Abstract
Tanzania settled its 2017 dispute with Barrick by creating Twiga Minerals, a joint venture in which the state holds a 16% free-carried interest and economic benefits are shared 50/50 after costs. Seven years on, the model has become the template: the 16% free carry now appears in new critical-minerals projects, mining's GDP share has passed 10%, mineral exports reached US$5.4 billion in 2025 and the Ministry of Minerals collected TZS 1.39 trillion in 2025/26, 16% above target. The paper asks the question the numbers do not answer: is 50/50 actually 50/50? A benefit-sharing formula is only as good as the cost base it is computed on, and the state's share of Twiga's benefits depends on costs that Twiga's operator controls. The paper specifies a benefit-sharing verification model that recomputes the state's entitlement from independently reconstructed production and benchmarked costs, a royalty-versus-share optimisation across price paths, an artisanal channel model for the sector that supplies much of Tanzania's gold, and a Bank of Tanzania gold-purchase model. The illustrative finding is that a 5% cost overstatement in a 50/50 structure moves about 2.5% of pre-tax cash flow from the state to the operator, which on Twiga's scale is tens of millions of dollars a year, and that the verification cost is a small fraction of that.
Introduction
The Twiga settlement of 2019 ended a standoff that had included a US$190 billion tax assessment, an export ban on concentrates and the halting of Barrick's Tanzanian mines. The deal gave the state 16% free-carried equity in a new company and a 50/50 split of economic benefits, and Barrick paid US$300 million. Since then Barrick reports US$4.24 billion contributed to the Tanzanian economy, US$888 million in 2024 alone, and Tanzania has extended the 16% structure to niobium, graphite and rare-earth projects.
The research question is one of verification. Benefit sharing is a profit-based instrument, and profit-based instruments are only as reliable as the audit behind them. The paper is written for the Ministry of Minerals, the Tanzania Revenue Authority, the Mining Commission and the state's representatives on Twiga's board.
The organising identity:
R^{pot} - R = \underbrace{\left(R^{pot} - R^{law}\right)}_{\text{policy gap: instrument mix, incentives}} + \underbrace{\left(R^{law} - R\right)}_{\text{compliance gap: cost base, transfer pricing, artisanal leakage}}Current tax issues
2.1 Verifying the cost base of a 50/50 split
The Twiga arrangement of 2019 ended one of the most damaging resource disputes of the decade. In 2017 Tanzania banned the export of gold and copper concentrate, issued Acacia Mining, then Barrick's majority-owned subsidiary, with a tax assessment of about US$190 billion, and passed laws allowing the renegotiation of what it called unconscionable contracts. Production at Bulyanhulu stopped and North Mara was curtailed. The settlement gave the state a 16% free-carried interest in a new company, Twiga Minerals, a 50/50 split of the economic benefits of the mines after costs, a US$300 million payment from Barrick, and the sale of the stockpiled concentrate through Twiga. Since then Barrick reports US$4.24 billion contributed to Tanzania's economy, US$888 million in 2024 alone, Bulyanhulu and North Mara have returned to Tier One status at more than 500,000 ounces a year, and Tanzania has written the 16% carry into law for all mining licences and extended the benefit-sharing template to niobium, graphite and rare-earth projects.
The 50/50 split is a profit-based instrument and, like every profit-based instrument, it is only as good as the profit figure it divides. Economic benefits are what remains of revenue after operating cost, sustaining capital, exploration, and the costs of running a company. Every one of those is incurred, procured, allocated and reported by the operator, which is also the majority shareholder and the manager of the mines. Related-party services from Barrick's group, fuel and reagent supply contracts, technical fees, and the allocation of capital between Bulyanhulu and North Mara all move the split. There is no public account of how Twiga's costs are audited on the state's behalf, of what audit rights the shareholders' agreement confers, or of whether they have been exercised.
This is not an allegation of misconduct; it is a statement about incentives and verification. An operator that reports costs honestly has nothing to fear from a cost audit, and the state that does not conduct one has no way of knowing which kind of operator it has. Part II shows that a cost overstatement of 5% on a 500,000-ounce operation at US$4,000 moves about US$30 million a year from the state to the operator, counting both the split and the corporate tax; a verification unit costs a small fraction of that. The case for the audit does not depend on suspicion; it depends on arithmetic, and it applies with equal force to the critical-minerals projects that now inherit the same template.
2.2 Free carry without cash
A free-carried interest gives the state a share of a company without the state paying for it or contributing to its capital. It is attractive precisely because it costs nothing up front. Its weakness is that it pays only when the company declares a dividend, and the timing and size of dividends are decided by the board, which the majority shareholder controls. A mine that is reinvesting in a new pit, repaying project debt, building working capital or simply retaining cash can go years without a distribution, and every one of those decisions is within ordinary management discretion.
For the state this means that a 16% carry is a claim on a residual whose timing it does not control. The benefit-sharing test in section 3.1 must therefore include when the state's half is paid, not only how much it is. If the operator recovers its own capital first, at its own pace, and the state's dividends begin only afterwards, the state's present-value share of the mine is well below 50% even if the arithmetic of the split is exact. Section 8.2 proposes a dividend-timing model of the cash-flow waterfall, and Part II shows that at a 60% payout ratio the ordering between royalty and share instruments reverses at low prices: the royalty is cash now, the share is cash when the operator decides. The remedy is a dividend policy in the shareholders' agreement that defines distributable cash by formula, so that the 16% is a right to a defined stream rather than to a board's discretion.
2.3 Royalty at 6% plus inspection fee
Tanzania charges a 6% royalty on gold plus a 1% clearing fee on the value of mineral exports, both introduced in 2017 (see section 8.1). Both are gross-value instruments: they take a share of revenue regardless of what it cost to earn. FERDI's mining-tax database estimated that the 2017 changes, together with the 16% carry, raised the average effective tax rate on a representative 3 g/t mine from 41% to 53% at a gold price of US$1,250. At US$4,000, the 7% of revenue taken at the top is a much larger sum, and a much larger share of total take, than the 2017 designers modelled.
Gross-value instruments have two virtues and one flaw. They are easy to verify, because revenue can be checked against refinery receipts and prices without any view of cost; and they pay from the first ounce, which gives the Treasury cash while the profit-based instruments wait for capital recovery. Their flaw is that they do not respond to cost: a mine at US$1,900 all-in cost and a mine at US$1,200 pay the same 7% at the same price, so the high-cost mine pays a far larger share of its margin, and at low prices a gross-value charge can push a marginal mine into suspension. The instrument mix now leans on gross-value charges for cash and on the benefit-sharing split for rent, and the two need to be tuned together, not separately. Section 3.3 sets out the optimisation; Part II's second worked example finds that the share instrument dominates on expected value and that the royalty's role is cash timing, which is a narrower role than the 2017 reforms assumed.
2.4 Artisanal gold and the mineral markets
Tanzania's artisanal and small-scale sector employs, on most estimates, more than a million people and produces a large share of the country's gold, historically sold through informal dealers and across the borders to Kenya and Uganda and onward to Dubai. In 2019 the government established mineral trading centres in the mining regions, where licensed dealers buy from small-scale miners under state supervision and royalty is collected at the point of sale. Recorded gold exports rose sharply in the following year, which the government and Barrick both attributed partly to the centres deterring smuggling. The Mining Commission issued nearly 9,000 small-scale licences between July 2025 and March 2026, STAMICO supports miners through demonstration centres and drilling services, and 99 areas have been identified for allocation to small-scale miners.
The centres are a state buying channel of the kind analysed in the Ghana and Burkina Faso papers, and the same model applies. Artisanal sellers compare the centre's price, net of royalty and the dealer's margin, with the price available across the border, and they are sensitive to the spread. The remaining leakage to Kenya, Uganda and the Gulf is therefore a function of that spread and of payment speed at the centres. The trading-centre purchase data since 2019 is long enough to estimate the channel-choice parameters, and Part II sets out the estimation. The 9,000 new licences raise a further question that section 8.6 addresses: what revenue each licence is expected to yield, so that licensing targets are set on revenue rather than on count.
2.5 Central bank gold purchases
The Bank of Tanzania buys domestic gold to build reserves, as several central banks in gold-producing countries now do. That creates a second state buyer alongside the trading centres, with its own price, its own procurement channel and its own data. If the two buyers set different prices, they compete for the same gold, and the spread between them is a discount that one of them is paying for no reason; if their data are not combined, neither has a complete view of domestic gold flows.
Coordinated, the two buyers form something no private party can match: a single, published state price at which artisanal gold is bought anywhere in the country, with the Bank taking reserves at that price from the centres' purchases, and a single ledger that reconciles purchases, refining, reserve additions and exports. The Zimbabwe paper in this series works through what that ledger looks like and why it matters for a currency backed by gold; Tanzania's shilling is not gold-backed, but the reserve-management logic is the same. Section 3.4 sets out the unified-buyer model and Part II the price rule.
2.6 Critical-minerals pipeline on the same template
Tanzania is applying the Twiga structure, a 16% free carry and benefit sharing, to a pipeline of projects outside gold: the Panda Hill niobium project, expected to deliver about US$768 million to the state over its life through royalties, taxes, levies and dividends; the Mahenge graphite project; the Wigu Hill rare-earth licence now held by STAMICO; and helium. The 2026 budget frames these as a new era of value addition, and mineral exports of US$5.4 billion in 2025, more than half of the country's export revenue, give the strategy momentum.
Each of these commodities has a cost structure the state has never audited. Niobium is sold as ferroniobium into a thin market with few published prices; graphite is sold on specification into battery supply chains with long-term offtake contracts whose prices are not public; rare earths are priced on reference indices that the largest buyers influence; helium is sold under contracts with industrial-gas majors. Benefit sharing on any of them requires the state to verify revenue as well as cost, which for gold is trivial and for these is not. The verification capacity the paper proposes for gold, in sections 3.1 and 3.2, must be built to transfer: a benchmark database per commodity, physical custodians identified before first production, and offtake contracts reviewed for related-party pricing. Section 8.6 sets out the transfer; the reform sequence in section 6 places it before first production, not after.
Models
3.1 Model A: benefit-sharing verification
The verification model does what the state's side of the Twiga board should have been doing since 2019: it recomputes the state's entitlement under the 50/50 split from a cost base the state has checked, rather than from the cost base the operator reports, and it measures the difference. The arithmetic is simple; the work is in the verification. Economic benefits under a 50/50 arrangement are the sum of what the state receives through royalty, taxes, fees and dividends, and what the operator receives after those. Write revenue R, verified cost C-hat, royalty and fees F, corporate tax T, and dividends D:
B^{state} = F + T + D, \qquad B^{op} = R - \hat{C} - F - T - D, \qquad \text{Gap} = \tfrac{1}{2}\left(R - \hat{C}\right) - B^{state}The gap is the state's shortfall: the half of verified benefits it is entitled to, less what it has received through every channel. The state's entitlement is recomputed using a verified cost base rather than the declared one. Verification proceeds cost category by cost category, because the categories differ in how they can be checked and in how much they vary legitimately across mines. Fuel can be checked against import parity to within a few percent; management fees against what comparable operators charge, with a wide tolerance; capital cost against a regression on capacity and type, with a wider one still. Verified cost is declared cost adjusted for benchmark deviations on each cost category j:
\hat{C} = \sum_j \min\!\left(C^{dec}_j,\; \tilde{C}_j\,(1 + k\,\hat{\sigma}_j)\right)where tilde-C is the benchmark for the category (import-parity fuel, contract reagent prices, peer management fees, capex per tonne of capacity) and k a tolerance set from the benchmark's own dispersion, so that a category that varies widely across honest mines is not flagged on normal variation. Every category above tolerance is a query to the operator, and the disputed amount is held in escrow until resolved. Escrow, rather than assessment, is the interim remedy because the arrangement is a joint venture governed by a shareholders' agreement, not a tax the state can assess unilaterally; holding the disputed sum protects the state's position while respecting the agreement. The model needs one thing that does not yet exist in public: an operational definition of economic benefits, written as accounting identities over Twiga's accounts and mapped to Barrick's published segment reporting, so that the split can be recomputed by anyone from published data as a first check. Section 8.2 makes that the first task.
3.2 Model B: production reconstruction
Verifying cost is half of verifying the split; the other half is verifying production, because benefits are revenue less cost and revenue is ounces times price. The reconstruction model of the series overview applies. Bulyanhulu is an underground mine, which changes the custodians: satellite volumetrics measure open pits, not stopes, so hoisted tonnes from the winder control system, ventilation and hoisting power, and explosives consumption from licensed suppliers take their place. North Mara has both open-pit and underground sections and uses both sets. As elsewhere in the series, declared output is tested against cyanide, electricity, refinery doré receipts and, for open pits, satellite pit and dump volumes:
A_i = \sum_m \omega_m \max\!\left(0, \frac{\ln \hat{Q}^{(m)}_i - \ln Q^{dec}_i}{\sigma_{m,s}}\right) + \delta H_iFor Tanzania the strongest independent record is the refinery. Doré from Bulyanhulu and North Mara is assayed on receipt, and the gold content received in a period cannot exceed the gold declared as produced in that period, once opening and closing stock are allowed for. This is an accounting identity, not a statistical benchmark, and a breach is a hard flag that generates an automatic query. Part II's third worked example shows an underground mine whose hoist, explosives and refinery records all sit above its declaration, with the refinery figure 2% above declared production; whether that is a stock effect or an identity breach is the audit question, and at US$4,000 the difference is worth about US$24 million of revenue, roughly half of it the state's.
3.3 Model C: instrument mix optimisation
The third model chooses the mix of instruments. Tanzania has three that reach the same mine: a gross-value royalty and fee, corporate tax, and the benefit-sharing split delivered through the state's equity. Each has a different response to price and cost, a different timing of cash and a different need for verification. The model chooses the royalty rate, the fee and the share to maximise the state's expected revenue across the price distribution, subject to two constraints that reflect the state's real situation: the investor must still earn its hurdle, or the next project is not built; and the state must receive a minimum cash amount even in low-price years, because the budget has commitments that do not fall with gold. Choose royalty r, fee f and benefit share s to maximise expected state revenue across a price distribution, subject to an investor hurdle rate and a floor on revenue in low-price years (the cash the Treasury needs regardless):
\max_{r, f, s} \; \mathbb{E}_p\left[F(r,f) + s\,(R - \hat{C} - F)\right] \quad \text{s.t.} \quad IRR^{post} \ge h, \qquad \Pr\!\left(\text{state revenue} < \underline{R}\right) \le \alphaThe model's original conclusion, that the state should rely more on royalty when its verification is weak, is refined by the worked example in Part II. On expected value, a higher share with a lower royalty delivers more to the state at every price, even after allowing for verification error, because the share reaches the rent that the royalty cannot. What the royalty buys is certainty of timing: it is paid from the first ounce, while the share is paid when the operator declares a dividend, and at a 60% payout ratio the royalty delivers more cash at low prices. Verification capacity therefore matters not because it determines whether to use the share instrument but because it determines how much of the share the state actually receives; and dividend policy matters because it determines when. The design that follows is a fixed dividend rule in the shareholders' agreement, a benefit-sharing audit with escrow, and then a lower royalty, in that order. Verification capacity is a fiscal parameter, and so is the dividend policy.
3.4 Model D: unified state gold buyer
The fourth model applies the channel-choice framework to Tanzania's two state buyers. The trading centres and the Bank of Tanzania together set the domestic price for artisanal gold, and each artisanal seller compares that price, net of royalty and dealer margin and payment delay, with the price available in Nairobi or Dubai net of transport and risk. Formal share of artisanal output is logistic in the spread between the state price and the Nairobi or Dubai price net of transport, payment delay and enforcement, as in the Ghana paper. Two features are specific to Tanzania. First, the two buyers are in the same government, so the spread between their prices is a cost the state imposes on itself; a single published price, with the Bank taking reserves at that price from centre purchases, removes it. Second, the centres collect royalty at the point of sale, so the royalty is part of the effective discount the seller faces, and its rate for small-scale sellers should be set by the same Laffer logic as the levy in the Ghana paper: low enough that the formal share stays high, because the fiscal value of formal gold, in income tax, foreign exchange and reserves, exceeds the royalty on the gold that leaves. Coordinating the two buyers on a single published price removes the intra-state competition that widens the spread; Part II gives the weekly price rule and the monthly reconciliation that make the single price operational.
Algorithms and illustrative results
4.1 Algorithm: annual benefit-sharing audit
INPUT Twiga accounts, refinery receipts, utility and customs records, benchmark database
RECONSTRUCT production; flag if refinery receipts or power-implied output exceed declared
FOR each cost category j: compare declared to benchmark; escrow the excess above tolerance
RECOMPUTE state entitlement on verified cost; compare to amounts received
ISSUE queries; release escrow on resolution; publish summary of the gap4.2 Illustrative results
For a 500,000-ounce operation at US$4,000 and all-in cost of US$1,500, pre-tax cash flow is about US$1.25 billion. A 5% overstatement of cost (US$37.5 million) moves half of that, about US$19 million, from the state's half to the operator's, plus the corporate tax on it. A 10% overstatement moves about US$40 million. A verification unit of eight to ten specialists with a physical-balance engine costs a small fraction of that. The instrument-mix model, on the same mine, prefers a 4% royalty and 55% share when verification error is under 3%, and a 6% royalty and 50% share when error is above 8%. All figures are illustrative.
4.3 Reading the results
The arithmetic in 4.2 is deliberately simple, because the point does not need a model. A 50/50 split of benefits computed after cost is a 50/50 split of whatever the cost base leaves, and the cost base is the operator's to report. Each dollar of overstated cost moves fifty cents from the state's half to the operator's and, because the overstatement also reduces taxable profit, removes thirty cents of corporate tax. On a 500,000-ounce operation at US$4,000, 5% is US$30 million a year, 10% is US$60 million. Against that, a verification unit of ten people with the engine and database this paper specifies costs a small fraction, and it would need to find an overstatement of under 1% to pay for itself.
None of this says that Twiga's costs are overstated. It says that the state does not know whether they are, because it has not looked with its own data, and that the cost of looking is trivial relative to what is at stake. The same reasoning applies to the critical-minerals projects that inherit the template, with the difference that their revenue is also hard to verify.
The instrument-mix paragraph in 4.2 needs the correction that Part II's second worked example supplies. On expected value, a higher benefit share with a lower royalty delivers more to the state at every price, even after verification error, because the share reaches rent that a gross-value royalty cannot. The royalty's advantage is timing: it is paid from the first ounce, while the share is paid when the operator declares a dividend. At a 60% payout ratio the royalty delivers more cash at low prices; at full payout the share wins everywhere. So the design question is not royalty versus share; it is dividend policy first, verification second, and the royalty rate last, set to the cash floor the Treasury needs in a bad year.
4.4 What the illustrative numbers do not show
The worked figures assume a representative mine with a single cost and price. Bulyanhulu and North Mara have different costs, different ore, and different capital histories, and the split is computed on Twiga as a whole, so the allocation of shared costs between them is itself a verification question. The arithmetic also assumes that the operational definition of economic benefits is known; it is not public, and section 8.2 makes writing it down, as accounting identities over Twiga's accounts, the first task of the verification unit. Until that definition exists, no one outside Twiga can say what 50/50 means, and the state's directors on its board may not be able to say it either.
Solutions, software, data science and AI models
| Issue | Solution | Data science or AI model | Software component |
|---|---|---|---|
| Unverified cost base | Annual benefit-sharing audit with escrow | Cost benchmark regressions; related-party graph; transfer-pricing tests | Benefit-sharing verification engine |
| Production declared by operator | Independent reconstruction | Refinery, reagent, power and satellite reconstruction; MCD scoring | Physical-balance engine (ONYX-class) |
| Deferred dividends on free carry | Dividend policy in the shareholders' agreement; state director analytics | Cash-flow waterfall model with covenant tests | Board analytics pack for state directors |
| Instrument mix | Periodic recalibration tied to verification capacity | Stochastic optimisation across price paths | Fiscal regime simulator |
| Artisanal leakage | Single published state price; same-day payment | Channel-choice model; mirror analysis of Kenya, Uganda, UAE imports | Unified buyer pricing service; trade-mirror dashboard |
| Two state buyers | Coordination protocol between Bank of Tanzania and trading centres | Flow reconciliation across both buyers | Shared gold-flow ledger |
| Critical-minerals template | Transfer of gold verification methods to niobium, graphite, helium | Benchmark databases per commodity | Multi-commodity audit workbench |
Software stack. Mining Commission, TRA and Bank of Tanzania systems as sources; Python with Airflow; PostgreSQL bitemporal warehouse; NumPy, SciPy, PyMC; scikit-learn, PyOD; Splink and Neo4j Community; Sentinel-2 pipeline; Superset; Docker and MLflow; hosted in Dodoma or Dar es Salaam with an encrypted replica. Swahili and English interfaces.
Governance of AI. Verification findings go to the operator as queries with a response window, then to the board and the Ministry. Escrow, not assessment, is the interim remedy, which respects the joint-venture agreement while protecting the state's share. Selection and scoring are explainable term by term.
Team. A verification unit of eight to ten across the Ministry of Minerals, TRA and the state's board representatives, with 12 months of external support. The same unit serves the critical-minerals projects.
5.7 The solution described as a system
The verification engine is the centre. It takes Twiga's management accounts by mine and cost category, the refinery's receipts by shipment, TANESCO's power records, the hoist, ventilation and explosives records for the underground sections, customs entries for reagents, and Sentinel-2 volumetrics for the open pits. It reconstructs production from each physical source and flags identity breaches, in particular any period in which refinery gold content exceeds declared production after stock adjustment. It benchmarks each cost category against a database built from public technical reports across East and West Africa and from the mines' own history, with a tolerance per category set from that category's normal dispersion. It recomputes the state's entitlement on the verified cost base, compares it with amounts received through royalty, fee, corporate tax and dividends, and produces a schedule of queries with the disputed amounts held in escrow. Its outputs go to the state's directors on Twiga's board, to the Ministry of Minerals and to the Tanzania Revenue Authority.
The board pack turns the engine's findings into shareholder action. The state holds 16% of Twiga and half its benefits, and its directors need the same information the operator's directors have. The pack presents the cash-flow waterfall (revenue, cost, sustaining capital, debt service, working capital, distributable cash), the dividend policy's application to it, the verification engine's queries, and the escrow balance. Its purpose is to make the dividend discussion a discussion about a formula rather than about the operator's discretion.
The fiscal regime simulator serves the instrument-mix review and the critical-minerals pipeline. It encodes the current regime (6% royalty, 1% clearing fee, 16% free carry, 30% corporate rate, 20% straight-line depreciation, benefit share) and any proposed alternative, loads the technical reports for Bulyanhulu, North Mara, Geita, New Luika and Nyanzaga and for the niobium, graphite and helium projects, and runs take, investor return and the timing of state cash across a price distribution. Its output is the frontier of section 8.4: for each combination of royalty and share, the expected value, the low-price cash floor and the investor's return, so that the trade-off is visible rather than assumed.
The unified buyer's pricing service sets one state price. It reads the LBMA fix, an estimate of the smuggler's margin from Nairobi and Dubai spreads and seizure data, and payment terms, and publishes a single state buying price weekly that the trading centres and the Bank of Tanzania both apply. The gold-flow ledger reconciles centre purchases, refinery output, Bank reserve additions and customs exports monthly, as in the Zimbabwe paper.
The multi-commodity workbench extends the engine. For niobium, graphite and helium it holds the physical custodians (concentrate assays, power, reagents, contract volumes), the offtake contracts and their pricing terms, and a benchmark database per commodity, so that when the first benefit-sharing calculation on a non-gold project is due, the verification capacity already exists.
Governance. Findings go to the operator as queries with a response window before they go anywhere else, because a joint venture is governed by agreement and the state's remedy is escrow, not assessment. Selection and scoring are explainable term by term. Summaries of the verification gap are published; commercial detail is not.
5.8 Cost, staffing and sequencing of the build
The verification unit is eight to ten people across the Ministry of Minerals, the TRA and the state's board representatives: a head with mining-finance experience, two mining engineers or metallurgists, two data engineers, two analysts with transfer-pricing training, a geospatial specialist and a lawyer for the shareholders' agreement and the operational definition of benefits. External support of about 12 months for the engine, the benchmark database and training. Open-source software; the cost is people, data agreements and the benchmark database, and it is repaid by a cost adjustment of under 1% on one year's Twiga costs.
Sequencing: the operational definition of benefits and the data-feed agreements first, because the engine cannot run without them; the production module next, because refinery, power and hoist data are obtainable and the identity check is the fastest win; the cost module and benchmark database through 2027; the board pack from the first audit; the unified price whenever the Bank and the Ministry agree, which needs no technology; the multi-commodity extension before the first non-gold project produces.
Reform sequence
| Phase | Period | Measure | Model |
|---|---|---|---|
| 1 | 2026 Q4 | Stand up the verification unit; obtain refinery, utility and customs data feeds | B |
| 2 | 2027 | First benefit-sharing audit of Twiga with escrow mechanism | A |
| 3 | 2027 | Unified state gold price across trading centres and the Bank of Tanzania | D |
| 4 | 2028 | Instrument-mix review tied to measured verification error | C |
| 5 | 2028 onward | Apply the template to niobium, graphite and helium projects before first production | A, B |
6.1 The phases explained
Phase 1, the verification unit and data feeds (2026 Q4). The unit is stood up and the shareholders' agreement's audit rights are exercised to obtain Twiga's management accounts; direct feeds are agreed with the refineries, TANESCO and the explosives suppliers; the operational definition of economic benefits is written as accounting identities and agreed with the operator. Deliverable: the unit, the feeds and the definition. Owner: the Ministry of Minerals with the state's Twiga directors. Risk: audit rights are narrower than assumed or contested; the state's position as shareholder and the Mining Commission's statutory powers are the two routes.
Phase 2, the first benefit-sharing audit (2027). The engine runs on the 2026 accounts: production reconstruction with the refinery identity check, cost benchmarking by category, recomputation of the state's entitlement, queries and escrow. Deliverable: the audit report, the escrow schedule and the first measured verification error, which is the parameter the instrument-mix review needs. Owner: the verification unit. Risk: the escrow mechanism is not in the agreement; it must be negotiated as part of Phase 1 or substituted by a reconciliation-and-adjustment clause.
Phase 3, one state gold price (2027). The Ministry and the Bank of Tanzania agree a single published buying price applied at the trading centres and by the Bank, with the Bank taking reserves from centre purchases at that price, and the gold-flow ledger starts. Deliverable: the price rule and the first monthly reconciliation. Owner: the Ministry of Minerals and the Bank. Risk: institutional reluctance to give up separate pricing; the model's estimate of the formal share lost to the spread is the argument.
Phase 4, the instrument-mix review (2028). With a measured verification error and a dividend policy in place, the simulator runs the royalty-and-share frontier and the government decides whether to lower the royalty in exchange for a defined dividend stream. Deliverable: the published frontier and any Finance Act change. Owner: the Ministry of Finance and the Ministry of Minerals. Risk: the review is conducted before the dividend policy exists, in which case a lower royalty simply defers state cash; the sequence guards against this.
Phase 5, the template for critical minerals (2028 onward). The verification engine, benchmark database and offtake-contract review are extended to niobium, graphite and helium before each project's first production, and the benefit-sharing definition for each is written at the time of the agreement rather than after. Deliverable: commodity-specific modules and definitions. Owner: the verification unit. Risk: projects reach production before the modules exist; the reform sequence places this phase early for that reason.
6.2 What can go wrong in the sequence
The first risk is that Phase 1 stalls on access to Twiga's accounts, which would leave the state as a shareholder that cannot audit its own company; the shareholders' agreement is the remedy if it contains the rights, and legislation is the remedy if it does not. The second is that the first audit produces a large adjustment and the relationship deteriorates as it did in 2017; the escrow mechanism, the query-and-response window and the publication of summaries rather than detail are designed to keep the dispute inside the agreement. The third is that the instrument-mix review lowers the royalty before the dividend policy is fixed, converting certain cash into deferred cash. The fourth is that the critical-minerals projects inherit the template without the verification, which would repeat for niobium and graphite the position the state has been in for gold since 2019.
Limitations
7.1 The limitations discussed
The agreement is not public. The paper assumes the state has, or can negotiate, cost-audit access to Twiga and an escrow or adjustment mechanism. Twiga's shareholders' agreement has not been published, and the state's actual rights may be narrower. If they are, the paper's first phase becomes a negotiation, and the state's leverage is its statutory position rather than its contractual one.
The definition of benefits is unknown. The 50/50 split has been described in public only in outline. Every calculation in this paper assumes a definition that treats benefits as revenue less verifiable cost; if the actual definition nets out items the paper does not anticipate, the worked examples change. Writing the definition down is the first task for that reason.
Benchmarks are thin. Cost benchmarking needs comparable mines, and the regional database the paper relies on does not yet exist in public form. The first audit will benchmark against operator-provided comparables and the mines' own history, which is weaker; the database improves with each audit and with sharing among the countries in this series.
Underground reconstruction is less precise. Hoist, ventilation and explosives records constrain tonnes moved; they constrain grade less well than open-pit volumetrics combined with reagent and power, and the refinery identity does most of the work at Bulyanhulu. A small opening-stock allowance can absorb a 2% discrepancy, so the identity check is decisive only for larger gaps or for sustained patterns.
Artisanal parameters are assumed. The trading-centre data since 2019 can estimate them and has not yet been used for that. Until it is, the case for a single state price rests on the logic of the model rather than on a measured spread elasticity.
The template's transferability. For niobium, graphite and helium the state must verify revenue as well as cost, in markets with few published prices and with offtake contracts that may be related-party. The paper's methods extend to those commodities in principle; in practice the benchmark databases have to be built from scratch, and the paper cannot say how long that will take.
Extended modelling: improvements and further research
8.1 Research update, September 2026: rates confirmed and a fiscal-regime benchmark
The rates queried in section 2.3 are confirmed. The Finance Act 2017 raised the royalty on gold and other metallic minerals from 4% to 6% and introduced a 1% clearing fee on the value of all mineral exports from 1 July 2017; the Written Laws (Miscellaneous Amendments) Act 2017 made the 16% non-dilutable free-carried interest mandatory for all mining and special mining licence holders and allows the state to acquire up to 50% in proportion to tax expenditures granted. The same reforms moved capital allowances to 20% straight-line and removed input VAT credit on raw mineral exports. FERDI's mining-tax database estimated that the 2017 changes raised the average effective tax rate on a representative 3 g/t mine at US$1,250 gold from 41% to 53%. Two things follow for the paper. First, at US$4,000 the gross-value instruments (7% of revenue between royalty and fee) are a much larger share of total take than the 2017 designers modelled, which strengthens the case in Model C for rebalancing toward the profit-based share as verification improves. Second, the 50% acquisition right tied to tax expenditures makes a tax-expenditure register for each mine a fiscal asset, not just a transparency exercise.
8.2 Model A, benefit-sharing verification: improvements
8.3 Model B, production reconstruction: improvements
8.4 Model C, instrument mix: improvements
8.5 Model D, unified state buyer: improvements
8.6 New models proposed
8.7 Further research
Sources for this section: Mondaq, Amendments to the Mining Act in Tanzania, February 2018; Addleshaw Goddard, New mining legislation passed in Tanzania, 2017; FERDI mining taxation database bulletin on Tanzania; Mining.com, Acacia agrees to pay higher taxes; Zambia Monitor on Tanzania's 2026 mining pipeline and small-scale licensing.
A. Worked example 1: the cost base of a 50/50 split
A 500,000-ounce operation at US$4,000 with a true all-in cost of US$1,500 has revenue of US$2,000 million, cost of US$750 million and economic benefits of US$1,250 million. If cost is overstated, the declared benefits fall, the state's half falls with them, and the state also loses corporate tax at 30% on the overstated cost.
| Cost overstatement | Declared benefits (US$ m) | State's declared half | State's entitled half | Shortfall on the split | Lost corporate tax | Total loss to state |
|---|---|---|---|---|---|---|
| 0% | 1,250 | 625 | 625 | 0 | 0 | 0 |
| 5% | 1,212 | 606 | 625 | 19 | 11 | 30 |
| 10% | 1,175 | 588 | 625 | 38 | 22 | 60 |
| 15% | 1,138 | 569 | 625 | 56 | 34 | 90 |
A verification unit of ten specialists with a physical-balance engine and a benchmark database costs, at most, a few million dollars a year. The break-even overstatement it needs to find is under 1%. The case for the audit does not depend on suspecting the operator; it depends on arithmetic.
B. Worked example 2: royalty versus share, and why the answer is about timing
On the same mine, compare a 6% royalty with a 50% benefit share against a 4% royalty with a 55% share, where the share is delivered through the state's dividend entitlement and the operator's payout ratio pi governs when it is paid.
| Price (US$/oz) | 6% royalty plus 50% share, full payout (US$ m a year) | 4% royalty plus 55% share, full payout | 6% and 50%, payout 0.6 | 4% and 55%, payout 0.6 |
|---|---|---|---|---|
| 2,000 | 60 + 95 = 155 | 40 + 116 = 156 | 60 + 57 = 117 | 40 + 69 = 109 |
| 2,500 | 75 + 213 = 288 | 50 + 248 = 298 | 75 + 128 = 203 | 50 + 149 = 199 |
| 4,000 | 120 + 565 = 685 | 80 + 644 = 724 | 120 + 339 = 459 | 80 + 386 = 466 |
On expected value with full payout, the higher share wins at every price. With a 60% payout the ordering reverses at low prices, because the royalty is cash now and the share is cash when the operator decides. Verification error adds to this: a 5% cost overstatement costs the state 2.5% of benefits under a 50% share and 2.75% under 55%. The instrument-mix conclusion is therefore not that royalty should rise when verification is weak, as section 4.3 loosely put it; it is that the share instrument dominates on expected value, and the royalty's role is to guarantee cash in low-price years and while dividends are deferred. The right design fixes the dividend policy in the shareholders' agreement and then lowers the royalty. Section 4.3 should be read with this correction.
C. Worked example 3: reconstruction at an underground mine
Bulyanhulu is underground, so satellite volumetrics do not apply. Hoisted tonnes from the winder system, ventilation and hoisting power, and explosives consumption replace pit volume. Assume the mine declares 320,000 ounces from 1.05 million tonnes at 10.5 g/t and 90% recovery. Hoist records show 1.12 million tonnes; explosives consumption at a norm of 0.9 kg per tonne implies 1.09 million tonnes; refinery doré receipts imply 326,000 ounces.
| Source | Implied tonnes | Implied ounces | z (sigma 0.06) |
|---|---|---|---|
| Declared | 1,050,000 | 319,000 | 0 |
| Hoist | 1,120,000 | 340,000 | 1.07 |
| Explosives | 1,090,000 | 331,000 | 0.61 |
| Refinery doré | not applicable | 326,000 | 0.36 |
All three third-party sources sit above the declaration, and the refinery figure exceeds declared production by 2%. That last point is an identity breach if the doré shipped in the period came from that period's production; the audit question is whether opening stock explains it. If it does not, the breach is a hard flag, and the difference of about 6,000 ounces at US$4,000 is US$24 million of revenue, of which the state's share through royalty, tax and the split is roughly half.
D. Algorithms in full
D.1 Benefit-sharing recomputation
ALGORITHM Recompute_Split
INPUT: Twiga management accounts by mine; refinery receipts; utility, customs and hoist records; benchmark database; shareholders' agreement definitions
STEP 1 Reconstruct production (D.2); if refinery > declared beyond opening-stock explanation: hard flag
STEP 2 For each cost category j: benchmark, tolerance k_j; C_hat_j = min(declared_j, benchmark_j*(1+k_j)); excess -> escrow schedule
STEP 3 Benefits_hat = revenue at reconciled realised price - sum C_hat_j
STEP 4 State entitlement = royalty + fee + CIT recomputed on C_hat + 50% of (Benefits_hat - royalty - fee - CIT) per agreement definition
STEP 5 Gap = entitlement - amounts received; issue queries with 60-day response window
STEP 6 On resolution: release or retain escrow; publish summary gap by categoryD.2 Reconstruction for open-pit and underground mines
ALGORITHM Reconstruct
IF open pit: sources = {throughput, cyanide, power, dore, satellite volumetrics}
IF underground: sources = {throughput, cyanide, power, dore, hoist tonnes, explosives}
FOR each source m: Q_hat[m]; z[m] = (ln Q_hat[m] - ln Q_dec)/sigma_m
H = 1 if dore > Q_dec + opening stock adjustment OR declared tonnes > hoist tonnes (underground) OR declared tonnes > power-implied*1.15
A = sum omega_m max(0, z[m]) + gamma d_MCD(ratios) + delta H
OUTPUT: band, flags, per-source decompositionD.3 Instrument-mix review
ALGORITHM Mix_Review (every three years)
INPUT: measured verification error from the last three audits (SD of cost adjustments as share of cost); dividend policy in force; price distribution
FOR candidate (royalty r, share s) pairs:
simulate state cash by year with payout ratio from policy; apply verification-error loss = err * cost * s
compute expected PV of state cash; 5th percentile of annual state cash; investor IRR
SELECT pair maximising expected PV subject to: 5th percentile cash >= floor; IRR >= hurdle
REPORT the frontier so that the policy trade-off is visibleD.4 Unified state buyer
ALGORITHM Unified_Price
WEEKLY: P_ref = LBMA fix; margin_smug = estimate from Nairobi/Dubai spreads and seizures
P_state = P_ref * (1 - margin_smug + epsilon), same at trading centres and Bank of Tanzania
cash pre-positioned at each centre from delivery forecast
MONTHLY: fit channel-choice on centre purchases; reconcile purchases, refinery output, reserve additions and exportsE. Data specification for Tanzania
| Data | Holder | Status | Action |
|---|---|---|---|
| Twiga accounts by mine | Twiga; state directors | Held under agreement | Audit access clause exercised |
| Refinery receipts | Refiners (domestic and export) | Partly held | Mandatory shipment reporting |
| Hoist, ventilation, explosives records | Mine; explosives suppliers (licensed) | Held by third parties | Statutory access via Mining Commission |
| Power by mine | TANESCO | Held | Direct feed |
| Trading-centre purchases | Mining Commission | Held since 2019 | Channel-choice estimation |
| Bank of Tanzania gold purchases | BoT | Held | Reconciliation ledger |
| Destination imports | UN Comtrade; Kenya, Uganda, UAE customs | Public with lag | Mirror dashboard |
| Technical reports | Barrick, AngloGold Ashanti, Shanta, Perseus | Public | Fiscal simulator calibration |
| Tax-expenditure register by mine | TRA; Ministry of Minerals | Not assembled | Build; it is the basis of the 50% acquisition right |
F. Sensitivity analysis
| Parameter | Range | Effect |
|---|---|---|
| Cost overstatement | 0 to 15% | Loss to state 0 to US$90 m a year on the example mine |
| Payout ratio | 0.4 to 1.0 | Decides whether royalty or share delivers more cash at low prices; crossover near 0.65 at US$2,500 |
| Verification error | 1% to 10% | Share-instrument loss from 0.5% to 5% of benefits |
| Royalty rate | 3% to 7% | Take at US$4,000 from 0.42 to 0.45 before share effects; low-price cash floor moves proportionally |
| Opening-stock allowance | 0 to 5% of annual output | Decides whether a 2% refinery excess is a flag or noise |