# Original baseline storage and head changes

Measured by `PYTHONPATH=diagnostics/coupled-improvement-2026-09-14/scratch/runtime .venv/bin/python scripts/diagnose_domain_seepage_soil_feedback.py --baseline-storage`. Native start/end head and daily STO-SS/STO-SY terms are extracted from the original G2 source; this is not a new simulation or a candidate result. Native and replay final heads agree exactly, and native storage sums equal the replay budget exactly (`baseline_storage_receipt.json`).

## Conclusion

The large control-year changes are already in the historical 2017 simulation. They cannot be attributed to the new seepage treatment. They are geographically broad, with the largest changes predominantly outside the earlier border-inspection rectangle, but meaningful changes also inside it. The rectangle is only a reporting area, not a boundary between lowlands and mountains.

Native net storage increases by 1.653709 billion m³ across both layers. Of that net increase, 74.78% occurs outside the rectangle and 97.10% occurs in diffuse-outlet cells. These are signed net shares, not shares of gross incoming water. Negative STO budget entries represent water entering storage; the script reverses that sign to report storage gain.

| Group | Upper cells | Median signed change, m | Upper p95 absolute change, m | Storage gain, both layers, million m³ |
| --- | ---: | ---: | ---: | ---: |
| all_active | 16689 | 1.69 | 19.49 | 1653.71 |
| diffuse | 15597 | 1.86 | 20.34 | 1605.76 |
| explicit | 1092 | 0.15 | 4.53 | 47.95 |
| inspection_window | 6598 | 0.66 | 8.09 | 417.10 |
| outside_inspection_window | 10091 | 2.95 | 25.09 | 1236.61 |
| upper_below_bottom_at_either_endpoint | 2904 | 12.63 | 36.77 | 82.59 |
| upper_above_bottom_at_both_endpoints | 13785 | 1.15 | 8.72 | 1571.12 |

The original inspection rectangle is x=340–450 km, y=4740–4800 km in EPSG:32643. Inside it, 652 upper cells rise more than 5 m over the year; the upper p95 change is 8.09 m. This is not solely a distant edge effect.

## Large head changes are not necessarily large water-table changes

Measured: 2904 upper cells have head at or below the upper-layer bottom at at least one endpoint. Their upper p95 head change is 36.77 m, versus 8.72 m for cells with upper head above the layer bottom at both endpoints. The endpoint test does not assert conditions throughout the intervening year.

The largest upper change is at row 146, column 224 (zero based), outside the rectangle: terrain 1545.35 m, layer bottom 1495.35 m, head 1266.35 → 1344.08 m. Its upper native storage change is 0.00 m³. Both heads lie below the modeled upper-layer bottom. The reported upper head there should not be interpreted as a water table moving through that upper layer, let alone as flooding. A raw maximum-head metric can therefore be dominated by dewatered-layer hydraulic states.

Nevertheless the cells with upper head above its bottom at both endpoints hold most of the net storage gain, so dismissing the whole signal as dry-cell numerics would also be wrong. Head-change and native-storage diagnostics should be reported together.

## Historical context and limits

`baseline_native_year_context.csv` extracts comparable endpoint changes from 2016–2019 in the original source. Their magnitude and sign vary between years. The 2017 replay is a historical comparison, not an equilibrium test; its annual change is not evidence that the new formulation failed to settle. Conversely, matching the historical run does not demonstrate physical plausibility of the original forcing, properties or initial conditions.

Candidate repeated-year conditioning should compare consecutive candidate cycles under the same repeated forcing. Keep the declared all-active-cell criterion, but diagnose failures separately for cells with head below layer bottoms and for actual native storage change. Do not silently relax the declared stopping criterion to hide either effect. A second conditioning control under repeated original forcing would distinguish baseline repeated-year drift from additional candidate effects; it has not been simulated by this diagnosis.

Files: `baseline_change_groups.csv`, `baseline_largest_changes.csv`, `baseline_native_year_context.csv`, and `baseline_storage_receipt.json`. Existing source model and engine inputs were unchanged.