Project status¶
What swatplus-builder establishes today — and what it does not — stated plainly.
Keeping this page prominent is deliberate. The whole point of the system is that claims track evidence; that discipline applies to the project's own claims too.
Where it stands¶
Alpha research software
In the dated 2026-07-02 objective-suite snapshot, 0 of 11 complete
workflows were promoted to the Gate-verified tier (stored as research_grade).
The gates were not relaxed to manufacture passes. A later focused positive
case is reported separately below and does not revise that snapshot.
This is not a failure of the pipeline — it is the pipeline working as designed. Most basins build and run the engine cleanly, and several calibrate with real, independently verified improvement — and still do not earn a Gate-verified workflow status, because a provenance, physical-realism, outlet-scope, or skill gate is not met. A blocked claim is classified evidence, not a crash.
What is solid today¶
- A complete gauge-to-model build in Python, with no desktop GIS.
- The locked calibration protocol: lock → calibrate → independently verified rerun.
- Claim governance: runtime gates emit a tier, and metrics never promote themselves.
- A machine-readable evidence bundle on every run, including typed, evidence-backed refusals.
- A 13-tool agent interface and a container baseline.
Later focused cases, honestly labeled¶
Two deliberately retained 2026-07-10 cases exercise contrasting package decisions. They are mechanism demonstrations, not a representative basin sample and not additions to the dated objective-suite snapshot.
| Basin | Benchmark NSE / KGE | Locked verification NSE / KGE | Withheld 2016--2019 NSE / KGE | Package decision |
|---|---|---|---|---|
01547700 |
0.2875 / 0.3522 | 0.3536 / 0.5856 | 0.3742 / 0.6260 | Gate-verified workflow tier (research_grade); terrain/lapse-derived subclaims remain blocked |
03349000 |
-0.4202 / 0.0841 | -0.0199 / 0.3548 | 0.0087 / 0.3010 | exploratory; final skill and transfer requirements did not pass |
The negative control preserves real metric improvement without promoting the calibrated skill claim. The positive case shows that a complete governed path can reach the package-defined tier while still retaining blocked subclaims.
Known limitations¶
- Drainage / outlet scope. For several basins the selected outlet carries only part of the network's flow, which blocks a basin-scale skill claim. Single-outlet delineation is the priority fix.
- Evaluation breadth. The two focused cases were selected to expose contrasting decisions and do not establish broad basin performance.
- Peak flows. The positive case still underrepresents some sharp observed peaks despite passing the package policy.
- Evidence schema. The on-disk evidence format is stabilizing toward a formally versioned schema; consume it defensively for now.
- Archival reproduction. Version 0.7.14 is the current hardened source state. A public checksummed evidence archive and archival DOI remain pending; the historical July run-engine checksum cannot be reconstructed.
Version
The current source version is 0.7.14
(pyproject.toml / swatplus_builder.__version__).
The contribution, stated precisely¶
swatplus-builder is an evidence-producing, claim-governed modeling workflow that makes both successes and limitations inspectable. The refusals are part of the result — not something hidden behind the headline number.