# Groundwater storage and aquifer representation — owner review

## Main conclusion

Continued storage gain is not, by itself, evidence of missing surface drains. Repeated 2017 forcing imposes **2.54 times** the actual historical mean recharge. The initial steady recharge is verified against the same native calendar and is not stale. The source model is being conditioned toward an unusually high prescribed-recharge regime. Whether that recharge is physically right remains a separate question.

The accumulation geography also matters: deep-water-table areas keep filling while near-surface head changes become much smaller. A single whole-domain adjustment verdict conceals that distinction. The remaining political-border pattern still cannot be treated as geological evidence: inherited conductivity, thickness and storage assignments remain structurally uncertain.

Measured by `scripts/report_seepage_storage_review.py` and this writer from retained native budget/head outputs and the audited recharge calendar. Source evidence and unchanged-input receipts are linked below. This investigation uses the original G2 `combined_feedback2_dec206_precise` generation underlying the earlier maps; conclusions do not automatically transfer to newer coupling experiments. No production model or geological replacement is adopted.

## 1. Longer runs and where the water goes

Each cycle repeats the same daily 2017 forcing, starting from the previous end state. These are not later historical years. Original cycles 1–6 remain in the earlier record; the continuation stores annual cell/layer package balances and signed horizontal/vertical exchange. Positive storage means more groundwater held, not deeper surface flooding.

| Case | Last cycle | Annual storage gain, million m³ | Whole-domain upper-head p95 change, m | Near-surface upper-head p95 change, m | Worst-month near-surface p95 change, m | Positive storage in cells with no drain discharge |
|---|---:|---:|---:|---:|---:|---:|
| control | 18 | 51.1 | 0.385 | 0.011 | 0.014 | 97.9% |
| r1000 | 18 | 50.9 | 0.383 | 0.011 | 0.015 | 95.6% |

Here, p95 is the change below which 95% of the selected cells fall; a small p95 does not rule out a few larger local changes. R1000 denotes the common-seepage case with a 1,000-day resistance to water leaving through the ground surface.

For R1000, annual storage gain is 85.8% lower than at cycle 6. The largest current high-accumulation cluster has a median upper-head depth of 73.5 m and gains 30.9 million m³ in the last cycle. Its full signed account is in `accumulation_clusters.csv`; the map popup shows recharge, mountain supply, pumping, net lateral/river exchange and outlet losses. Deep cells do not start discharging through a surface drain simply because more such drains are assigned.

The fixed near-surface reference cohort consists of cells with an upper head within 5 m of mean ground at any retained month-end in cycle 6 of either unchanged control or R1000. It is a model-derived diagnostic cohort, not an observed flooding boundary. Five metres is an assumed reporting threshold, not a hydraulic parameter. Worst-month statistics compare matching months in consecutive repeated cycles. `head_adjustment_groups.csv` also retains both-layer statistics for countries, inherited zones, the original inspection window and Georgiev. Group water accounts overlap across grouping families and must not be added together.

The original whole-domain stopping criteria remain unchanged. A declining storage trend or a locally small head change does not override a failing global test. Neither establishes observational validity. The continuation is a bounded high-recharge stress test; representative conditioning should use the actual historical forcing sequence or a justified representative climate before a historical prediction is assessed.

### Initial state and boundary evidence

[Boundary audit](boundary_audit/findings.md): the native graph is connected. The outside/inactive faces have no lateral exchange; sources and sinks still act through wells, rivers, drains, recharge and EVT. This is an assumption to compare with hydrogeology, not proof that an arbitrary open boundary should be added. The lower layer remains mathematically confined even where calculated heads fall below its base, which needs physical qualification before interpreting saturated thickness or storage there.

The true calendar mean recharge recomputed from `boundary_audit/native_recharge_calendar.csv` is 4.900 million m³/day; 2017 is 12.457 million m³/day. `native_recharge_calendar_receipt.json` independently compares the actual steady field against every transient native recharge field. The stale-initial-recharge hypothesis is ruled out for this generation; the validity of the prescribed recharge itself is not established by that identity.

## 2. Aquifer representation and the border

[Independent evidence inventory and proposed alternatives](aquifer_evidence/findings.md) distinguishes actual observations from source availability. BGS well tables and cross-sections provide local stratigraphic constraints, but local coordinates, section-projected positions and datum must be resolved before contacts enter the model grid. Kazakhstan intake coordinates provide retrieval anchors, not a continuous aquitard surface or hydraulic-test values. The KGZ geological map does not supply equivalent KAZ coverage.

The supported conceptual direction is a connected alluvial system with varying proportions of coarse material and low-permeability interbeds. Apply that interpretation across the domain; allow parameters to vary where physical evidence supports variation. The political border and crude condition polygons must not choose material contacts.

[Uniform-conductivity comparison](property_null/findings.md) reports the isolated sensitivity. Thickness, storage and lower-layer confinement remain inherited; this is not a homogeneous aquifer or an adopted geological replacement. At matched cycle 12, the KGZ part of the original inspection window has a median head difference of -0.16 m in the upper layer and -4.12 m in the lower layer relative to unchanged conductivity. Effects vary in direction and size across the domain. This demonstrates sensitivity to the inherited conductivity recipe, especially in deeper heads; it does not identify conductivity as the sole cause of the visible border contrast. Both comparison cases are still adjusting. Measured by the paired comparison script from matched saved heads; summary recomputed here from its CSV.

## Review decisions proposed

- Keep the natural-seepage formulation as the shared candidate; do not tune it to hide the inherited aquifer-property seam.
- Treat the repeated-2017 experiment as an unusually high-recharge sensitivity test. For subsequent historical assessment, select and justify conditioning with the actual forcing calendar.
- Use the connected-alluvial/interbed concept as a hypothesis for spatial qualification. Owner constraint, 2026-09-28: no further aquifer-geometry information will become available. Use the Morris/Alameddin evidence only as a local constraint. Develop a deliberately simple domain-wide representation and test plausible parameter ranges; additional boreholes or mapped contacts are not prerequisites. Do not adopt a fabricated whole-domain aquitard surface.
- Review lower-layer confinement and outer groundwater-edge assumptions through physical consistency checks and bounded sensitivity tests using the available information. No arbitrary boundary head or uniform-property aquifer is proposed for adoption here.

Soil-feedback implementation and any production adoption await the next decision; they were not part of the authorized steps 1 + 2.

## Reproduce and inspect

Continue with `scripts/continue_seepage_storage_review.py --arm control` or `--arm r1000` using the FloPy runtime named in `contract.md`; run `scripts/report_seepage_storage_review.py` and `scripts/write_storage_aquifer_review.py` after completed cycles. The evidence and boundary subfolders contain their executed probes and receipts. The interactive map is `index.html`; intermediate rendered HTML is not committed. Numerical checks and source preservation are distinct from physical acceptance.