# Whole-domain seepage investigation

Measured by `scripts/run_domain_seepage_candidate.py`; this report is recomputed from its retained annual results by `scripts/report_domain_seepage_candidate.py`. This is a groundwater-only E0 candidate, not a production adoption or fully coupled flood model.

## Main conclusion

The border discontinuity cannot be accepted as a real geological boundary. The [source reconstruction](outlet_evidence/aquifer_regionalization.md) establishes that inherited footprint and crude condition-evidence geometry select aquifer properties as well as the previously diagnosed drainage differences. The new experiment removes the diffuse-drain selector across the domain; it deliberately retains aquifer properties so the drainage effect can be isolated. Correcting seepage alone does not qualify the remaining aquifer structure.

The annual adjustment test below determines whether the runs have settled; solver success does not. Soil-pressure feedback remains an implementation gap. The mapped outlet inventory supplies investigation leads, not surveyed drains or validated discharge capacities. The aquifer-regionalization defect is registered as `MD-032`.

## 1. Common seepage rule

All 15597 diffuse-boundary cells now use the same terrain-band formulation and scenario resistance. Country and the old model footprint do not select parameters. All 16689 active upper cells are covered by diffuse seepage or a retained explicit boundary. The 1092 existing line/northern-drain/river cells retain their distinct processes provisionally; the outlet audit does not newly qualify their hydraulics.

Full-domain terrain exposed 320 cells whose low terrain bands fall below the upper model layer. Initial runs were rejected by MF6; logs are retained under `terrain_geometry_rejection`. A country-independent geometry fallback uses model-average ground for those cells, flags them on the map, and leaves the aquifer geometry unchanged. This is a conservative numerical simplification requiring later local refinement; it is not a measured outlet elevation.

Measured border audit: 315 cross-country active neighboring pairs; 313 have different upper horizontal conductivity and 313 different vertical conductivity. Read `docs/records/model-domain-review-2026-09-23/lowland_mf6_candidate/run.py:224–243`: aquifer-zone assignment is also restricted to the old footprint, with default properties outside it. The separate [regionalization audit](outlet_evidence/aquifer_regionalization.md) traces the screening zones to their source geometry. Those inherited contrasts remain in this isolated seepage comparison; no unsupported geological homogenization was made. A remaining head seam therefore cannot be attributed solely to seepage or accepted as real geology. Whole-domain seepage consistency is implemented; whole-domain hydrogeological consistency is not yet established.

## 2. Adjustment to new outlets

Each cycle repeats identical daily 2017 forcings and carries both layer heads forward. Cycles are artificial conditioning repetitions, not consecutive historical years. Initial state is the original pre-2017 solution. The first control cycle uses the original packages and replays 2017; monthly heads must match the native record to <0.001 m. The control is then repeated under the same conditioning as the candidates, so later-cycle outlet effects can be separated from inherited model adjustment.

Adjustment requires both layers' p95 end-head changes <0.05 m, each layer's maximum <0.5 m, and annual diffuse discharge change <1%. A short run or normal solver termination is not adjustment. Current retained results:

| Arm | Completed cycles | Upper p95 annual head change (m) | Lower p95 annual head change (m) | Adjustment passed |
| --- | --- | --- | --- | --- |
| control | 6 | 2.746 | 2.707 | False |
| r250 | 6 | 2.618 | 2.608 | False |
| r1000 | 6 | 2.787 | 2.767 | False |
| r4000 | 6 | 3.015 | 3.006 | False |

0 candidate cycle records meet the full adjustment test. Consult `adjustment_history.csv` for maximum changes and discharge changes; `border_head_gradient_diagnostics.csv` reports cross-border head differences without treating smoothness as a physical-validity test; do not label a failing arm equilibrated. Budgets and independent DRN-formula comparisons are checked during extraction before raw outputs are discarded. Red head-exceedance cells can already discharge through drains and EVT; the colour does not mean zero drainage. Map annual seepage is external potential discharge, not routed flow or ponded-water depth. `matched_outlet_effects.csv` compares each candidate with the control at the same cycle and season; its annual volume difference repeats on each seasonal row and must not be summed across months.

The [baseline storage diagnosis](soil_feedback/baseline_storage_diagnosis.md) separates water stored in the aquifer from large head changes in dewatered upper-layer cells. Neither solver convergence nor small annual discharge change alone establishes stable water tables. Recomputed from `soil_feedback/supplementary_cycle_adjustment.csv`, the latest candidate cycles still have upper-head p95 changes of 2.52–2.84 m among cells whose upper heads are above the layer bottom at both endpoints, and whole-domain annual storage gains of 323–394 million m³. Endpoint status does not prove year-round saturation. Thus the remaining adjustment is not merely a dry-cell reporting artifact. The experiment stops at its declared initial cap and returns no equilibrated prediction; more conditioning and qualification of the inherited physical structure remain necessary.

The [EVT audit](outlet_evidence/evt_audit.md) checks native package coverage, depth-response rules, potential rates and executed control discharge. It finds no country-dependent coverage or extinction-depth switch. Potential demand varies through the inherited nearest-HRU residual-PET mapping; this audit does not qualify the complete land-atmosphere coupling.

## 3. Soil feedback

[Bounded soil diagnosis](soil_feedback/findings.md) and its `contact_summary.csv` identify recharge into cells whose previous head intersects the soil profile. Existing source states do not support a physically qualified pressure-dependent flux calculation. Saturated soil can still transmit downward water: withholding every exposed flux is NOT a defensible correction. An exact signed hypothetical withholding ledger retains water in an explicitly unallocated account; it is an accounting test, not physical storage or a coupled simulation. A reusable live soil-credit hook exists for upward supply, but does not calculate downward drainage pressure response. Step 3 remains a diagnosed implementation gap, with the concrete two-direction live-state route documented; no corrected soil-feedback run is claimed.

## 4. Outlet evidence

[Outlet audit](outlet_evidence/findings.md) inventories mapped drains, ditches, channels and existing terrain/winter leads. [Reconnaissance targets](outlet_evidence/observation_priorities.csv) include both countries, the border and Georgiev; markers are cell centres for investigation, not measured outlet locations. Inverts, capacity and downstream stages remain unqualified. Hand-drawn KGZ polygons are rough search hints only for this investigation. They do not select the new seepage parameters or serve as validation targets. However, the regionalization audit establishes that the inherited aquifer properties still depend on zones built from those polygons; that dependence remains an unresolved defect in these runs.

## Interpretation and next scientific boundary

Do not choose resistance by visual smoothness or adopt these runs as flooding predictions. The experiment removes the diffuse-drain history selector, tests adjustment under repeated forcing and exposes unresolved aquifer regionalization, outlet and live soil-feedback constraints. Production remains unchanged. No observations, field surveys, hydraulic connections or discharge capacities have been invented.

## Reproduce

Use `PYTHONPATH=diagnostics/coupled-improvement-2026-09-14/scratch/runtime .venv/bin/python scripts/run_domain_seepage_candidate.py build`, then `run --arm control` and each of `r250`, `r1000`, `r4000`. The default cap is six repeated annual cycles and stopping is conditional on all adjustment criteria. Use `--resume --cycles N` to extend a completed arm without replacing its earlier results. Run the outlet/soil scripts named in their findings and then `scripts/report_domain_seepage_candidate.py`. Existing result folders preserve each cycle; raw run directories are owned disposable copies with read-only source-file hardlinks. Source checksums are checked after the run.