v1.0.0

First release. Everything below shipped together.

Added

Data pipeline

• Synthetic dataset generator producing 883,231 rows across five related tables
  (telemetry, machines, errors, maintenance, failures) from a fixed seed.
• DataIngestion with format detection, DataValidator producing a structured
  ValidationReport, and an eight-dimension exploratory analysis script.
• DataPreprocessor: merges the five tables, engineers 63 rolling, lag, and
  time-since-last features, builds 24-hour-horizon binary labels, performs a
  chronological train/validation/test split, fits StandardScaler on the
  training split only, and slides 24-step windows into (N, 24, 63) tensors.

Model

• LSTM architecture (128 → 64 → 32 → 1, 149,825 parameters) with a hand-written
  GradientTape training loop implementing class weighting, early stopping,
  learning-rate reduction, and checkpoint resume explicitly.
• ModelEvaluator reporting AUC, precision, recall, F1, and a confusion matrix —
  and deliberately not accuracy, which is meaningless at a 1:864 positive rate.
• Threshold sweep over the precision-recall curve with cost weighting, including
  a lowest_cost_is_degenerate flag after a cost-optimal threshold selected on
  validation was found not to transfer.
• Recall measured over failure events as well as hourly sequences, with
  lead-time statistics.

Inference

• Predictor, which reuses the training feature code rather than
  reimplementing it, and verifies at load time that model, scaler, and feature
  contract describe the same thing.
• Feature reconciliation for categorical columns absent from a scored batch,
  using per-family defaults rather than a blanket zero fill.

GenAI layer

• Grounded report generation over LangChain, supporting OpenAI, Google Gemini,
  and a local Ollama model. Every figure a model quotes is supplied from the
  prediction record.
• Multi-turn maintenance assistant that declines questions the data cannot
  answer.

API and dashboard

• FastAPI service with nine endpoints, an exception hierarchy mapped onto status
  codes, and report generation isolated so a slow language model cannot delay a
  prediction.
• Point-in-time assessment: every prediction endpoint accepts an as_of
  timestamp and hides all later data — telemetry, errors, and maintenance alike.
• Streamlit dashboard as a pure HTTP client that holds no model and performs no
  scoring.

Infrastructure

• Two Docker images (API 2.87 GB, dashboard 803 MB) with multi-stage builds,
  non-root users, and health checks that report readiness rather than liveness.

• docker compose topology with health-gated startup ordering.

• GitHub Actions CI: lint, format, and type checks with pinned tooling; unit and
  integration tests; a dependency audit; image builds; and a smoke test asserting
  the API starts degraded rather than crashing when no model is mounted.

• Contributor-facing documentation: CONTRIBUTING.md, SECURITY.md,
  CODE_OF_CONDUCT.md, and this changelog. Development conventions, the layering
  rule, and the correctness invariants are now documented in CONTRIBUTING.md
  and docs/architecture.md.

• Optional dependency extras for the LLM providers, which are imported lazily and
  were previously undeclared: pip install -e ".[ollama]", ".[google]", or
  ".[llm]".

• Two configuration tests that run without a trained model, so they execute in
  CI: one asserting the risk bands ascend and that RISK_BAND_HIGH equals
  PREDICTION_THRESHOLD, one asserting the served threshold matches the
  committed evaluation report. The equivalent assertions previously lived in a
  test that skipped whenever no model artifact was present — which is every CI
  run.

Changed

• Training is seeded end to end, so published metrics are reproducible rather
  than merely reported. The model was retrained under the seed —
  test F1 0.9086 at threshold 0.3415, 8 of 8 failure events caught with a
  median 23.5 hours of warning.
• Early stopping and checkpoint selection monitor val_f1 rather than val_auc,
  which saturates under this class imbalance and selects on noise.
• Model selection moved onto a dedicated validation split so the test set is
  scored exactly once, at a threshold chosen without observing it.
• Dashboard risk colours meet WCAG AA contrast at their rendered size, asserted
  numerically by tests rather than judged by eye.

Fixed

• An abseil symbol collision between TensorFlow and Apache Arrow that deadlocked
  the process at 0% CPU with no traceback. Import order is now load-bearing and
  documented at each site that depends on it.

• The API image shipped Streamlit, which nothing in src/ imports; dependencies
  are now split per service.

• /fleet's cache is keyed by as_of and bounded by an LRU, having previously
  been a single unbounded slot that would serve a present-day answer to a
  request about a past date.

• All 159 outstanding type errors resolved, making mypy blocking in CI.

• Concurrent /fleet requests stampeded the cache. Route handlers are
  synchronous, so FastAPI runs them in a threadpool; checking the cache and
  filling it were unsynchronised, so every request arriving for an uncached
  as_of while another was computing missed too and recomputed the same
  answer. Measured: four concurrent requests for one cold timestamp produced
  four full fleet scorings and made every caller wait 57.8 s for work that
  takes 13.4 s once. Now serialised behind a compute lock, with cache reads on
  a separate short-lived lock so a hit still returns in milliseconds while a
  cold scoring is in flight.

• ModelEvaluator.evaluate() wrote invalid JSON on single-class input. The
  guard was written as except ValueError, which scikit-learn stopped raising in
  1.6 — it now warns and returns nan. The fallback had become unreachable, so
  nan reached models/metrics.json, where json.dump writes it as a bare
  NaN token that no strict JSON parser accepts. Now tests the precondition
  directly.

• The confusion matrix could change shape. confusion_matrix() inferred its
  labels from the data, returning a 1×1 matrix when only one class was present,
  despite being published and read as [[tn, fp], [fn, tp]]. Pinned with
  labels=[0, 1], which is identical for every two-class input.

• make setup failed on every platform except Apple Silicon, having
  hardcoded a Homebrew ARM interpreter path. It now delegates to
  scripts/setup.sh, which already resolved python3.12 → 3.11 → 3.10 from
  PATH. That script is also now independent of the working directory it is
  invoked from.

• Corrected published figures that had drifted from the artifacts they describe:
  the README summary quoted the pre-retrain model's false-alarm count and lead
  time, docs/RESULTS.md reported an early-stopping epoch that contradicted
  models/training_history.json, and test counts were stale in six places.

• Corrected two docstrings that described behaviour the code does not have —
  most importantly create_sequences(), whose description of the label pairing
  implied a leak the implementation does not contain.

---