Farmers and local water agencies often report how much groundwater they pump, and those reports are hard to check. Across the North China Plain, extraction is usually inferred from agricultural electricity bills or satellite evapotranspiration rather than measured at the well. Yet the aquifer keeps its own record. Pumping at one well lowers the water level at nearby observation wells by an amount that physics largely fixes. This proposal builds an auditor on that fact. Every pumping report, electricity reading and water-level reading is sealed with a timestamp when it is captured and signed by whichever party produced it. The auditor then asks whether the sealed pumping reports are physically consistent with the sealed water levels. The same analysis says exactly which wells or groups of wells the monitoring network can actually tell apart when assigning blame.
In plain terms, the claim is this. An immutable audit trail, meaning an append-only log in which altering any past entry is detectable, only guarantees that records were not changed after the fact. It cannot guarantee they were true when written. The aquifer supplies the independent truth check that the log lacks. The log supplies what the aquifer check lacks: protection against a database insider who, once a discrepancy appears, edits the water levels or the pumping figures until they agree. Neither piece is sufficient alone. We test whether the combination detects and localizes under-reporting that simpler monitoring misses.
Core Claim
Our reading of the literature is that the work so far sits on either side of this idea without touching it.
On the hydrology side, Real-Time Management of Groundwater Resources Based on Wireless Sensors Networks links telemetry from 42 wells to a MODFLOW model. MODFLOW is the standard finite-difference simulator of groundwater flow. That paper uses a response matrix, a linearized map from each well's pumping rate to drawdown at each observation point, to optimize pumping yields. It assumes reported pumping is honest and keeps everything in an ordinary central database.
On the audit side, AUDITA seals signed messages into Merkle-committed ledgers. A Merkle commitment is a single hash that fixes a whole batch of records, so any later edit is exposed. AUDITA then computes graded causal blame. Its results give us two motivating facts. First, unsigned logs framed innocent parties in 100% of forgery probes. Second, its Theorem 4 states that an auditor restricted to reading records cannot reach full correctness unless it can independently compute the ground truth of the domain. AUDITA reasons only over discrete events. Fine-Grained Traceability for Transparent ML Pipelines anchors Merkle roots of machine-learning training events and applies nothing to physical sensors.
We take the completeness barrier literally. For groundwater, the domain's ground-truth capacity is the response matrix. The auditor's power is therefore bounded by how well the observation network can resolve pumping, and that bound can be computed before any audit runs.
The central relation is simple. Let $h$ be the vector of sealed observed heads (water levels) at monitoring wells, $h_0$ the heads expected without pumping, $A$ the response matrix, $q$ the vector of reported pumping rates, and $\epsilon$ observation and model noise with covariance $\Sigma$. The consistency residual is
$$ r = h - h_0 - A q . $$
If the wells under-report their pumping by an amount $\delta$, the residual's mean shifts by $A\delta$. The violation is detectable roughly when the Mahalanobis size of that shift clears the noise:
$$ \delta^\top A^\top \Sigma^{-1} A \delta \gtrsim \chi^2_{\text{crit}} . $$
This is the standard power condition for a Gaussian mean-shift test. Two consequences follow directly from the matrix $A^\top \Sigma^{-1} A$. First, its directions of near-zero curvature are violations the network cannot see. Second, clusters of wells whose columns of $A$ are nearly parallel cannot be told apart. We call such clusters "blame cells." The auditor issues verdicts at the resolution of blame cells, never at the resolution of individual wells it cannot distinguish.
Blame cells are the specific answer to a framing threat. A dishonest reporter could shift reported pumping onto a neighbor in a direction the network cannot see. Refusing to issue verdicts below blame-cell resolution means that shift cannot pin guilt on the neighbor.
Research Program
Edge sealing. Each data source signs and hash-chains its own records at capture. Pressure transducers are signed by the monitoring agency. Electricity meters are signed by the utility. Pump self-reports are signed by the concession holder. Hourly batches are summarized as Merkle roots and anchored with an external timestamping service. A public ledger is optional; the property that matters is that records from mutually independent signers are fixed before anyone sees the residual.
The volumes are small. The CERN groundwater dataset A dataset of groundwater level in Chinese Ecosystem Research Network (2005 – 2014) records hourly transducer data aggregated to daily means. We therefore expect storage overhead far below the roughly 1.2 GB that FG-Trac reported for a 100-epoch training ledger, because FG-Trac also stored gradient information.
Consistency engine. The engine takes the sealed records and computes the residual with $A$ derived from a calibrated MODFLOW model. It runs a generalized likelihood-ratio test for localized mean shifts, one per candidate blame cell. It also computes the blame-cell partition and a map of minimum detectable under-reporting for each well.
Blame is kept simple on purpose. In the confined-aquifer regime, drawdown superposes linearly, so apportioning drawdown among wells is simply the column-weighted share. The hard problem is detection and identifiability, not attribution. Heavier machinery from continuous causal attribution, such as Causal Attribution Analysis for Continuous Outcomes, is not needed in this regime. That paper handles continuous outcomes but only binary causes, and it assumes trusted data. It becomes relevant only where nonlinearity breaks superposition, which we treat as a stress test.
Network design. Given the blame-cell map, we greedily choose where to add a small number of piezometers so as to split the largest cells or raise the smallest detectable violation. The residual scores also feed a budget-limited inspection schedule for physical meter checks.
Baselines
The do-nothing reference accepts sealed self-reports at face value and flags nothing. It fixes the floor for detection and sets the false-accusation rate to zero by construction.
The electricity-proxy check is the strongest simple competitor because it is what regulators in Cangzhou already have. It converts utility electricity logs to pumping volumes, using the proxy of Predicting Groundwater Level Dynamics… Using Stacking Ensemble Learning, and flags wells whose self-reports fall below the proxy by a tuned margin.
The learned-forecaster residual is the strongest existing modeling method. It trains the stacking ensemble from that paper and flags wells whose observed heads fall below forecast. On Cangzhou's third confined aquifer, that paper reports a test MAE of 3.25 m, RMSE of 4.20 m and $R^2$ of 0.88, which is 16.2% lower MAE than its best single learner, SVR. That is a forecasting figure. We found no reported result on Lacuna for detecting pumping misreports with this forecaster or any other method, so the detection comparison is new.
All methods see the same sealed records, the same training window (2018–2021) and test year (2022) in real-data analyses, and the same false-accusation budget. Thresholds are tuned on a held-out validation split with equal search effort. Beating a baseline means higher localization accuracy at a matched 5% false-accusation rate, and a smaller minimum detectable under-reporting volume.
Two ablations isolate the ingredients. The unsealed ablation runs the identical consistency engine on an editable database, against an insider who may alter records after seeing residuals. It isolates what sealing contributes. The no-blame-cell ablation forces per-well verdicts. It measures how often naive localization frames a neighbor.
For the inspection schedule, the reference is uniform random inspection at equal budget. The Stackelberg allocation of Audit Games serves as a structured competitor that treats wells as independent. The online signaling policy of To Warn or Not to Warn reported up to 77.31% utility gain over Stackelberg baselines on medical access logs. That gain is not transferable as a target here, and we cite it only as evidence that budget-aware scheduling matters.
Why not just…?
Why not just use a well-secured central database? For the honest-but-noisy case, it would do. The consistency engine does not need cryptography, and we say so openly. The seal matters only against the party who controls the database and can edit water levels after a discrepancy appears. AUDITA shows that unsigned logs are trivially forged under such an actor. The unsealed ablation measures this directly. If insider edits cannot restore consistency without leaving detectable traces anyway, the sealing layer should be dropped from the recommendation. The mechanism we need is independent signing plus timestamping, which does not require a blockchain.
Why not just flag wells whose heads fall below a learned forecast? This is the real threat. The forecaster is accurate at regional scale and needs no MODFLOW model. However, its 3.25 m error is several times larger than the local drawdown a single under-reporting well produces. It has no notion of which well caused a deficit, so it cannot localize or protect neighbors from blame. It is also trained on the very reports being audited. The head-to-head detection experiment settles this. If the forecaster matches the physics-based test on localization, the contribution collapses to the sealing layer alone.
Why not just rely on GRACE or InSAR? Satellite gravimetry resolves storage change over basins of the size tracked in Unprecedented large-scale aquifer recovery through human intervention, far coarser than a single depression cone. InSAR subsidence (High-Resolution Monitoring… Linfen-Yuncheng Basin) records compaction that has already happened. Both are useful as a regional prior on $h_0$, and neither can assign a specific irrigation season's deficit to a specific concession.
Why not just inspect meters at random? At more than 52,400 wells in Cangzhou alone, random inspection catches little per unit of budget. The experiment asks how much residual-guided inspection raises the violation catch rate at equal budget.
Proposed Experiments
Controlled experiment (Heihe). We rebuild the Daman-district MODFLOW setup from the Heihe study as the ground truth, pumping at magnitudes around its reported optimized yields (for example, about 0.57 × 10⁸ m³/a under its loosest constraint S1). We inject under-reporting into 1–10% of wells. The attackers are of four types: a constant fractional under-reporter, a seasonal smoother, a null-space shifter who knows $A$ and moves reported pumping onto neighbors, and a database insider who edits heads after seeing residuals.
Model mismatch. The auditor uses the linearized response matrix while the truth comes from full nonlinear MODFLOW. This tests robustness in the unconfined regime where superposition fails.
Metrics. Detection AUC, blame-cell localization accuracy, false-accusation rate, and minimum detectable under-reporting volume. We check the measured detection threshold against the $\chi^2$ prediction above.
Real-data check (Cangzhou). We take the 131 monitoring wells in the confined layer and use the electricity-derived pumping as the "report." We then ask whether the consistency engine finds spatially coherent residual clusters and whether their timing lines up with the diversion-driven recovery documented there.
Transfer check (North China Plain). On the more than 2,000 wells of the North China Plain inventory, we compute blame-cell maps to show how identifiability differs between urban recovery zones and the rural districts that are still lagging.
Decisive checks. The proposal succeeds if three things hold. First, the physics-based test localizes violations the stacking-forecaster residual misses at matched false-accusation rate. Second, the unsealed ablation collapses under the insider attacker while the sealed version does not. Third, enforcing blame-cell resolution reduces framing of innocent neighbors to near zero under the null-space attacker. Failure on the first check means the aquifer is not a strong enough witness at realistic network density. That is itself a useful, publishable bound on what pumping audits can promise.
References
- Real-Time Management of Groundwater Resources Based on Wireless Sensors Networks
- AUDITA: certified auditing and causal attribution of adverse outcomes in autonomous multi-agent systems
- Fine-Grained Traceability for Transparent ML Pipelines
- A dataset of groundwater level in Chinese Ecosystem Research Network (2005 – 2014)
- Causal Attribution Analysis for Continuous Outcomes
- Predicting Groundwater Level Dynamics and Evaluating the Impact of the South-to-North Water Diversion Project Using Stacking Ensemble Learning
- Audit Games
- To Warn or Not to Warn: Online Signaling in Audit Games
- Unprecedented large-scale aquifer recovery through human intervention
- High-Resolution Monitoring and Driving Factor Analysis of Long-Term Surface Deformation in the Linfen-Yuncheng Basin