Abstract
Mali's 2023 Mining Code raised the state's potential stake in projects to 30% (10% free-carried plus a 20% purchase option), added a 5% local-investor stake, raised royalties and the corporate rate, shortened licences to 10 years and removed exploitation-phase exemptions. A 2023 contract audit was followed by settlements with Resolute, Allied Gold, Hummingbird, B2Gold and Barrick; the Finance Minister reports CFA 761 billion recovered against a target of 400 billion, and mining payments to the Treasury rose to about US$1.4 billion in 2024 even as output fell. The Barrick dispute over Loulo-Gounkoto, settled in November 2025 with a US$253 million arrears payment and a 10-year licence renewal in February 2026, cost a year of production at the country's largest mine. National output fell by roughly a quarter. This paper models the trade-off between fiscal take and investment: a fiscal-regime simulator across price paths, an exploration-response model, a settlement-valuation method for arrears claims, and an indirect-method audit for cost and production. The illustrative result is that the 2023 terms lift the government's present-value share of a representative mine from about 46% to 63%, and that the marginal exploration project's return falls below the hurdle at prices under about US$3,000 an ounce. The code is sustainable at today's prices and not at yesterday's; the paper's contribution is a rule for adjusting it before that matters.
Introduction
Mali did what many resource-rich states talk about: it audited its mining contracts, rewrote the code, and collected. The results are real. Treasury receipts from mining rose by more than half in 2024 despite lower output, and back-tax settlements exceeded target. The costs are also real: production down about 23%, Loulo-Gounkoto idle for much of 2025, exploration spending by juniors sharply lower, and a sovereign cut off from concessional finance since 2020 leaning harder on a single commodity.
The research question is whether a fiscal regime this tight can be kept without eroding the base it taxes, and what instruments would let Mali capture rent at high prices while keeping the marginal project alive at low ones. The paper is written for the Ministry of Economy and Finance and the Ministry of Mines, and for the state-owned entities that now hold minority stakes and will need to audit their own partners.
The organising identity:
R^{pot} - R = \underbrace{\left(R^{pot} - R^{law}\right)}_{\text{policy gap: rate structure, stabilised old contracts}} + \underbrace{\left(R^{law} - R\right)}_{\text{compliance gap: cost inflation, transfer pricing, artisanal exports}}Current tax issues
2.1 Take versus investment
Mali's fiscal terms for gold were, until 2023, among the more generous in West Africa. The 2019 code carried forward a structure built in the 1990s to attract capital to a landlocked country with no infrastructure: a royalty of 3% on the value of production plus a 3% special tax, a corporate rate of 25 to 30% with a five-year holiday in some conventions, a state free carry of 10% with the option to buy a further 10%, and exemptions from most indirect taxes during the exploitation phase. Those terms did their job. Sadiola, Morila, Loulo, Gounkoto, Fekola and Syama were financed and built by foreign capital, and by 2020 gold was more than three-quarters of Mali's exports.
The 2023 code changed every one of those parameters at once. The state's stake rose to a mandatory 10% free carry plus an option over a further 20%, with 5% more reserved for Malian private investors, so that total domestic ownership can reach 35%. Royalty rates rose, the corporate rate rose to 35% for the sector, exploitation-phase exemptions were removed, licences were shortened from 30 years to 10, a contribution of 0.75% of quarterly revenue to a local development fund was imposed, and Malian nationals must hold senior positions. The government's stated goal is to raise the mining sector's contribution to GDP toward 20%.
The consequence for investment follows from arithmetic rather than ideology. Combining the 2023 terms with Mali's security risk, its exclusion since 2020 from regional debt markets and concessional lenders, and the operational risk that the Barrick dispute made visible, the effective hurdle rate for new capital in Mali is now among the highest in West Africa. Existing mines continue to operate, because their capital is sunk and their cash margins at US$4,000 gold are wide. What has stopped is the marginal decision: expansions are deferred, juniors have cut exploration, and the licence relinquishment rate has risen. Gold mines take eight to twelve years from discovery to first pour. The tax base of 2032 to 2036 is being decided by exploration budgets set in 2025 and 2026, and those budgets are being cut. The paper's first model quantifies that trade-off; its third model proposes a rule that keeps the revenue at high prices while restoring the marginal decision at low ones.
2.2 Transition of stabilised contracts
Most of Mali's producing mines operate under conventions signed before 2023 that contain stability clauses: undertakings that the fiscal terms in force at signature will apply for a fixed period, usually the life of the initial licence, regardless of later changes in the law. The 2023 code asserts that it applies to all operators, and the Finance Minister has said all companies will now fall under it. The two positions cannot both be right, and the settlements of 2024 and 2025 resolved the tension case by case rather than by rule. Resolute, Allied Gold, Hummingbird, B2Gold and Barrick each reached agreements whose terms are only partly disclosed.
The Barrick case illustrates the cost of ambiguity. Loulo-Gounkoto's licence was renewed by decree in February 2026 for ten years, following the November 2025 settlement, yet neither party has said whether the renewed licence carries the 2023 ownership and fiscal terms or the earlier framework. That single unanswered question determines whether the state's stake in the country's largest mine is 20% or up to 30%, what royalty applies, and whether exploitation-phase exemptions continue. It also determines how much of the Finance Minister's projected US$1 billion a year of additional revenue is real.
For revenue administration, ambiguity of this kind is itself a loss. It invites litigation over every assessment; it delays dividends on the state's own shares because the dividend base cannot be agreed; it makes multi-year revenue forecasting impossible, which weakens the budget process and the case for concessional finance; and it gives every operator an incentive to hold back investment until the question is settled. The register proposed in section 5 and specified in Part II, listing for every licence the regime in force and its expiry, is the cheapest reform in this paper and the one on which all the others depend.
2.3 Arrears claims and settlement valuation
The 2023 audit of mining contracts produced claims against most operators for back taxes, unpaid dividends and disputed exemptions. The Finance Minister reports CFA 761 billion recovered against a target of CFA 400 billion. The largest single claim, about CFA 300 billion or US$512 million against Barrick, was settled for a payment of US$253 million together with Barrick's withdrawal of its ICSID arbitration and the state's release of detained employees and legal proceedings. The other settlements are undisclosed.
A claim is a bundle of items with different legal strength: a royalty-base dispute where the law is clear may be worth nearly its face value; a transfer-pricing claim built on a benchmark the operator contests may be worth a third; a claim resting on a reading of a stability clause that an arbitral tribunal has rejected elsewhere may be worth little. Without a published method for valuing each element, a settlement at 49% of face value can be presented as a capitulation by critics at home and as extortion by investors abroad, and the state has no way to show it was either well judged or well negotiated.
The method matters for the next negotiation as much as the last. Operators now know that Mali will use detention of staff, licence threats and export blockages as leverage, and they price that into their hurdle rates. A published valuation method, applied by a unit that is not the negotiator, would let Mali make the same recoveries with less collateral damage: the operator would know what the state considers the claim worth and why, the state could show that a settlement above expected value was a gain, and the production loss that the Loulo dispute caused, a year of lost royalty, dividends and supplier income at the country's largest mine, would be counted as a cost of the strategy rather than ignored. Section 3.4 sets out the method and Part II works an example.
2.4 Cost inflation and transfer pricing
The 2023 code shifts the composition of the state's take toward profit-based instruments: a higher corporate rate and a larger equity share, alongside the royalty. Profit-based instruments are the right choice economically because they respond to margin, but they change the operator's incentive. Under a royalty, the operator's tax bill depends only on revenue, which the state can observe from refinery receipts and prices. Under corporate tax and dividends, the bill depends on costs, which the operator incurs, procures, allocates and reports. Every dollar of cost that can be shifted into the Malian entity reduces the state's take by 35 cents of tax plus 30% of the after-tax residual.
The channels are well known from mining audits elsewhere. Related-party procurement: fuel, reagents, tyres, spare parts and mining contractors supplied through a group company at a mark-up. Management, technical and marketing fees charged from headquarters. Interest on shareholder loans at rates above the group's cost of capital. Allocation of shared costs between mines in different countries so that the higher-tax jurisdiction bears more. Hedging arrangements or sales to related refiners that realise a price below the daily fix. Accelerated or inflated capital cost, which increases depreciation and the cost-recovery base for the state's equity.
Mali's Direction Générale des Impôts has limited independent data with which to test any of these. It does not hold arm's-length benchmarks for fuel and reagent prices, a regional database of capital cost per unit of capacity, or a beneficial-ownership map of mine suppliers, and it has few staff with mining engineering or transfer-pricing training. The 2023 audit was carried out largely by external consultants. The models in section 3.5 and the audit engine in section 5 are designed to give the DGI a standing capacity to do what the consultants did once, and to do it every year on every mine.
2.5 Artisanal gold and the northern trade
Estimates of Mali's artisanal and small-scale gold output range widely, from twenty to more than fifty tonnes a year, because the sector is unrecorded by construction. Production is concentrated in the Kayes and Sikasso regions in the south-west and, increasingly, in the north and centre where the state's control is contested. Output moves through traders in Bamako to Dubai, largely by air, either declared at low values or not declared at all. The UAE's recorded imports of gold from Mali have for years exceeded Mali's recorded exports to the UAE by a wide margin.
The fiscal loss is only part of the problem. Gold mined in areas held by armed groups finances those groups, through taxes they levy at the pit, protection payments on transport, and direct control of some sites. Any counter that pays well for gold without asking where it came from is, in those areas, a conflict-finance channel. Any rule that excludes gold from those areas pushes it entirely into the informal channel. The design problem is therefore harder than in Ghana or Burkina Faso: the state needs a buying channel that draws artisanal gold into the formal economy in the areas it controls, with provenance checks strict enough to exclude gold that finances its enemies and loose enough not to be evaded wholesale.
The companion papers on Ghana and Burkina Faso set out the channel-choice model and its estimation. Burkina Faso's experience is the most relevant to Mali: SONASP's buying counters lifted recorded artisanal output from 8 to about 42 tonnes in one year, in a country with comparable security conditions, by paying promptly at close to world prices. Mali has no equivalent institution. The paper's recommendation, in section 6, is a licensed buying-counter network with a low levy, provenance rules tied to a register of artisanal sites in state-controlled areas, and mirror-data cooperation with the UAE to measure what still escapes.
2.6 State equity without audit capacity
The 2023 code's centrepiece is equity: a 10% free carry, an option over a further 20%, and 5% for Malian investors. Equity looks like the strongest form of participation. In practice a minority shareholder receives dividends only after the operator, as majority shareholder and manager, has decided how much profit there is, how much capital to reinvest, how much debt to service and how much to distribute. Each of those decisions is legitimate management discretion, and each can be used to defer or reduce the minority's cash indefinitely.
The state therefore needs three things it does not yet have. First, audit rights over the operating company's costs, exercised by a technical unit that can read a mine plan and a procurement contract. Second, a dividend policy written into the shareholders' agreement, so that distributable profit is defined by formula rather than by board resolution. Third, directors on the mine's board who are briefed by that technical unit and who understand the cash-flow waterfall. Without them the free carry is a right to a share of whatever the operator declares, which is worth much less than 30% of the mine.
The purchase option compounds the problem. Buying a further 20% requires the state to value the mine, which requires exactly the cost and production verification it lacks, and to finance the purchase at a sovereign cost of capital that is high while Mali is outside concessional markets. Exercising the option on the operator's numbers, at the operator's valuation, financed at 12% or more, could leave the state paying more for its equity than the equity returns. Section 3.1 treats the option explicitly and Part II shows how to value it; section 5 proposes the technical unit that would make the state's shareholdings worth their face value.
Models
3.1 Model A: fiscal regime simulator
The simulator answers one question for any fiscal regime: across the range of gold prices the country might face, how much does the state receive, and does the investor still earn enough to build the mine? Everything else in the paper depends on that answer, so the model is set out in full.
A mine is described by its output profile q-t in ounces, its all-in sustaining cost c-t per ounce, its initial capital K-t spread over the construction period, and its life T. These come from the operator's published technical report, which lists reserves, grade, recovery, throughput and cost by year. The regime is described by a royalty function r(.) that may depend on price and on margin, a corporate rate t, a depreciation rule that converts capital into annual deductions D-t, a free-carried state share e-f that costs the state nothing, a purchased share e-p bought at price B in year t-B, and any additional instrument such as a rent tax. For a representative mine with price path p, output q, all-in cost c and capital K, under royalty r, corporate rate t, free-carried state share e-f and purchased share e-p bought at price B:
G_t = r\,p_t q_t + t\,\Pi_t + \left(e_f + e_p\right)(1-t)\,\Pi_t - \mathbb{1}[t = t_B]\,B, \qquad \Pi_t = p_t q_t (1-r) - c_t q_t - K_tGovernment take share is the present value of G over the present value of pre-tax cash flow, discounted at a common rate so that timing differences between a royalty (paid from the first ounce) and a dividend (paid after capital recovery) are counted. The investor's return is the internal rate of return on the cash flows that remain after every government item, including the dilution of its share by the state's equity. The simulator runs the cash-flow identity across a distribution of price paths rather than a single price, and reports take, investor IRR and the probability that the project would have been sanctioned, meaning the share of price paths on which the investor's IRR exceeds its hurdle rate.
Three features of the model deserve comment. First, royalty and free carry do not respond to margin: at any price they take the same share of revenue and of after-tax profit respectively. That is why the 2023 regime's take share rises as prices fall, which section 4.2 shows. Second, the corporate tax and the purchased equity do respond to margin, but only on the cost base the operator declares, which is why the audit models in section 3.5 are not optional extras. Third, the purchased share is a real option: the state can choose whether and when to exercise it, and its value depends on the purchase-price formula in the code, on the timing window, and on the state's cost of finance. Treating it as a fixed 30% stake, as public discussion usually does, overstates the state's position when finance is expensive and understates it when the option can be exercised cheaply after a mine has proved itself.
3.2 Model B: exploration response
The fiscal simulator values the mines that exist. The exploration model values the mines that will or will not exist in ten years, which is where a high fixed take does its damage. Exploration is discretionary spending by companies choosing among countries. The decision depends on the expected value of a discovery to the investor, which is the geological prospect times the price outlook times the share of project value the investor keeps after the state's take, less the cost and risk of operating in the country. Raise the take, and the investor's share of any discovery falls, and so does the exploration budget allocated to Mali relative to Senegal, Guinea or Côte d'Ivoire.
The model links three things: exploration spending X, the reserves that spending adds, and the production those reserves yield after a lag. Future production depends on exploration spending X, which responds to the expected post-tax value of discovery. With elasticity epsilon of exploration to the investor's share of value, and a discovery function converting spending to reserves:
X_t = X_0 \left(\frac{(1 - g_t)\, V_t}{(1 - g_0)\, V_0}\right)^{\varepsilon}, \qquad \Delta \text{Reserves}_t = \phi\, X_t^{\,\theta}, \qquad Q_{t+L} = \psi\, \Delta\text{Reserves}_tg is government take, V project value, L the lag from discovery to production (8 to 12 years for gold), phi and theta the parameters of the discovery function, and psi the conversion of reserves into annual production. The exponent theta is below 1 because exploration has diminishing returns: the best prospects are drilled first. The elasticity epsilon is the model's most important and least known parameter. International studies of mining exploration put it between about 0.6 and 1.8 with respect to the investor's share of value; section 8.3 explains how to estimate it on Mali's own licence data, separating the fiscal effect from the security effect, which in Mali is large and independent.
The model's output is a long-run Laffer curve for the code. In the short run, raising take from 46% to 63% raises revenue by roughly the ratio of the two, because output is fixed. In the long run it lowers exploration, which lowers reserves, which lowers production in the late 2030s, and the revenue at that time is 63% of a smaller base. Whether the long-run revenue is higher or lower than under the old regime depends on epsilon, theta and the lag; Part II works the numbers for a range of values. The policy use of the model is not to find the exact optimum, which the data cannot support, but to show the direction and rough size of the effect and to put a number on what a lower take in low-price years buys in future production.
3.3 Model C: adaptive take rule
The pattern in Mali, and across the region, is that fiscal terms are tightened in booms, when high prices make every regime look too generous, and litigated in busts, when the tightened terms make marginal mines uneconomic and operators demand relief. Each cycle costs a year or two of investment, a round of arbitration, and a reputation for instability that raises the hurdle rate for the next cycle. The adaptive rule replaces that political cycle with a formula.
The idea is that the state's marginal share should depend on how profitable the mine actually is, measured as margin, the share of revenue left after cost. When margins are thin the state takes a low share, which keeps the mine open and leaves the operator enough to keep exploring; when margins are wide the state takes a high share, capturing the rent that high prices create. The instrument that implements this is a resource rent tax, a margin-linked royalty band, or a combination. Rather than a fixed code that is renegotiated by crisis, define the state's marginal share as a function of the realised margin:
g(m) = g_{min} + \left(g_{max} - g_{min}\right)\cdot \frac{1}{1 + e^{-\kappa\,(m - m^{*})}}, \qquad m = \frac{p - c}{p}g-min is the take at thin margins, set by the marginal project's hurdle rate; g-max is the take at wide margins, set by comparison with the highest-take regimes that still attract capital, such as Tanzania's 50/50 split; m-star is the margin at which the take is halfway between the two; and kappa controls how quickly the take moves between them. At thin margins the state takes g-min, which keeps the mine open and the exploration budget alive; at wide margins it takes g-max. The band is a smooth function, not a step, so that a small change in price never causes a large change in take.
Two cautions. The margin must be computed on a verified cost base, or the band creates exactly the incentive to inflate cost that section 2.4 describes: an operator that can push its declared margin below m-star pays g-min. The audit engine of section 3.5 is therefore a precondition, not a companion. And the band must be legislated as the regime, not layered on top of the 2023 code's fixed royalty and equity, or it simply adds to an already high take. Part II shows the band in place of the 2023 royalty, holding the corporate rate and equity unchanged: at US$4,000 and above the state receives the same 58% as under the fixed terms; at US$2,000 it receives 48% instead of 55%, and the marginal project's return rises from 4% to 6%. The band is insurance the state provides to investors at low prices, paid for by nothing at high prices.
3.4 Model D: settlement valuation for arrears
An arrears claim is not a debt; it is a probability-weighted bundle of legal positions. Its value to the state is what the state would expect to receive if it pursued each element to a decision, net of what pursuit costs. A published method for computing that value gives the negotiator a floor, gives the reviewing authority a standard, and gives the public a way to judge the outcome. An arrears claim C is a bundle of items i, each with a probability of success in arbitration q-i, expected recovery share w-i, and litigation cost and delay. Its expected value is:
EV(C) = \sum_i q_i\, w_i\, C_i\, e^{-\delta T_i} - \text{Cost}_{lit} - \lambda \cdot \Pr(\text{production loss}) \cdot L_{prod}The first term discounts each item's expected recovery for the time an award takes, typically three to five years. The second is the direct cost of pursuing the claim: counsel, experts, tribunal fees. The third is the term that the Loulo dispute made visible and that most claim valuations omit: the probability that pursuit disrupts production, times the revenue the state loses when it does. At Loulo-Gounkoto that loss was a year of royalty on roughly 680,000 ounces, a year of dividends on the state's 20%, and a year of income tax from suppliers and employees, together well above US$150 million. A negotiating strategy that raises the probability of that loss must count it.
The probabilities q-i are the method's judgement calls, and they should be set from evidence rather than assertion: the record of ICSID and UNCITRAL awards in mining tax disputes since 2000 gives base rates of success and recovery by claim type. Section 8.5 sets out how to build that prior. Part II works an example on a US$512 million claim and finds an expected value near US$47 million once production loss is counted; the actual settlement of US$253 million therefore looks like a good outcome for the state, achieved at a cost that a published method would have avoided. A settlement above EV is a win; a public method makes the win legible and makes the next one cheaper.
3.5 Model E: cost and production audit
The audit models close the loop. The simulator and the adaptive rule both compute the state's entitlement from cost and production; the audit models test whether declared cost and production are true. They are label-free by necessity: Mali has completed too few mining audits to train a supervised classifier, and the audits it has completed were selected by consultants on grounds that are not recorded. The models therefore test each declaration against records held by parties whose interests are not the operator's, and against benchmarks built from other mines.
Production is reconstructed as in the companion papers: from cyanide consumption, which is bought from a small number of importers whose customs entries are on record; from electricity, whether purchased from EDM or generated on site from imported fuel; from doré weights and assays at the refinery; and, for open pits, from the change in pit and waste-dump volume measured by satellite stereo imagery. Each source gives an independent estimate of ounces produced; the declaration is scored on how far below those estimates it sits, in units of each estimate's known noise. Costs are tested category by category. Fuel is compared with import parity for the month of purchase; reagents with published contract prices; management fees with what comparable operators charge; interest with the group's cost of debt; and capital cost with a regional benchmark regression of the form set out below.
\ln \text{Capex}_i = \alpha + \beta_1 \ln \text{Capacity}_i + \beta_2\, \text{Type}_i + \beta_3\, \text{Year}_i + u_i, \qquad \text{query if } u_i > 2\hat{\sigma}Algorithms and illustrative results
4.1 Algorithm: annual regime review
ANNUALLY:
run Model A on each producing mine with realised price and cost; record take share
run Model B with current take; project exploration and 2035 output
IF projected 2035 output < floor: propose lower g_min or higher m_star via Model C
IF realised margin > m_star: confirm g_max applies; no ad hoc renegotiation
publish take share, exploration index and projected output with the budget4.2 Illustrative results
Representative mine: 400,000 ounces a year, US$800 million capital over three years, all-in cost US$1,500, 12-year life, 8% discount rate.
| Regime | Royalty | Corporate rate | State equity | Government take share (PV) at US$4,000 | at US$2,500 |
|---|---|---|---|---|---|
| 2019 code, typical | 6% | 25% | 20% | 46% | 52% |
| 2023 code | 10% | 35% | 30% | 63% | 69% |
The take share rises at lower prices because royalty and free carry do not fall with margin. That is the design flaw the adaptive rule fixes: at US$2,500 the 2023 regime takes 69% of a project earning barely above cost, and the marginal exploration project is not funded. At US$4,000 a 63% share is defensible by any international comparison.
4.3 Reading the results
The table in 4.2 rewards a slow reading, because the number that matters is not the take share at US$4,000 but the way it moves as the price falls. Under the 2019 code the state's share is almost flat across prices, at 41 to 43%; the corporate tax and the 20% equity respond to margin, and the low royalty barely disturbs them. Under the 2023 code the share rises from 55% at US$4,000 to 69% at US$2,000. The royalty doubled and the equity rose by half, and neither instrument cares what the mine cost to run; as the margin shrinks, they take a larger and larger slice of what remains. A code that takes more of a mine's cash flow the less cash flow it has is a code that closes mines at the bottom of the cycle.
The investor's return tells the same story from the other side. At US$4,000 the 2023 terms leave an investor with 36%, which is high by any standard and explains why existing operators keep producing and why the state was able to collect its arrears. At US$2,500 the same terms leave 15%, which is the hurdle for most mid-tier producers, and at US$2,000 they leave 4%, which is a return no board would approve for a new project in Mali. Gold was at US$2,000 in late 2023; nothing guarantees it will not be there again during the decade in which today's exploration decisions turn into production.
The margin band's column shows what the paper proposes. At US$3,500 and above it is indistinguishable from the fixed code: the state gives up nothing at the prices it currently enjoys. At US$2,500 it gives up three points and at US$2,000 seven, and in exchange the investor's return at those prices rises to a level at which exploration continues. Whether that trade is worth making depends on the exploration elasticity, which section 3.2 and Part II treat, and on the security discount that Mali cannot remove by fiscal policy. What the table settles is narrower: the 2023 code's take at low prices is not a design choice anyone would make deliberately; it is a by-product of using price-insensitive instruments to raise a price-sensitive target.
4.4 What the illustrative numbers do not show
The representative mine is a stylisation. Actual Malian mines differ in cost (Fekola is among the lowest-cost mines in West Africa; Syama's sulphide operation is among the higher-cost), in remaining life, in the share of their capital already recovered, and in whether a stability clause shields them from the 2023 terms. The take share on a mine that has recovered its capital and has five years of life left is much higher under any regime than on a new project, because there is no capital to deduct; the 2023 code's effect on such a mine is a windfall for the state with no investment consequence, which is part of why the government's revenue rose in 2024 while output fell. The regime's effect on new projects, where the paper's concern lies, is what the mine-by-mine runs proposed in section 8.2 would show.
The illustrative results also hold the price fixed within each run. A real mine faces a path, and the order of prices matters: a project that starts production into a falling market recovers capital slowly and its investor bears a high effective take for longer. The two-factor price process in the series overview, section 4.2, and the 10,000-path simulation specified in Part II address that; the numbers in 4.2 are the single-price slice of that distribution and should be read as such.
Solutions, software, data science and AI models
| Issue | Solution | Data science or AI model | Software component |
|---|---|---|---|
| Take versus investment | Adaptive margin-linked take; published annual regime review | Fiscal simulator across price distributions; exploration-response model | Regime simulator shared by Finance and Mines |
| Stabilised-contract ambiguity | Public register of which regime applies to each licence | Rules-as-code encoding of each convention | Licence and fiscal-terms register |
| Arrears settlements | Published settlement-valuation method | Decision-tree expected value with production-loss term | Claims valuation tool |
| Cost inflation and transfer pricing | Benchmark-based cost audit; advance pricing on key inputs | Capex benchmark regression; related-party graph; fuel and reagent price tests | Mining audit workbench |
| Artisanal exports | Licensed buying counters with same-day payment and low levy; mirror-data enforcement | Channel-choice model (as in the Ghana paper); Dubai import mirror analysis | Buying-counter system; trade-mirror dashboard |
| State equity oversight | Shareholder audit rights exercised through a technical unit | Production reconstruction; dividend reconciliation | Physical-balance engine (ONYX-class) |
Software stack. Existing tax system and mining cadastre as sources; Python with Airflow; PostgreSQL bitemporal warehouse; NumPy, SciPy, PyMC for simulation; scikit-learn, PyOD; Splink and Neo4j Community; Sentinel-2 satellite pipeline with rasterio; Superset; Docker and MLflow. French-language interfaces throughout. Hosting within Mali with an encrypted replica.
Governance of AI. The regime review is published so that operators can plan and the public can judge. Audit scores select; assessments are officer decisions with appeal. Settlement valuations are reviewed by an independent panel before signature.
Team. A mining fiscal unit of eight across Finance, Mines and the state holding entity, with 18 months of external support. The unit's first job is the licence register, without which nothing else can be forecast.
5.7 The solution described as a system
The table in 5.1 maps issues to components; this section describes how the components work together as one system, because a mining fiscal unit that builds them separately will build them twice.
The register is the foundation. Every other component reads from it. The licence and fiscal-terms register records, for each mining title, the regime in force with its effective and expiry dates, the stability clause if any, the state's equity and option terms, and any settlement that modified them. It is stored bitemporally, meaning that it records both when a term was in force and when the state learned of it, so that a revenue forecast made in 2026 can be reproduced in 2028 even after the terms have been amended. Encoding the terms is legal work as much as data work; it needs a lawyer who has read every convention and a data engineer who can express the result as rules. Until it exists, the simulator has nothing to simulate and the audit engine has nothing to check against.
The simulator turns the register into forecasts. It reads the regime for each mine from the register, the mine's parameters from its technical report, and a price distribution from the two-factor model, and it produces the state's expected revenue by mine and year with intervals, the take share, the investor's return, and the probability that each new project proceeds. Its outputs feed the budget's revenue forecast, the annual regime review, and the Ministry of Mines' assessment of which exploration licences are likely to become mines. It is a few thousand lines of Python and runs in minutes on a laptop; its value is entirely in the parameters, which is why the register and the technical-report database matter more than the code.
The audit engine tests what the simulator assumes. The simulator takes declared cost and production as inputs. The engine tests them: it reconstructs production from customs, EDM, refinery and satellite data, scores each mine's declaration, benchmarks each cost category against the regional database, and maps related-party suppliers through the beneficial-ownership graph. Its outputs are audit rankings with per-mine decomposition, hard flags with the identity breached, and a schedule of cost adjustments that feeds back into the simulator's estimate of the state's entitlement. It runs annually on every mine and monthly on production, and its results are what the state's directors on each mine's board take into the dividend discussion.
The claims valuation tool is the audit engine's downstream. When an audit produces a claim, the tool values it: item by item, with probabilities from the arbitration-outcome database, discounted for time, net of litigation cost and of the production-loss term. It is a spreadsheet with a database behind it, and its importance is procedural rather than technical: it produces a number that an independent panel reviews before the negotiator sits down, so that the settlement is measured against something.
The buying-counter system and the trade-mirror dashboard serve the artisanal sector. The counter system records every purchase (seller, weight, assay, price, payment time, provenance declaration) and computes the levy; it is the same design as GoldBod's in Ghana and SONASP's in Burkina Faso, and could be procured from either. The trade-mirror dashboard compares Mali's declared gold exports with UAE, Swiss and Indian import records by month, to bound the artisanal leakage and to measure whether the counters are reducing it.
Hosting, language and control. All components run in Mali, in French, on infrastructure the state controls, with an encrypted replica outside the country against loss. No taxpayer-identifying data leaves the jurisdiction. The regime review, the settlement valuations and the audit summaries are published; the underlying commercial data is not.
5.8 Cost, staffing and sequencing of the build
The unit that runs this is eight people: a head with mining-tax experience, a lawyer for the register, two data engineers, two analysts with mining-finance or engineering backgrounds, a geospatial specialist, and a liaison to the state's board representatives. External support of about 18 months covers the register encoding, the simulator and audit-engine build, the benchmark database, and training. The software is open source throughout; the cost is people and data, and an indicative budget is in the low single-digit millions of dollars over two years, against mining receipts of about US$1.4 billion a year. Recovery of a single under-declared cost category at one large mine would fund the unit.
Sequencing follows dependency. The register comes first, because nothing else can be forecast or audited without it; three to six months. The simulator follows, because it can run on public technical reports while audit data access is negotiated; six to nine months. The audit engine's production module comes next, because customs and EDM data are the easiest to obtain; the cost-audit module and the benchmark database follow, because they need regional data assembled; twelve to eighteen months. The claims tool and the buying-counter system are independent of the others and can run in parallel. The regime review is published at the first budget after the simulator is live, and every year thereafter.
Reform sequence
| Phase | Period | Measure | Model |
|---|---|---|---|
| 1 | 2026 Q4 | Publish licence-by-licence fiscal terms register; state which regime applies to Loulo and each renewal | Register |
| 2 | 2027 | Adopt a settlement-valuation method by decree; apply to remaining audit claims | D |
| 3 | 2027 | Amend the code to add a margin-linked take band in place of fixed royalty increases for new projects | C |
| 4 | 2027 to 2028 | Cost and production audit of the five largest mines; benchmark database | E |
| 5 | 2028 | Licensed artisanal buying counters with a low levy; mirror-data enforcement with UAE authorities | Ghana Model A |
6.1 The phases explained
Phase 1, the register (2026 Q4). The Ministry of Mines, the DGI and the Ministry of Economy and Finance jointly publish a register listing every mining title, the code and convention under which it operates, the state's equity and option terms, the stability clause and its expiry, and any settlement. The single most important entry is Loulo-Gounkoto's, because the renewed licence's terms are unstated and the country's largest mine cannot be forecast until they are. Publishing the register resolves the ambiguity of section 2.2 by forcing the state to decide, mine by mine, which regime applies; that decision is the state's to make, and making it in public is what ends the litigation risk. Deliverable: a published register and the encoded rules behind it. Owner: a joint unit reporting to both ministers. Risk: the decision on stabilised mines is contested by operators; mitigated by publishing the legal basis for each entry.
Phase 2, the settlement method (2027). A decree adopts the claims-valuation method of section 3.4 and applies it to the audit claims still outstanding. The method's probabilities are set from the arbitration-outcome database, its discount rate from the state's cost of finance, and its production-loss term from the Loulo experience. An independent panel, drawn from the Cour des comptes and outside experts, reviews each valuation before the negotiator receives it. Deliverable: the decree, the database, and published valuations for settled claims. Owner: the Ministry of Economy and Finance. Risk: the method is seen as tying the negotiator's hands; mitigated by treating the valuation as a floor, not a target.
Phase 3, the margin band for new projects (2027). An amendment to the 2023 code replaces the fixed royalty increase with the margin-linked band of section 3.3 for new mining titles and for renewals, leaving existing stabilised conventions unchanged until expiry. The band's parameters, g-min, g-max, m-star and kappa, are set from the simulator's runs on the projects Mali hopes to attract, at the 25th-percentile price of its distribution. Deliverable: the amendment and the published parameter calibration. Owner: the Ministry of Mines with the National Assembly. Risk: the band is seen domestically as a concession; mitigated by showing, from the simulator, that it costs nothing at current prices.
Phase 4, cost and production audit (2027 to 2028). The audit engine runs on the five largest mines: production reconstruction from customs, EDM, refinery and satellite data, then cost benchmarking category by category against the regional database. Findings become queries to operators, then adjustments to the state's entitlement, then, where warranted, claims valued by the Phase 2 method. Deliverable: audit reports, adjustment schedules, and a benchmark database shared with Ghana and Burkina Faso. Owner: the DGI's mining unit. Risk: data access is refused or delayed; mitigated by statutory data-sharing provisions in the same amendment as Phase 3.
Phase 5, artisanal counters (2028). Licensed buying counters with same-day payment at close to world prices and a levy set by the channel-choice model with a formal-share floor, in the regions the state controls, with provenance rules tied to a register of artisanal sites. Mirror-data cooperation with the UAE measures what escapes. Deliverable: the counter network, the levy calibration, and a published monthly leakage estimate. Owner: a new state buying entity or a mandate to an existing one. Risk: counters in contested regions become conflict-finance channels; mitigated by restricting the network to state-controlled areas and by provenance checks, accepting that some gold will remain informal.
6.2 What can go wrong in the sequence
The register can be delayed by the political difficulty of stating which regime applies to Loulo; if it is, every later phase waits. The settlement method can be adopted and ignored; the independent panel is the safeguard. The margin band can be legislated as an addition to the 2023 terms rather than a replacement, which would raise take further and defeat the purpose; the amendment's drafting must be explicit. The audit engine can be starved of data by operators who cite confidentiality; the statutory provision is the remedy, and the state's position as shareholder gives it a second route through board audit rights. The counters can fail as Ghana's 2020 levy failed, if the levy is set before the price and payment terms are right; the model's parameter estimates from Burkina Faso's experience are the guard against that.
Limitations
7.1 The limitations discussed
Stylised codes. The simulator in section 4 encodes each code as three numbers. The actual 2023 code has royalty tiers by product and price, a special tax, the local development contribution, transition rules for existing conventions, and provisions on the purchase option's price and timing that the stylisation ignores. Part II's algorithm D.1 sets out the full encoding; until it is done, every take figure in this paper is approximate to within several points, and the comparison between regimes is more reliable than the level of either.
The exploration elasticity. The model's long-run results depend on a parameter estimated elsewhere, on other countries, in other decades. Mali's exploration is also depressed by insecurity, by the exit of Western juniors after 2020, and by the loss of access to regional capital markets, none of which a fiscal model captures. A panel estimate on Mali's own licence data, with a security index as a control, is the only way to know how much of the exploration decline is fiscal; section 8.3 sets it out, and until it exists the exploration results should be read as a range, not a point.
What the state cannot reach. Some of Mali's gold is mined in areas the state does not control, and no instrument in this paper reaches it. The paper's artisanal recommendations apply to the south and west; the north and centre are a security problem before they are a fiscal one. An honest statement of the fiscal base is the share of the country's gold that any instrument can reach, and section 8.7 proposes measuring it.
Undisclosed settlements. The paper's account of the 2024 and 2025 settlements rests on what Barrick and the Finance Minister have said publicly. The other settlements' terms, and the allocation of the CFA 761 billion among them, are not public. The claims-valuation method is proposed partly because the state cannot show, with the information it has released, that the settlements were well judged.
The state as shareholder. The paper assumes the state can obtain cost-audit access through its board seats. The shareholders' agreements that govern that access are not public, and the state's rights may be narrower than assumed. If they are, the audit engine's cost module depends on the statutory provision proposed in Phase 4, and the timeline lengthens.
Price. Every result is conditional on a price distribution centred near current levels. A sustained fall to US$2,500 would change the fiscal picture more than any parameter in the paper, and it would change it in the direction the paper warns of: the fixed 2023 terms would take more of a smaller margin, and the marginal project would not be built.
Extended modelling: improvements and further research
8.1 Research update, September 2026
Loulo-Gounkoto's licence was renewed for ten years by decree on 13 February 2026, following the November 2025 settlement in which Barrick paid US$253 million in arrears, withdrew its ICSID arbitration and regained operational control. Barrick has not given a timeline for full production; ramp-up requires remobilising the workforce, maintenance and rebuilding supply chains. Neither party has stated whether the renewed licence sits under the 2023 code's ownership and fiscal terms or the earlier framework. The Finance Minister has said all companies will now fall under the 2023 code and that the audit is expected to add about US$1 billion a year in future revenue, without publishing the calculation. Gold prices retreated from above US$5,000 in January to the low US$4,000s by mid-2026. The paper's central concern, take at high prices versus survival at low ones, is therefore live in the current year.
8.2 Model A, fiscal regime simulator: improvements
8.3 Model B, exploration response: improvements
8.4 Model C, adaptive take rule: improvements
8.5 Model D, settlement valuation: improvements
8.6 Model E, cost and production audit: improvements
8.7 New models proposed
8.8 Further research
Sources for this section: Skillings, Mali renews Loulo-Gounkoto licence, February 2026; Skillings, full implementation of Mali's revised mining code; Ecofin Agency, Mali expects extra $1 billion a year after sector audit.
A. Worked example 1: take and investor return across the price cycle
Representative mine: 400,000 ounces a year, all-in sustaining cost US$1,500, capital US$800 million over three years, 12-year life, 8% discount rate. Regimes are stylised: the 2019 code as 6% royalty, 25% corporate rate, 20% state equity; the 2023 code as 10% royalty, 35% corporate rate, 30% state equity; and the paper's proposal, the 2023 code with the royalty replaced by a margin band from 6% to 10% centred on a 45% margin.
| Price (US$/oz) | 2019 code: take, IRR | 2023 code: take, IRR | 2023 with margin band: take, IRR |
|---|---|---|---|
| 2,000 | 0.41, 8% | 0.55, 4% | 0.48, 6% |
| 2,500 | 0.42, 21% | 0.57, 15% | 0.54, 17% |
| 3,000 | 0.43, 31% | 0.58, 23% | 0.56, 24% |
| 3,500 | 0.43, 39% | 0.58, 30% | 0.57, 31% |
| 4,000 | 0.43, 46% | 0.58, 36% | 0.58, 37% |
| 5,000 | 0.43, 58% | 0.58, 47% | 0.58, 47% |
At US$4,000 and above, the margin band gives the state the same 58% as the fixed 2023 terms. At US$2,000 it gives up seven points of take (0.48 against 0.55) and in exchange lifts the investor's return from 4% to 6%; at US$2,500 it gives up three points and lifts the return from 15% to 17%. The band is not a concession at high prices; it is insurance at low ones, paid for only when the insurance is needed. Note that the stylised take share for the 2019 code differs from the 46% quoted in section 4 because this run uses a three-year capital profile rather than the simpler treatment there; the ordering and the size of the difference between codes are the same.
B. Worked example 2: exploration response and the 2035 base
Using the exploration-response model of section 3.2 with elasticity 1.2 and a discovery function with exponent 0.7, indexed to the 2019 regime:
| Government take | Exploration spending index | Reserves added index |
|---|---|---|
| 46% (2019 regime) | 100 | 100 |
| 55% | 80 | 86 |
| 63% (2023 regime at high prices) | 64 | 73 |
| 69% (2023 regime at low prices) | 51 | 63 |
With a discovery-to-production lag of eight to twelve years, reserves added in 2026 to 2028 become production in 2035 to 2040. A regime that holds take at 63% therefore implies about a quarter less new production in the late 2030s than the 2019 regime would have, before any security effect. The margin band, by keeping take near 48 to 54% in low-price years when exploration decisions are most sensitive, recovers most of that. The elasticity is the key unknown; section 8.3 sets out how to estimate it on Mali's own licence data.
C. Worked example 3: valuing an arrears claim
A claim of US$512 million made up of four items, each with an assumed probability of success in arbitration and expected recovery share, discounted for three years at 10%:
| Item | Claim (US$ m) | Probability of success | Recovery share if successful | Expected value (US$ m) |
|---|---|---|---|---|
| Royalty base | 180 | 0.6 | 0.8 | 86 |
| Transfer pricing | 150 | 0.4 | 0.6 | 36 |
| VAT refunds | 90 | 0.7 | 0.9 | 57 |
| Dividends | 92 | 0.5 | 0.7 | 32 |
| Sum before discount | 512 | 211 | ||
| After three-year discount | 157 | |||
| Less litigation cost | -20 | |||
| Less expected production loss (0.5 probability of a one-year stoppage costing US$120 m royalty and US$60 m dividends) | -90 | |||
| Expected value of pursuing to award | 47 |
On these assumptions the state should accept any settlement above about US$47 million and would be well served by one near US$150 million, which is roughly the discounted expected value before the production-loss term. The actual settlement of US$253 million exceeds both, which supports the state's negotiating result; the cost was that the production loss occurred anyway. A published method would have made the same result available without the stoppage.
D. Algorithms in full
D.1 Licence-by-licence regime encoding
ALGORITHM Encode_Fiscal_Terms
FOR each mining licence L:
regime[L] = {code version, royalty schedule (tiers by price and product), CIT rate, depreciation rule,
free carry, purchase option (share, window, price formula), local development fund rate,
stability clause (scope, expiry), transition provisions, settlement terms if any}
source[L] = {decree number, convention date, settlement document}
effective_from[L], effective_to[L]
STORE in bitemporal table keyed by (licence, effective date, knowledge date)
VALIDATE: every producing mine has exactly one regime in force per date; unresolved cases listed for Ministry decision
PUBLISH register (terms only, not commercial data)D.2 Annual regime review
ALGORITHM Annual_Regime_Review
INPUT: encoded regimes; technical-report parameters per mine; price distribution; exploration panel data
STEP 1 For each mine: simulate take and IRR over price paths under the regime in force
STEP 2 Estimate exploration elasticity on licence panel: ln X = a + eps*ln((1-g)V) + beta*security + year effects
STEP 3 Project reserves and production to 2040 under the current take and under the margin band
STEP 4 If projected 2035 output < policy floor: propose g_min reduction or m_star shift; quantify revenue cost
STEP 5 If realised margin > m_star for any mine: confirm g_max applies; no ad hoc renegotiation
STEP 6 Publish with the budget: take share by mine (aggregated by size class), exploration index, projected outputD.3 Settlement valuation
ALGORITHM Value_Claim
INPUT: claim items with amounts; case-outcome priors from ICSID/UNCITRAL mining-tax awards since 2000;
expected arbitration duration T; litigation cost; production-loss scenario (probability, royalty, dividend, supplier tax)
FOR each item: EV_i = q_i * w_i * C_i * exp(-delta * T)
EV = sum EV_i - litigation cost - Pr(loss) * L_prod
REPORT EV, its 10th/90th percentile under prior uncertainty, and the settlement floor
SUBMIT to independent review panel before signature; publish EV after settlementD.4 Cost audit
ALGORITHM Cost_Audit
FOR each cost category j (fuel, reagents, power, labour, contractors, management fees, capex):
benchmark_j from West African regression on public technical reports: ln cost_j = f(capacity, type, year, country)
tolerance k_j = 1.5 * residual SD of the regression
if declared_j > benchmark_j * (1 + k_j): query; hold the excess in the audit adjustment schedule
RELATED PARTIES: build supplier graph from beneficial-ownership registers; any supplier linked to the mine's group with price above benchmark is a transfer-pricing query
REALISED PRICE: reconcile declared realised price per shipment to LBMA fix on shipment date; hedge contracts disclosed and tested
OUTPUT: adjusted cost base; recomputed royalty (where cost-linked), CIT and dividendsE. Data specification for Mali
| Data | Holder | Status | Action |
|---|---|---|---|
| Licence terms, conventions, settlements | Ministry of Mines; DGI | Held, unencoded | Encode per D.1 |
| Technical reports | Operators (public filings) | Public | Load parameters per mine |
| Exploration permits and expenditure | Mining cadastre | Held | Panel dataset |
| Security events by district | ACLED (public) | Public | Security index |
| Mine monthly returns | Ministry of Mines; DGI | Held | Standardise categories |
| Power, reagents, refinery receipts | EDM, customs, refiners | Partly held | Statutory access |
| Artisanal exports and UAE imports | Customs; UN Comtrade | Public with lag | Mirror dashboard |
| Beneficial ownership | Registry (RCCM) | Partial | Complete for mining suppliers |
F. Sensitivity analysis
| Parameter | Range | Effect |
|---|---|---|
| Exploration elasticity | 0.6 to 1.8 | Reserves index at 63% take from 86 to 58 |
| Discovery exponent | 0.5 to 0.9 | Widens or narrows the gap between regimes by about a third |
| Band centre m-star | 0.35 to 0.55 | Take at US$2,500 from 0.51 to 0.56; at US$4,000 unchanged |
| Band slope kappa | 6 to 20 | Smoothness only; revenue variance falls with lower kappa |
| Security index | current to halved | Exploration index rises 15 to 30 points independently of fiscal terms |
| Arbitration priors | success 0.3 to 0.8 | Claim EV from 90 to 230 before losses |