# Natural-seepage candidate: implemented exploratory comparison

Measured by `scripts/build_natural_seepage_candidate.py` and recomputed here by `scripts/report_natural_seepage_candidate.py` from retained cell-series arrays and summary files. The interactive comparison is `index.html` beside this record; regenerate using the committed HTML template.

## What was implemented

Country-independent elevation-area seepage on 6090 existing diffuse-boundary cells in the border inspection window. Ten equal-area DEM bands per cell discharge through standard native DRN records: Q=sum(A*f/R*max(h-z,0)). Explicit line drains, rivers and northern outlets are preserved. No new agricultural drains or production changes. Candidate labels refer to assumed resistance in days; lower resistance means easier discharge. `flat_r1000` uses model-top elevation with the same total conductance as `terrain_r1000`, isolating surface geometry (including source DEM/model-top differences).

Both controls reproduce original daily heads exactly; read `april/replay_results.json` and `september/replay_results.json`, key control. All native arms terminate normally. Largest layer-relative budget residual: 1.84e-12. Largest independently reconstructed candidate DRN flux discrepancy: 4.55e-10 m3/day. Only the two diffuse DRN files differ in each matched candidate. This establishes implementation, not physical validation.

## Results on the cells actually changed

Counts are cells above MEAN model ground at least once during the 31 saved days, not observed flooding. Flow is mean total potential seepage from those cells. Median head change is candidate minus matched control at the final day on the same changed cells.

| window | arm | country | ever_above_mean_ground | mean_seepage_m3_day | median_final_head_delta_m |
| --- | --- | --- | --- | --- | --- |
| april | control | KAZ | 1238 | 3846527.0 | 0.0 |
| april | control | KGZ | 393 | 3651850.0 | 0.0 |
| april | flat_r1000 | KAZ | 1349 | 3096456.0 | 0.2 |
| april | flat_r1000 | KGZ | 1477 | 2401307.0 | 3.31 |
| april | terrain_r250 | KAZ | 948 | 5675278.0 | -0.11 |
| april | terrain_r250 | KGZ | 636 | 3782740.0 | -0.0 |
| april | terrain_r1000 | KAZ | 1292 | 3597298.0 | 0.05 |
| april | terrain_r1000 | KGZ | 1326 | 2617184.0 | 2.93 |
| april | terrain_r4000 | KAZ | 1435 | 2302821.0 | 0.44 |
| april | terrain_r4000 | KGZ | 1483 | 2171519.0 | 11.46 |
| september | control | KAZ | 55 | 59234.0 | 0.0 |
| september | control | KGZ | 2 | 233627.0 | 0.0 |
| september | flat_r1000 | KAZ | 58 | 30187.0 | 0.0 |
| september | flat_r1000 | KGZ | 3 | 1947.0 | 0.0 |
| september | terrain_r250 | KAZ | 14 | 1624400.0 | -0.03 |
| september | terrain_r250 | KGZ | 2 | 919939.0 | -0.01 |
| september | terrain_r1000 | KAZ | 35 | 523893.0 | -0.01 |
| september | terrain_r1000 | KGZ | 2 | 279503.0 | -0.0 |
| september | terrain_r4000 | KAZ | 53 | 143075.0 | -0.0 |
| september | terrain_r4000 | KGZ | 3 | 73961.0 | 0.0 |

## What the comparison demonstrates

Measured on the changed KGZ cells: the middle resistance raises the April median final head by 2.93 m; the weaker outlet raises it by 11.46 m. The result is strongly dependent on an unmeasured outlet resistance. Replacing the inherited drains can expose high heads, but cannot by itself establish which high heads are realistic. The weak-outlet arm is not suitable for adoption on this evidence. Seepage can also occur below the cell-average ground elevation through the lower terrain bands, so counts of red cells are not a measure of total discharge.

The formulation uses the documented DRN head-minus-outlet relation, with multiple elevation bands to approximate increasing wet area. See [USGS DRN documentation](https://water.usgs.gov/nrp/gwsoftware/ModelMuse/Help/drn_object_pane.html) and [the MODFLOW surface-seepage example](https://modflow6-examples.readthedocs.io/en/latest/_examples/ex-gwf-drn-p01.html). Those sources support the numerical mechanism, not the local resistance values. Unlike the routed MODFLOW example, these isolated replays do not pass the removed water to streams.

## Physical interpretation and limits

This is a tested FREE-OUTLET potential-seepage formulation, not a completed flood-inundation or routed surface-water model. Water removed by these DRN entries leaves this groundwater replay. Actual receivers, their capacity, backwater, ponding storage, reinfiltration and changed surface feedback have not been qualified. Existing line/rivers retain their inherited uncertainties. A common formula removes country as the selector of diffuse drainage but does not erase real or inherited aquifer-property differences.

The terrain bands allow low patches to discharge before the cell-average ground is reached. Consequently fewer above-average-ground cells do not necessarily mean less seepage. Broad within-cell slope may also cause this representation to drain too readily because the groundwater head itself is represented by one value per cell. Measured: 672 cells have DEM p90-p10 relief above 20 m; 68 have absolute mean-DEM versus model-top differences above 2 m. Those are diagnostic flags, not adopted exclusion thresholds. Surface-model vegetation/building bias and absolute elevation errors are not corrected. The ten-band maximum sampling-curve approximation error at R1000 is 0.000582831 m/day across the sampled head range; this is numerical approximation error, not total physical uncertainty.

The resistance bracket (250, 1000, 4000 days) is ASSUMED, not measured locally or selected by calibration. Each window starts from its ORIGINAL native state. Candidate head changes include storage adjustment; these are not equilibrated seasonal forecasts or full-history reruns. The April and September comparison is not a validation against 2017 observations.

Owner clarification: the MCHS polygons are crude hand-drawn KGZ-only indications of where to investigate, not confirmed upwelling boundaries or calibration targets. Their absence does not establish dry ground, including within KGZ. KAZ was not assessed by this source. They are an optional map overlay only. The earlier `condition_evidence_comparison.csv` is retained as historical descriptive output and is no longer regenerated or used to assess fit. No observational goodness-of-fit score is claimed. Further process diagnosis is in [process investigation](process_investigation/findings.md).

## Decision

Keep this as an exploratory candidate and sensitivity envelope. Do not adopt one resistance or replace all KGZ drainage merely because the map looks smoother. Before adoption: qualify low-ground elevations and connected receivers in representative landforms; obtain shallow groundwater and collector/spring discharge observations; represent retaining depressions explicitly; run a fully coupled, sufficiently equilibrated comparison and independent scientific review. No production pointer or owner ruling was changed.

## Reproduce

With `PYTHONPATH=diagnostics/coupled-improvement-2026-09-14/scratch/runtime` and `.venv/bin/python`, run `scripts/build_natural_seepage_candidate.py build`, then `run --window april` and `run --window september`. Run `scripts/report_natural_seepage_candidate.py` to regenerate this record and the comparison. The native runner refuses to overwrite previous raw run folders. Terrain input identities and checks are in `terrain_receipt.json`; native hashes and initial-state identity are in each window's replay receipt; detailed checks in `verification.json`; native cell series and package summaries are retained.