The Rock DoesNot Lie

Physical-Financial Reconciliation and the audit of deferred stripping in open-pit mining

K53.5MAdjustments recovered · zero false positives

Conventional audit tests financial records against other financial records. Because the taxpayer authors every document in a stripping computation, that method is structurally blind. This paper makes the ore body the primary object of audit and reduces the financial claim to a testable hypothesis.

PaperMWP-2026-02
AuthorSujoy Maitra
JELH26 · H83 · Q38 · M42
EngineONYX-DS reference implementation
Pit cross-section — Kavarra Stage 3SURVEYED VS CLAIMED
THE EXCAVATED VOID — AS THE ROCK RECORDS IT FOUR CHANNELS PROBE THE SAME FLOOR — 35.3M BCM THE CLAIM ASSERTS ROCK THAT WAS NEVER MOVED
ORIGINAL SURFACE SURVEYED FLOOR CLAIMED FLOOR 3.2M BCM UNSUPPORTED
Indicators failed0 of 9
Risk score0.00
AdjustmentK0.0M
01

Why ledger auditing cannot work here

Deferred stripping — capitalising waste-removal costs under IFRIC 20 — is among the largest recurring discretionary judgements in extractive taxation, and among the least auditable. It rests on quantities that originate outside the financial system: strip ratios, component boundaries, bank cubic metres.

Because the taxpayer controls the generation of every document, a rational misstater produces a set that is internally consistent by construction. Conventional tests detect only errors of execution. Detection requires an anchor outside that control — and a modern mine supplies several: dispatch logs built for maintenance, surveys that certify contractor payments, metered diesel disclosed as scope-1 emissions, and a pit photographed from orbit every few days.

What IFRIC 20 requires

A stripping asset may be recognised only where three conditions hold together. Each is also a manipulation surface — the paper catalogues twelve schemes (F1–F12), of which four are operationalised here.

Condition 01 · ¶9

Probable future benefit

Improved access to ore must be probable. The boundary against ordinary production cost is where F7 operates — booking post-production cost as development stripping.

Condition 02 · ¶9

An identifiable component

Access must relate to a specific component. Define it as the entire final pit and amortisation stretches across decades; re-slice opportunistically and the base moves at will.

Condition 03 · ¶9

Reliably measurable cost

Costs must be measured, then split between inventory and asset on the excess of actual over expected strip ratio (¶13–14). Both pool and ratio are manipulable — F4 and F3.

Download the paper PDF
02

Headline result

On a physically consistent synthetic mine-year with four manipulation schemes injected at realistic intensities, the engine identifies every scheme with quantified, evidence-cited adjustments — while returning zero findings on an honest control derived from identical physics.

K53.5M
Fraudulent case

Seven of nine indicators fail, one warns. Five findings issue, each citing rule, uncertainty and evidence. Risk score 0.99.

K0.0M proposed on the honest control
zero findings, risk score 0.23

The specificity result matters as much as the sensitivity one. Administrations fear wrongful assessment; taxpayers purchasing assurance fear the opposite error. The engine leaves a clean position alone.

The mountain keeps better books than the CFO.

03

Four channels, one physics

PFR asks one question: does there exist an excavation history consistent with all admissible evidence under which the claim is true? Verification builds a maximum-likelihood reconstruction — the pit digital twin — with propagated uncertainty, then tests the claim against the supportable set it implies.

Each channel is generated by a different function of the firm, for a non-tax audience. Co-ordinated falsification of all of them is assumed infeasible; partial falsification shows up as cross-channel residuals, and those residuals are the detection signal.

01

Survey differencing

Monthly surfaces yield excavated solids with error-propagated confidence intervals. These surveys certify contractor payments — falsifying them means underpaying your own contractors.

02

Satellite DEM differencing

Public archives observe the pit directly and are not producible by the taxpayer at all — an independent volume check that also dates the excavation.

03

Haul-energy feasibility

Diesel consumed against the haul energy the claim requires. Fuel reconciles to supplier invoices and ESG disclosures, so the figure carries exposure elsewhere.

04

Block model tribunal

Conditional simulation over the taxpayer's own block model yields the distribution of supportable expected ratios — testing the assertion against the mine's own geology.

Non-production is itself evidence Uncertainty propagated in standard-error units Runs pinned for bit-identical re-execution
04

The money identity

Total disallowed deductions decompose exactly into three independently evidenced components. Each maps to one scheme and one finding, and the findings tie arithmetically to the recomputation — the property that survives objection and appeal.

Δ  =  I  +  Φ  +  s − κc)+
IIneligible content
F4 pool stuffing
ledger ontology
ΦFictitious volume
F5 phantom volume
survey / satellite / energy
κs−κcRatio suppression
F3 ratio inflation
block-model tribunal

Each term rests on a different channel, so a finding cannot be defeated by attacking one source. The split is exact by construction, and the acceptance-test suite asserts the tie-out.

05

Findings on the fraudulent case

One mine-year of the synthetic Kavarra pit: a 60×40×24 block model on 10 m benches, calibrated so the Stage 3 pushback's true life-of-component ratio is 4.9:1, with twelve months of excavation, monthly surveys, fleet cycles at 218 t payload and physics-derived diesel.

From that single truth, two submissions are derived — one honest, one with four schemes injected. Surveys, satellite, fuel and dispatch reflect true physics in both.

The honest control

Run identically, the honest submission passes all nine indicators. Supportable ratio recomputes to 4.8:1 (95% interval 4.6–4.9) against the 4.9:1 used — the 90th percentile of its own geology, aggressive but inside tolerance. Risk 0.23. Zero findings, K0.0M proposed.

FindingSchemeAdjustmentBasis
DS-2026-014F3 ratio inflationASSERTED 6.8:1 VS TRUE 4.9:1K17.3MRecomputed κs K59.3M vs K42.0M claimed
DS-2026-015F5 phantom volume38.5M BCM CLAIMED VS 35.3M MINEDK12.5M3.2M BCM unsupported, at pool unit cost
DS-2026-009F4 pool stuffingDEWATERING, CAMP, COMMUNITY COSTSK11.8MIneligible content, IFRIC 20 ¶12
DS-2026-011F7 boundary abuseBOOKED AT A SATELLITE PITK7.4MRamp-curve commercial-production test
DS-2026-017F8 consequentialK4.5MAmortisation base vs component reserves
TotalRisk score 0.99K53.5MFirst three tie exactly to the money identity

Table 1. Fixed seeds, fully reproducible. Engine wall time under one second per case, excluding the Monte Carlo tribunal (~15s at 1,500 realisations).

Volume
z = 6.2survey residual

Claim exceeds the rock

3.2M BCM above survey-differenced volume, corroborated by satellite DEM differencing and a 16.8 ML diesel shortfall against the haul energy claimed.

Ratio
>99.9th percentile

Outside its own geology

The asserted 6.8:1 ratio lies above the 99.9th percentile of the distribution built from the taxpayer's own block model.

Disclosure
104.5vs 113.0 Mt

Two ledgers disagree

Material moved per the sustainability report contradicts the tax computation; the component register's metadata post-dates the year it governs.

06

Evidentiary posture and limitations

01

Deterministic first. The machine-learning layer prioritises but never asserts. Every finding carries a rule citation, an uncertainty statement and an evidence reference.

02

Thresholds in standard-error units. Measurement-noise objections are answered in advance rather than litigated afterwards.

03

Adversarial cost multiplies. Coherent deception now requires co-ordinating surveys, fuel records, dispatch data, investor-facing reserve statements — and a satellite record that cannot be manipulated at all.

04

No systems integration required. No data leaves the administration; the jurisdiction rule pack is authored with, and retained by, counterpart staff.

Detection requires an anchor outside the taxpayer's control — or outside its practical capacity to co-ordinate.

Research agenda

01 — Retrospective field pilot with a partner administration 02 — Public benchmark corpus with per-scheme ROC 03 — Extension to underground and cost recovery 04 — Formal multi-channel deception game
Limitations. Results are simulation-based. The synthetic generator, however faithful to mining physics, cannot capture the messiness of real productions — datum errors, format archaeology, contested cut-offs — and the injected schemes, while drawn from practitioner experience, are stylised. Detection power against an adversary who partially co-ordinates channels, sensitivity to survey cadence and satellite revisit in cloud-affected geographies, and the behaviour of the non-production score under genuinely incomplete archives all require field study. Establishing field performance is explicitly future work.

Selected references: IFRIC 20 (2011) · Journel & Huijbregts (1978) · Chilès & Delfiner (2012) · Nigrini (2012) · IGF/OECD BEPS in Mining · TADAT Field Guide · Maitra (2026), MDRIA MWP-2026-01.

Availability. Platform design document, reference implementation with acceptance tests, interactive demonstrations, evidence-request register and deployment package available from the author. All results reproducible with fixed seeds. © 2026 Maitras.ai.