MinxiongHydroCast v0.1.6

MinxiongHydroCast v0.1.6 — Legacy Documentation Retirement

Release date: 2026-08-18

Release tag: v0.1.6

GitHub Release name: MinxiongHydroCast v0.1.6

Highlights

• Removes the retired AIWeatherForecast predecessor document and its active documentation-index
  route.
• Establishes MinxiongHydroCast as the sole current project identity in user-facing documentation.

Compatibility and safety

This is a documentation and governance patch. It changes no package behavior, data-root content,
runtime state, or external artifact contract. The v0.1.4 historical changelog entry remains intact
as evidence of the prior migration decision.

Verified release evidence

• Documentation references contain no active AIWeatherForecast route.
• Ruff, the full test suite, package build, and isolated installed-wheel smoke checks pass.

MinxiongHydroCast v0.1.5

MinxiongHydroCast v0.1.5 — Safe Data-root Relocation

Release date: 2026-08-18

Release tag: v0.1.5

GitHub Release name: MinxiongHydroCast v0.1.5

Highlights

• Adds mhc data relocate-root for the metadata phase of an external data-root move or copy.
• Rewrites catalog root metadata and CWA collection absolute output paths, then recalculates the
  resulting collection artifact records.
• Creates a fresh dataset verification report whenever the dataset catalog changes.

Compatibility and safety

The command defaults to a no-write preflight. --apply is required to write changes and retains
the original JSON documents below the new data root. It does not move or delete payload files.
Unexpected checksums fail the preflight unless an operator explicitly acknowledges them with
--refresh-artifact-checksums after investigation.

Verified release evidence

• Collection, event-evidence catalog, dataset-catalog, verification-report, idempotency, and
  unrelated-checksum rejection paths are covered by synthetic tests.
• Ruff, the full test suite, package build, and isolated installed-wheel smoke checks pass.

MinxiongHydroCast v0.1.4

MinxiongHydroCast v0.1.4 — Data-root and Deployment Governance

Release date: 2026-08-18

Release tag: v0.1.4

GitHub Release name: MinxiongHydroCast v0.1.4

Highlights

• Makes MINXIONGHYDROCAST_DATA_ROOT the canonical external artifact setting while retaining
  the prior research-root setting as a temporary compatibility fallback.
• Separates the documented workspace into Git checkout, durable runtime, and durable data assets.
• Records verified private deployment metadata after installation and rejects a package-version
  mismatch.

Compatibility and safety

Existing event-evidence and dataset catalogs that contain research_root remain readable. New
catalog writes use data_root; artifact paths stay relative and checksummed. No runtime, raw data,
model artifact, or backup is moved or committed by this release.

Verified release evidence

• The full test suite passes.
• Ruff, deployment-shell syntax validation, and the event-discovery CLI help check pass.
• The source distribution and wheel build successfully.
• The installed-wheel demo/API smoke test passes in an isolated environment.

MinxiongHydroCast v0.1.3

MinxiongHydroCast v0.1.3 — Version Metadata Alignment

Release date: 2026-08-18

Release tag: v0.1.3

GitHub Release name: MinxiongHydroCast v0.1.3

Highlights

• Aligns the Python package version, runtime __version__, citation metadata, issue-template
  example, changelog links, and README release table at 0.1.3.
• Makes the patch-level release state explicit without claiming new forecast capability.

Verified release evidence

• The full test suite passes.
• The source distribution and wheel build successfully.
• The installed-wheel CLI smoke test verifies the v0.1.3 release tag matches the wheel version.

Research and safety boundary

This is a metadata-only patch release. The v0.1.2 research findings, Persistence comparison, and
forecast-publication gate remain unchanged. The project still does not claim forecast readiness or
provide an emergency decision service.

MinxiongHydroCast v0.1.2

MinxiongHydroCast v0.1.2 — Deterministic Optical-flow Baseline

Release date: 2026-08-17
Release tag: v0.1.2
GitHub Release name: MinxiongHydroCast v0.1.2

Highlights

• Adds a deterministic CPU OpticalFlowNowcaster based on FFT phase correlation and global
  integer translation.
• Adds event-level and lead-time comparisons with Persistence using RMSE, MAE, CSI, POD, and FAR.
• Adds a public-safe aggregate report for Persistence, Optical Flow, and Tiny U-Net that excludes
  private archive and checkpoint paths.
• Adds optical-flow evaluation artifacts to the external research dataset build without changing
  the forecast promotion gate.

Verified release evidence

• pytest -q: 316 passed.
• Ruff, Python compilation, wheel/sdist construction, and installed-wheel CLI smoke tests pass.
• The formal five-event independent aggregate is reproducible.
• Optical Flow remains a research benchmark: independent RMSE is 13.657048 versus Persistence
  9.258762, and CSI is 0.135459 versus 0.152099.

Research and safety boundary

This release does not claim forecast readiness or improve the publication gate. The formal dataset
still needs broader human-reviewed weather-regime diversity, and all official raw data, private
event evidence, trained weights, and credentials remain outside the repository.

Future GitHub Releases should use the exact display-name format MinxiongHydroCast vX.Y.Z; the
release workflow validates that the display name matches its tag.

v0.1.1 — Portable Demo and Reusable Contracts

v0.1.1 — Portable Demo and Reusable Contracts

MinxiongHydroCast v0.1.1 turns the Minxiong reference implementation into a lighter,
reproducible, contributor-ready official-source hydrometeorological data-quality platform. It
does not expand forecast capability or weaken any observation, model, notification, or publication
gate.

Release date: 2026-08-01

Highlights

• Adds a credential-free docker compose up --build demo with deterministic synthetic
  observations, an operator dashboard, health/readiness endpoints, Prometheus metrics, and an
  intentionally blocked forecast endpoint.
• Reduces the base wheel to NumPy, Pydantic, and Requests; scraping, reporting, and modeling use
  explicit capability extras.
• Replaces more than 30 installed executables with one grouped mhc command while retaining
  transition aliases for legacy command tokens.
• Adds strict packaged region profiles, a non-authoritative Minxiong working boundary, an example
  profile, and profile-driven regional feature coverage.
• Publishes a source-adapter Protocol and shared contract checks for stable source identity,
  adapter version, timezone-aware fetch time, provenance, and SHA-256 integrity.
• Adds contributor guidance, issue forms for bugs/features/source drift, a pull-request template,
  a clean-wheel CI smoke test, and a PyPI Trusted Publishing release workflow.

Verified release evidence

• The deterministic test suite passes for all non-network tests; loopback HTTP behavior was also
  verified through the built Compose stack.
• Ruff, Python compilation, wheel construction, Compose validation, and installed-wheel smoke
  tests pass.
• The installed wheel exposes only mhc and packages both region profiles and their GeoJSON
  boundaries.
• The Compose demo returns /healthz=200, /readyz=503, readiness metric 0, and an unavailable
  experimental forecast, as required for synthetic data.
• The 60-second dashboard walkthrough contains 30 decoder-validated GIF frames.

Install and try the demo

Install the base wheel from PyPI:

python -m pip install "minxiong-hydrocast==0.1.1"
mhc collect --region minxiong --mode demo --once
mhc serve --host 127.0.0.1 --port 8080

Or run the isolated synthetic stack from a source checkout:

docker compose up --build

Live official-source collection still requires locally protected CWA/WRA credentials and the
operational runbook. The Docker demo never contacts live sources.

Upgrade notes

Standalone minxiong-hydrocast-* executables are no longer installed. Use the grouped commands:

Previous executable	v0.1.1 command
minxiong-hydrocast-operations	mhc collect
minxiong-hydrocast-serve	mhc serve
minxiong-hydrocast-dataset-build	mhc dataset build
minxiong-hydrocast-event-discover	mhc event discover
minxiong-hydrocast-event-review	mhc event review
minxiong-hydrocast-evaluate-baselines	mhc model evaluate
minxiong-hydrocast-backup	mhc operations backup

The derived snapshot dataset is now region_features instead of minxiong_features. The generic
API route is /api/v1/features/region; /api/v1/features/minxiong remains as a read-only
compatibility route for this release.

Install optional capabilities only where needed:

python -m pip install "minxiong-hydrocast[scraper]"
python -m pip install "minxiong-hydrocast[model]"
python -m pip install "minxiong-hydrocast[report]"

Safety and research boundary

• Synthetic records remain demo_fixture and cannot satisfy readiness or forecast publication.
• Region-profile geometry states its authority; the Minxiong polygon is an operational working
  bound, not an official administrative boundary.
• Five formal radar events remain insufficient for broad weather-regime claims.
• Persistence remains the mandatory benchmark, and forecast publication remains blocked because
  Tiny U-Net does not consistently pass CSI and lead-time gates.
• This release contains no official raw data, credentials, private event notes, model checkpoint,
  or trained weights.

Next public work

• Optical-flow baseline
• Reviewed weather-regime diversity
• v0.2 dataset-building protocol and event manifest

v0.1.0 — Observation Service

MinxiongHydroCast v0.1.0 establishes the first supported product boundary: a localhost-only,
official-source hydrometeorological observation service for Minxiong/Chiayi, with reproducible
research interfaces kept behind explicit publication gates.

Highlights

• CWA rain-gauge, WRA rainfall-warning, and WRA IoW flood-sensor ingestion with strict schemas,
  freshness, provenance, and immutable SHA-256 snapshots.
• Bounded retry for transient/malformed WRA pages and changing measurement/catalog join
  transactions, with structured retry telemetry.
• Read-only API, health/readiness, Prometheus metrics, real operator dashboard, verified
  backup/restore, and rolling shadow evaluation.
• Reproducible radar-event research, human review boundary, Persistence benchmark, and weighted
  Tiny U-Net evaluation.
• Python 3.11/3.13 CI, CodeQL, Dependabot security updates, secret scanning/push protection, and
  protected main.

Safety status

Forecast publication and automated risk notification remain disabled. The weighted Tiny U-Net
improves aggregate RMSE on the independent events but does not consistently pass CSI and
lead-time gates. The 2026-07-29 shadow window passed success/readiness targets but remained blocked
by a 50.98-minute maximum gap and no confirmed heavy-rain period.

Public package boundary

This release contains code, schemas, deployment templates, synthetic samples, aggregate results,
and a model card. It does not contain credentials, live snapshots, host configuration, official
raw research downloads, CCTV, unpublished labels, checkpoints, or trained weights. The MIT
license applies to repository code only.

See the complete release notes,
architecture, and
baseline results.