Derisking Carbon-Project Investment: The Mombak Case

Framing the investment decision

Proceed to focused diligence, while withholding a conclusion on new investment. Public evidence supports examining risk allocation, historical quantification and early delivery. It is insufficient to calculate Mombak's project equity value, verify the entire contractual cash waterfall or establish that every proposed treatment is effective.

The case centres on the Mombak reforestation programme associated with the World Bank's 2024 Amazon Reforestation-Linked Outcome Bond. Its Phase I asset is Turmalina in Pará, Brazil. The offering document identifies Mombak Angico-Branco Florestal S.A. as its project company.^[WB4] The Isometric registry separately identifies Turmalina within Reforesting the Brazilian Amazon Project 1.^[ISO1] Independent verification confirms Turmalina as the first project instance.^[SCS1] Specific certificate allocation and contractual eligibility still require reconciliation.

Three investment exposures

A bondholder holds an IBRD obligation, with a fixed component and conditional carbon-linked payments. A project investor bears its contractual share of development cost, operating shortfalls and residual cash flows. A carbon buyer bears delivery, eligibility and claim-use risks under its purchase agreement. One party's protection can create an obligation for another.

The working decision is whether a prospective investor should fund further Mombak-related project exposure, fund only diligence and milestone-based commitments, or decline. The analysis uses information available at the report cutoff; it does not price a secondary-market purchase of the existing bond.

What would change the decision

The next commitment should depend on a reconciled asset-to-cash record: legal rights and applicable safeguards; parcel and registry identity; buyer acceptance of the relevant methodology and certificates; delivery invoices and collection rights; and an executable budget covering operations and long-term stewardship. These are proposed investor requirements, not findings that Mombak has failed them.

Unresolved eligibility and rights remain conditions for commitment. A numerical haircut does not resolve them. The report moves from existing protection and observed carbon outcomes to the risk register (pages 6–7), then conditional financial consequences, incremental treatments and the evidence needed to assess their effectiveness. Unknowns remain explicit; exhaustive risk discovery is not claimed.

Understanding the existing protection

The 2024 issuance was a $225 million IBRD bond, mobilising approximately $36 million for Mombak's reforestation activities. The bond proceeds support IBRD's wider development activities; the entire principal was not invested in Mombak.^[WB1]

Foregone conventional coupons are converted into project funding through arrangements involving HSBC. The disclosed structure envisages 11 milestone-dependent disbursements through December 2026, with undisbursed amounts returned to bondholders in January 2027. It links carbon-related payments to part of the proceeds from a Microsoft purchase agreement. The issuer's case study describes an agreement for up to 1.5 million units at pre-agreed prices through 2032. These are financing and purchase commitments, not evidence that the full amounts have been disbursed, delivered or collected.^[WB2]

Financing features: WB1–WB3. Certification and buffer: ISO1–ISO3. Residual-exposure interpretation: Andi analysis.

Defining the baseline and obligation horizon

Existing offtakes, buffers and documented field controls belong in the baseline. Monitoring, reporting and verification (MRV) expenditure is incremental only to the extent it adds to that baseline.^[MOM2][SCS1]

The PDD distinguishes 50-year crediting, a 100-year durability commitment, and financial planning for 150 years including post-crediting monitoring. These source labels do not establish a single reconciled contractual timetable or a verified funded reserve.^[MOM2] An investment exit must leave continuing obligations assigned and financed.

Recomputing the carbon evidence

The July 2026 reforestation issuance is distinct from Mombak's April enhanced-weathering issuance.^[ISO2][ISO4] The following outputs cover 31 March 2023 to 23 October 2025, although the hosting GHG entry headline shows only the 2025 vintage.^[DATA][ISO1]

First five rows: public simulation and removal summaries [DATA]. Final three: registry and verified GHG statement [ISO1, MOM1, SCS1]. Stage-specific values; supplier allocation is not sales revenue.

The mean stock difference reproduces the exported gross mean. Moving to the gross 16th percentile subtracts 236.299 tonnes, or 1.051%. The benchmark then reduces the conservative estimate by a further 225.351 tonnes. Together, those steps reduce the gross mean by 2.054%. These are accounting adjustments within the modeled historical estimate, not measured percentage reductions in investment risk.

Separating reconciliation from certification

The final allocation reconciles exactly: 2,286.012+19,485.518=21,771.530. The buffer is approximately 10.5%; it is an allocation to pooled protection, not a calibrated reversal probability.

One small upstream difference remains. Subtracting the GHG statement's displayed emissions of 182.89 and 63.01 tonnes from the exported 22,017.528 gives 21,771.628 tonnes, approximately 0.098 tonnes above the verified total. Two-decimal rounding of those two emissions figures alone cannot explain the difference. The complete precise bridge was not reproduced; the verified final total is retained.^[MOM1][DATA]

Quantifying uncertainty and model limits

The GHG statement describes 1,000 Monte Carlo draws for carbon quantification uncertainty. The public regression-diagnostics export contains 1,000 simulation rows; each reports 129 model observations. SCS states that it independently replicated uncertainty distributions and key calculation steps.^{[MOM1][DATA][SCS1]}

A separate public export contains 10,000 simulated counterfactual fractions. Andi recalculated its mean and percentiles from every row, using linear interpolation between ordered values:

Matches performance_benchmark_summary.csv to numerical precision. Draws represent simulated uncertainty, not 10,000 independent field observations. [DATA]

The gross-removal distribution has a central range of 22,242.880 to 22,720.429 tonnes; after the benchmark, it is 22,017.528 to 22,505.200 tonnes. These span the 16th to 84th percentiles, the central 68% of each supplied modeled distribution. They are not 95% confidence intervals, future delivery forecasts, or probabilities that investment proceeds will be received.^[DATA]

Checking independence and prediction error

From 129 public prediction rows, Andi reproduces bias of -0.407 tCO₂e/ha and RMSE of 5.368 tCO₂e/ha for above-ground biomass carbon. The published 95.21% MAPE reproduces when eight zero-observed rows are excluded. That percentage is sensitive to small denominators and does not measure aggregate credit error.^[DATA]

The PDD uses leave-one-out cross-validation to select a model. Its separate equivalence check did not pass both required one-sided tests: the reported p-values were 0.064 and 0.007 against a 0.05 threshold. It then used a regression fallback.^[MOM2] SCS records that Isometric permitted an exception concerning the independent test dataset.^[SCS1] Arithmetic reproduction therefore supports internal consistency; it cannot establish independent generalisation or eliminate model misspecification.

Identifying the underlying risks

Each row identifies an exposure, its measurement route and the evidence needed to move from a conditional scenario to a project estimate. “Open'' means this review has not established the required evidence; it does not allege a defect.

Andi risk hypotheses informed by ISO1–ISO3 and IC1. Public protocol and certification evidence establish an assessment framework; they do not close land-title, consent, cash-flow or long-run outcome diligence.

Shared drivers matter. A drought can simultaneously reduce growth, increase fire risk, raise operating expenditure and postpone issuance. The financial model must capture that combined state once, rather than add separate expected losses for overlapping consequences.

Identifying the financing risks

Andi analytical coverage, informed by the disclosed transaction (WB1–WB4) and registry evidence (ISO1). No finding of default, fraud, rights violation or regulatory non-compliance is made.

Preserving the uncertainty

These risks are quantified through observed balances, contractual sensitivities, conditional cash-flow tests or explicit unresolved conditions. No calibrated basis was established for forward probabilities of fire, default, non-issuance or incremental treatment success. A numerical score would not supply that missing evidence.

Testing the bond's residual exposure

The bond offers an observable example of protection at one layer alongside retained exposure at another. The following conditional calculation uses the published full-disbursement coupon schedule. It assumes all 11 disbursements occur, no catch-up or excess-collection uplift, and full intermediary receipts unless the failure branch is selected. It describes annual payments per $100,000 face value, not annualized investment yield.^[WB4]

Dollar payments. The last three columns include $1,745 fixed annual interest; the variable amount is conditional. Last row applies in each year 2030–2033. Inputs: WB4, final terms PT-46 and PT-49. Recomputed in the report model.

What the contract does and does not protect

For 2030, full realization produces $5,476 variable interest; half produces $2,738. If the corresponding intermediary receipt fails, IBRD need not pay the carbon-linked amount. Deferred shortfalls earn no interest and remaining amounts are cancelled at maturity.^[WB4] The principal obligation is a different exposure.

This is why better forest performance and improved credit accounting cannot alone eliminate financial risk. The required protection must reach the collection account and contractual recipient. Conversely, paying for operational treatments cannot create value for a bondholder beyond the applicable payout rules. A project-level improvement and a bond-level improvement are not interchangeable.

Keeping expected payments separate from observations

The first carbon-linked coupon is scheduled for July 2028, beyond this report's cutoff.^[WB2] Its absence in September 2026 is not a default signal. Earlier fixed interest is a separate obligation.

Actual draw completion, eligible Microsoft deliveries and intermediary collections are not established here. The early deliveries discussed on page 9 cannot be assigned to the bond without certificate, contract and receipt matching. The 2024 documents anticipated standards including Verra; reconcile later Isometric certificates with buyer acceptance and governing confirmations.^[WB3][ISO1] This table neither values the bond today nor estimates project equity returns.

Assessing derisking evidence

Public evidence supports several steps along the proposed causal chain. It does not identify which intervention produced how much of the observed improvement.

Evidence: [DATA, ISO1, MOM2, MOM3, BUY1, WB2, WB4]. Status judgments are Andi's assessment of the evidence reviewed, not a new assurance opinion.

Interpreting the early deliveries

Mombak reports deliveries to Google, McKinsey and McLaren Racing more than two years ahead of expectations. Climeworks separately acknowledges receiving early credits.^[MOM3][BUY1] Both are transaction participants; buyer quantities, collected prices and bond allocation remain unverified.

The first reporting period ended on 23 October 2025; the SCS opinion followed 245 days later and issuance another six days later, a 251-day interval.^[SCS1][ISO1] This measures elapsed time, not avoidable processing delay. Conservative initial schedules, favourable conditions or cohort selection could explain early delivery; the evidence does not isolate a causal MRV effect.

Designing additional derisking

The proposed treatments follow the scoped risk register and consequence assessment. The public estimates above provide a baseline; unresolved risks remain conditions for commitment. The assessment returns to risk identification whenever diligence reveals a new dependency or a treatment creates a new exposure.

Evaluating treatment alternatives

Evaluate avoidance, reduction, transfer and retention separately. A guarantee changes who bears loss; a reserve funds loss; an information exercise reduces uncertainty; none necessarily prevents the physical event. Require a quote and a defined beneficiary before crediting insurance or a guarantee with financial value.

The normalized tests on pages 11–12 examine only MRV readiness and field resilience. Other treatments remain conditional proposals pending scope, cost and enforceability evidence. Any financing improvement must also preserve credit eligibility, including additionality reassessment where required by the applied Isometric Reforestation v1.1 protocol.^[ISO3]

Comparing treatment packages

To compare treatment packages without assuming Mombak's undisclosed finances, normalize a prospective investor's planned net receipt to $1 million in year 4. This is cash after existing operating and contractual obligations, not gross credit revenue. Use an illustrative 8% discount rate, an 80% receipt realization and a two-year delay: the unchanged-course present value is $504,136.

Assume MRV readiness brings collection forward one year at a $25,000 cost. Assume field resilience recovers 10 percentage points of the planned net receipt at a $50,000 cost. Combining them costs $75,000. These are hypothetical effects and budgets. The empirical evidence on pages 4–5 and 9 does not calibrate them.

USD thousands. Same starting exposure, horizon and discount rate in every row. “Combined fails'' pays the full cost and achieves neither assumed improvement. Results: report model, unweighted conditional scenarios.

Present value after incremental cost, USD thousands.

The combined package adds $33,389 in this state, yet leaves a $197,505 gap to the untreated full, on-time receipt benchmark. The gap includes remaining volume and delay effects plus the cost of treatment. Derisking improves this conditional outcome; it does not restore the original proposition.

Testing conditions and treatment failure

The combined package is sensitive both to the starting exposure and to whether its assumed effects occur. These are unweighted conditional scenarios, not samples from a probability distribution.

Varying the untreated starting conditions

Change in PV, USD thousands, after the $75,000 combined package. Its nominal effects remain ten receipt percentage points and one year of acceleration, capped at full, on-time receipts. These nine states do not supply a success probability.

Paying for the full package when an effect fails

Worked state: 80% receipt realisation, two-year starting delay, 8% discount rate. USD, rounded. Same initial exposure as page 11.

Purchasing MRV alone costs $25,000 and adds $15,331; paying $75,000 but obtaining only its timing effect loses $34,669. The $50,000 difference is the extra cost paid. Likewise, field-only purchase adds $13,017 at a $50,000 cost; that effect alone loses $11,983 at the full $75,000 cost. The two comparisons therefore use consistent benefits and different cost commitments.

With both effects, the maximum total upfront cost is $108,389; a further $25,000 implementation burden reduces the net gain to $8,389. At a $75,000 cost, break-even needs a 5.094-point receipt gain with one-year acceleration, or 11.902 points without acceleration. Actual effectiveness and cost evidence must support these thresholds before they inform commitment.

Assigning responsibility for investment decisions

The following governance plan assigns responsibility for an investor's diligence and any subsequent commitment. Its evidence requirements and responses are proposals, not observed Mombak practices.

Converting the case into an investment decision

Before an investment decision, require the governing project confirmations, permitted standards, actual disbursements, delivery targets, ownership and payment records, reconciled accounts and funded stewardship plan. This public-only research has not obtained the complete set. Establish the investor's hurdle, liquidity limit and loss capacity before ranking alternatives.

A rights or eligibility failure can make declining the exposure rational even when a spreadsheet shows positive value. An attractive conditional treatment result can justify obtaining a quote or running a bounded test. It cannot by itself authorize an investment.

Assessing effectiveness and future evidence

The integrated assessment supports partial empirical validation. Selected historical accounting outputs reproduce and delivery is corroborated. The causal risk reduction and net investor benefit of additional treatments remain unresolved. This is an evidence limit, not a finding that those treatments fail.

The assessment connects public evidence to potential causes of loss, financial consequences and proposed treatments. It compares incremental costs with conditional benefits, retains residual exposure and identifies the observations needed to inform an investment decision.

Learning within the public-evidence constraint

The next public update can record new monitoring and issuance events, retain forecasts before outcomes arrive, and test performance against later cohorts. Treatment evaluation additionally needs assignment and timing, comparable baseline conditions, all-in cost and net cash attributable to the same investor. An untouched holdout or later independent cohort would strengthen generalisation testing.

This exploratory assessment is informed by evidence already seen. It does not demonstrate superior investment performance from the analytical approach. Conditional treatment comparisons are possible, but empirical evidence of incremental effectiveness remains insufficient. A useful scenario result and a demonstrated investment benefit require different evidence.

Appendix: reproducing the financial tests

Inputs and interpretation

The normalized model assumes a $1 million net receipt scheduled in year 4 from a hypothetical future commitment; an 8% discount rate; receipt realization of 60%, 80% or 100%; and delay of zero, one or two years. No probability weights are assigned. Dollar values share a constant purchasing-power basis, with no additional inflation, tax or financing layer introduced.

The receipt is defined after all existing claims, costs, reserves and stewardship provisions. This definition is necessary for comparison; its level and sufficiency are not established for Mombak. An actual valuation must model annual BRL and dollar cash flows, taxes, financing, existing pledges, delivery obligations and continuing stewardship explicitly.

Equations

Let R be the planned net receipt, f the realized fraction, d the delay, r the discount rate, g the recovery in receipt fraction, a the acceleration in years and C total upfront incremental cost:

MRV readiness uses (g,a,C)=(0,1,25,000); field resilience uses (0.1,0,50,000); combined uses (0.1,1,75,000). Full-cost failure cases retain C=75,000 while setting either or both effects to zero. No benefit is allowed above full realization or before the original scheduled receipt.

Limits on transfer to Mombak

The recovery parameter represents improvement in investor net receipts, not a directly assumed percentage improvement in tree survival or gross tonnes. Field and MRV evidence must first be translated through applicable credit accounting and contracts. The shared discount rate avoids counting the same risk reduction again through a lower hurdle rate. These are scenario present values, not Mombak's project NPV or an expected return. In the worked state, the timing-only gross benefit is $40,331 and the volume-only gross benefit $63,017; accelerating recovered volume adds an interaction.

Appendix: interpreting the empirical data

The carbon additionality benchmark compares restoration with continued pasture. A treatment-effect comparison instead asks what the same restoration activity would achieve with and without a defined extra safeguard. The biomass-model equivalence check on page 5 addresses a third question: agreement between prediction and inventory. None of these comparisons can substitute for another.

Nine exports from official calculation run mlr_1KPTWWH361S0QX4M; source DATA. Mean and percentile removal summaries are exported outputs; they were not reconstructed from raw joint carbon draws.

The measurement IDs are unique and have no blank cells. The file flags 4,379 records inside the area and 461 outside. No unsupported filter was imposed. The inventory and prediction sets differ; their unmatched IDs remain explicit. T0 and T2 labels do not overlap, and the PDD states that nonpermanent, randomized plot locations prevent direct repeated-measure growth analysis.^[DATA][MOM2]

The raw exports lack coordinates and incremental treatment assignments, costs and receipts. Control-parcel matching, spatial inputs and the complete emissions bridge were not reproduced. Publicly printed financial links failed to resolve as transcribed; that does not establish that their contents are private or absent.^[SCS1]

Maintaining reproducibility and boundaries

Source files, hashes, two numerical models and shared-input checks underpin the calculations. The financial model has 53 arithmetic/boundary checks and the empirical reanalysis 29 checks; these counts are not independent empirical tests. Consistency checks reconcile inputs, allocations and full-cost versus individual-package comparisons.

The PDD development sensitivity table is distinct from the final official export. Andi uses the official export for the historical totals and does not mix parameter labels or uncertainty discounts across versions.^[MOM2][SCS1] Selected calculations are reproducible, while a complete spatial/model rerun and independent human domain review remain outside this assessment.