages on a single JSON schema. GPU end-to-end latency: mean 47.8 ms, P95 59.4 ms.
• Model weights and configs uploaded to VforVitorio/f1-strategy-models on Hugging Face, plus the N16 undercut artefacts that were missing from v0.8.1.

0.8.1 - 2026-03-13

Strategy ML suite: pit-stop prediction and undercut intelligence.

• N15 Pit Stop Duration: HistGradientBoostingRegressor at P05 / P50 / P95 on the normal physical window of 2.0 to 4.5 s. P50 MAE 0.487 s vs baseline 0.555 s. Coverage P05 to P95 is 70.5% on the test set.
• N16 Undercut Success: LightGBM binary classifier on 1,032 labeled pair-laps (2023 to 2025) with DRY_COMPOUNDS filter. AUC-PR 0.6739, AUC-ROC 0.7708, Platt-calibrated threshold 0.522. SHAP top features: pos_gap_at_pit, pace_delta, circuit_undercut_rate, tyre_life_diff.
• N12B Causal TCN Overtake archived as a valid negative result (AUC-PR ~0.10 vs N12's 0.5491). Confirms feature-engineered LightGBM wins on this dataset.
• Roadmap lists N17 to N24 for the upcoming NLP radio pipeline.

0.7.0 - 2026-03-05

ML foundation phase closes out. Two predictive models trained, validated on held-out 2025 data and exported under data/models/.

• N06 Lap Time Predictor: XGBoost delta-lap-time model with circuit clustering features, trained on 2023 to 2024 and tested on 2025. MAE 0.392 s. Features include fuel-corrected lap time, tyre life, compound, circuit cluster and race phase.
• N07 to N10 Tire Degradation Predictor: Temporal Convolutional Network in PyTorch with per-compound fine-tuning (SOFT / MEDIUM / HARD) and MC Dropout for uncertainty (N=50 forward passes at inference). Calibration JSON exported alongside the model weights.
• src/ module integration deferred to v0.9.0 (post-notebook phase). Tire compound mapping to C1 through C5 flagged as a future enhancement.

0.6.0 - 2026-02-12

Data engineering phase closes out. End-to-end pipeline from raw FastF1 telemetry to a clean feature-rich dataset ready to feed the ML models.

• Repo restructure: previous notebooks and code moved to legacy/ to preserve the original work. New structure built around the TFG architecture: notebooks/data_engineering/, notebooks/strategy/, src/strategy/, src/agents/, src/telemetry/.
• N01 download pipeline extended to support the 2025 season alongside 2023 to 2024. FastF1 naming inconsistencies aliased (Miami_Gardens, Spain vs Barcelona) for canonical cross-season names.
• N03 circuit clustering: K-Means with k=4 fitted on 2023 to 2024 and serialized with joblib. 2025 inference runs kmeans.predict() on the saved model without refitting. Las Vegas missing speed-trap data imputed with training means from the scaler.
• N04 feature engineering: 48-column dataset across ~45,000 clean racing laps. Fuel-corrected degradation (0.055 s/lap from Pirelli literature), sequential lap features, rolling 3-lap degradation rate via polyfit clipped to +/-2 s/lap, race-context fields, circuit cluster merge from N03. 2025 saved as a held-out test set.
• Dataset published to VforVitorio/f1-strategy-dataset on Hugging Face; scripts/download_data.py pulls everything locally.

0.1.1 - 2026-04-09

First CLI release (R1 milestone). Distributed as the f1_strat_manager-0.1.1-py3-none-any.whl wheel.

• Seven-agent multi-agent system (N25 to N31) on LangGraph.
• f1-sim CLI simulation with Rich Live rendering.
• No-LLM mode (ML + Monte Carlo simulation only).
• OpenF1 radio corpus with Whisper transcription.
• F1 strategic guard-rails baked into every sub-agent.
• Lazy Hugging Face data download on first run.
• Eight ML models (pace, tire degradation, overtake, safety car, pit duration, undercut) plus the NLP pipeline (sentiment, intent, NER) and RAG over FIA regulations.