C-RISK Methodology

C-RISK measures how ready a country is for a conservation debt-for-nature swap. The swap is the load test rather than the market: it is the one instrument that puts weight on every joint at once, the sovereign balance sheet, the movement of capital, the quality of the asset, execution on the ground, and precedent, which is why it is the yardstick. The readiness score is a weighted mean of five slices, fiscal capacity, the deal environment, the conservation asset, delivery capacity, and traction, renormalized over the slices that carry data. Two gates, active armed conflict and sovereign distress, then cap the readiness call. Underneath, each country is still scored 1 to 5 on the dimensions that feed those slices, the gates, and context: conflict, governance, capital flow freedom, sovereign fiscal capacity, donor environment, market policy, protected-area legal durability, tenure, and operator-government relations. Governance is read as context and carries no weight in the headline.

A high score is not a green light and a low one is not a wall. The score names the binding constraint, so what it tells a reader is what would have to be structured around, not whether to go.

C-RISK is designed for a specific audience: the funder, family office, DFI, or coalition deciding whether to deploy conservation capital in a given country. The rating agencies rate sovereigns for bond investors. C-RISK rates countries for the question funders actually face - can capital flow in, get deployed, and earn back over a 25-year horizon. The two views overlap on sovereign fiscal capacity but diverge at everything else.

Rating the hardest instrument produces a read that carries down to the easier ones. A country that could carry a swap can carry a carbon project, a concession, a payment-for-ecosystem-services scheme, or a ten-year grant portfolio. A country scoring 2 will make each of those slower and more expensive in the same specific way, and the binding constraint the score names is usually the same constraint those instruments hit. The score is therefore usable by readers who will never arrange a swap, provided the slice is read for what it actually measures. Take the slice that matches the exposure.

• Fiscal capacity. Sovereign stress reaches conservation before it reaches the headlines: protected-area budgets cut, ranger payroll slipping, pressure to license extraction inside gazetted land, and currency and repatriation friction on committed money. Read it as early warning wherever the work leans on a government counterparty, a co-funding line, or a state agency.
• Deal environment. Whether capital can move to conservation at all, and whether official-sector cover exists. This is the read for carbon developers and for anyone whose funding has to cross a border and come back out.
• Conservation asset. Whether the parks are managed, scored from keystone protected-area performance rather than from designation on a map. This is the read on the estate being backed, bought from, or partnered into.
• Delivery capacity. Whether anyone in the country can execute at the scale under discussion. This is the grantmaker's read: whether money converts into work on the ground or sits.
• Traction. Whether comparable capital has closed already, and what other parties had to structure around to get it done.

The two gates carry across in the same way. Active armed conflict and sovereign distress cap the readiness call whatever the weighted number says, and for a non-swap reader they are the continuity and insurability question rather than a verdict on the country.

The readiness score is the weighted mean of five slices, renormalized over the slices that carry data, so a slice with no data abstains and is left out of the mean rather than scored zero. The weights put the fiscal call first, because for a debt-for-nature swap the binding constraint is almost always whether the sovereign can carry or restructure the instrument. The country score published on the country pages is this readiness score times 20 and nothing else, with no second composite and no other component, so the two are one model on two scales and move together by construction.

Two gates sit on top of the score. A gate caps the readiness call regardless of the number, and only ever lowers it, so the point of view is not averaged away. Cutoffs were locked against the 54-country cohort in August 2026.

The numeric score maps to one of four calls, which a gate can then cap. The binding constraint shown as the headline reason is the lowest-scoring populated slice, or the highest-priority tripped gate.

The dimensions behind the slices split into two sourcing tiers, carried through from v1: a data-driven tier auto-scored from structured public data (conflict, governance, donor environment, and the sovereign fiscal sub-scores), and an editorial tier that synthesises published evidence (capital flow freedom, market policy, protected-area legal durability, tenure, and operator relations). Country pages show the tier a score came from, so the mix of data and judgement behind a slice is visible.

The country pages publish the readiness score on a wider scale: country score = readiness × 20, and nothing else. Because each of the five slices floors at 1.00, the attainable range is 20.0 to 100.0, not 0 to 100. The number is the audit form and the four-tier band is the citation form. Either can be used in isolation; both are always published together. The tiers below are the readiness calls above carried onto that scale, so a country's tier and its call are one statement rather than two.

The tier thresholds are absolute, not percentile-based. A country's tier does not change because other countries' scores changed. This is deliberate. A percentile-based system would mean a country could drop a tier because another country improved, which is nonsensical for an allocation decision. We report continental percentile alongside the absolute tier for context only.

Where the edges come from. All three edges are the readiness call bands carried onto the 20-100 scale. Since v2.0, the country score is the readiness score times 20, so the country-score tiers and the readiness call published above band the identical number and their edges belong in the same place. The call bands sit at readiness 4.0, 3.0 and 2.0, which on this scale are 80, 60 and 40. They were validated and confirmed unchanged in a separate review that explicitly accepted an empty top band, on the reasoning that lowering an edge to populate a band over-claims. The tiers did not start there. All three were originally cut against the retired four-component composite, at 65.0, 50.0 and 35.0, and they were corrected in two steps: Tier 1 moved to 80.0 in v2.1, and Tier 2 and Tier 3 followed to 60.0 and 40.0 in v2.2. The inherited Tier 1 edge did exactly what that review refused: it put nine countries in the top tier, from Rwanda at 76.2 down to Botswana at 67.8, whose own readiness call is "Ready with structuring" and none of which reach 4.0. The inherited Tier 2 and Tier 3 edges did a milder version of the same thing to 24 more. Tier 1 is now empty on both surfaces, and every country's tier agrees with its call. That is the honest consequence of a position already taken, not a new claim about Africa.

Tier 3 holds 33 of 54 countries, and we have left that concentration in place on purpose. The concentration is real. It sits in Tier 3 rather than Tier 2 only because the edges moved to meet the readiness calls, not because any country moved. We checked whether it is an artefact of abstention, since renormalising over populated slices lifts thin-data countries that the old composite punished. It is not. Restricted to the 44 countries that carry all five slices the concentration is worse, not better, and the ten abstaining countries are the most dispersed in the cohort, sitting at both ends rather than in the middle. The cause is arithmetic: averaging five slices each scored 1 to 5 costs about 45 percent of the dispersion a single slice carries, and 17 of those 44 countries hold both a slice at 4.0 or above and a slice at 2.5 or below, which cancel inside the mean. Re-cutting the edges to spread the cohort evenly across the tiers would fit a shape that was a property of the retired blend, which contained a log-scaled count that could reach zero, and not a finding about the continent. We have not done that, and the edges we do use are not a compromise with it: they are the readiness call edges, and the distribution they produce is bottom-heavy rather than flattened, which is the opposite of what fitting a distribution would yield. We are not moving edges to flatter a distribution. For discrimination inside a tier, read the number, the binding constraint, and the count of populated slices, all of which are published per country.

This breaks comparability with stored tiers. The monthly snapshots carry the tier label as it was computed at the time, and snapshots back to 1 May 2026 carry labels cut at 65.0. A stored "Tier 1" from before this change is not the same claim as a "Tier 1" after it. The underlying country-score number in those snapshots is untouched and remains directly comparable.

Two gates sit outside the weighted mean and can cap the tier a country lands in. They are the readiness model's gates, applied to the country score so that the tier and the readiness call cannot tell different stories about the same country.

Three properties of the caps, each deliberate.

They move the tier, never the number. A capped country's score is unchanged, and so is its rank. Only the published tier changes. A reader who wants the uncapped reading has it: the pre-cap tier is published alongside the capped one.

They only ever lower. A gate cannot lift a country into a better tier, and where a country is already at or below the cap the gate does nothing.

The asymmetry is intentional. Active conflict caps harder than sovereign distress because a war is not a structuring problem. Distress is: it caps a country out of the top band without pretending the country is undeployable. The same two gates cap the readiness call identically, because since v2.3 the tier and the call are the same four bands under the same names rather than two vocabularies that had to be mapped onto each other. Note that while Tier 1 is empty the sovereign-distress cap cannot bind, since no country sits above it. It moved no country before the v2.1 edge change either. It is left in place so that it re-arms on its own should any country clear 80.0.

Where a portfolio is scored in aggregate rather than country by country, the tier is read from the score alone. A grant-weighted portfolio spans many countries and therefore has no single gate set to apply, so the caps are a per-country step and not part of the band function itself.

For press, pitch decks, and third-party reference, use any of:

Full form: "Kenya, C-RISK 70.0 (Tier 2, Ready with structuring)"

Short form: "Kenya, C-RISK 70.0"

Band-only: "Kenya is a Tier 2 country on the C-RISK country score"

Canopy requests attribution on any public reproduction of the country score or its components. See the corrections policy.

The slice definitions, sources and weights are set out above. What follows is the part that decides what a country's number rests on: how each slice is reduced to one figure, and what happens when it cannot be.

The nine dimensions still exist and are still scored 1 to 5. They no longer combine into a single weighted composite; instead each one feeds a slice, a gate, or context, as the "Feeds" column shows. Sovereign fiscal capacity leads the fiscal slice, conflict becomes a gate, and governance and the site-durability dimensions move to context.

The Governance baseline dimension is editorially scored 1 to 5 against global percentile context plus a country-level institutional read. As of v1.3 (May 2026), every country also carries a structured quantitative backbone alongside that editorial score, as reference data rather than a replacement. A reader auditing the 1-to-5 read can compare it directly against the underlying global governance percentile and an independent corruption-perception triangulation without re-deriving anything.

The quantitative backbone does not feed the editorial governance score directly. The 1-to-5 read is intentional editorial judgement against percentile context plus a country-level institutional read; the backbone is the public global data made structured and citable. Where the editorial read and the computed percentile diverge materially, the country is flagged for editorial review at the next refresh. As of the May 2026 build, two countries flag this way: Cabo Verde (top editorial marks against a mid-range percentile, where the editorial scale is Africa-relative rather than globally-anchored at the top end) and Zimbabwe (a low editorial score against an even lower percentile, where the editorial read may be generous on reform talk).

Both fetchers run inside the monthly snapshot job on the first of each month. Substantive updates only occur once or twice per year (the governance indicators refresh in autumn, the corruption index early in the year); intervening months produce no diff. Fetched values are stored in structured form and integrated into each country's record, with the source-year tags carried for traceability, and the backbone is captured in each monthly snapshot.

The backbone draws on established, publicly available cross-country governance and corruption-perception indices. The specific series, access endpoints, and the percentile computation are shared with funders and partners under diligence.

Fiscal capacity is the 0.30 lead of the readiness score, and it is also Gate 2. It reads whether the sovereign can carry or restructure a debt-for-nature swap: a country in or near sovereign distress cannot issue one, and distress transmits to parks budgets, which are typically first cut when sovereigns are squeezed. The slice is served from the Sovereign Fiscal Lens, the gated standalone page backed by data/sovereign.json, so C-CARBON and the C-FIN country brief read the same fiscal record. The rest of this section is how that lens is scored.

The fiscal lens is computed from four scored sub-scores, each rated 1 to 5: rating and spread carry the market view, IMF backstop carries the official-sector distress reading, and fundamentals produces an independent credit view from primary fiscal data. They combine as a weighted mean, not a simple average, with the market and official-sector signals carrying the most weight (rating 0.30, IMF backstop 0.30, spread 0.25, fundamentals 0.15), re-weighted over whichever sub-scores are populated. A fifth reading, multilateral pipeline, is kept alongside as context and is not part of the score: it measures the catalytic capital a conservation instrument could ride, a deal-flow signal rather than a measure of default risk.

The major agencies' current ratings map to the 1-to-5 scale in fixed bands: investment grade at the top, stepping down through the speculative bands to default or active restructuring at the floor. The exact band cutoffs are shared with funders and partners under diligence.

The sovereign's Eurobond spread over the US 10-year Treasury maps to the scale in fixed bands: the tighter the spread the higher the score, down to a floor for distressed levels or where no liquid quote exists. The exact spread cutoffs are shared under diligence.

Each of the seven components is scored 1 to 5 against thresholds calibrated to the observed distribution of African sovereigns rather than to generic global cutoffs, so a mid-range score reflects a typical African fiscal position rather than a globally benign one. The components run from the public debt burden and the cost of servicing it, through the primary balance, inflation, and reserve cover, to the share of debt denominated in hard currency. Lower debt, cheaper servicing, stronger balances, tamer inflation, deeper reserves, and a smaller hard-currency debt share all score better. The exact per-component cutoffs, and the recent sovereign-distress cases they are calibrated against, are shared with funders and partners under diligence.

FX-denominated debt share - what it captures. Currency composition is a structural fragility metric. A country with 30 percent of its public debt in hard currency and 70 percent in local currency takes a 30 percent currency depreciation as a 9-percent-of-debt-stock revaluation; a country with 80 percent of its debt in hard currency takes the same depreciation as a 24-percent-of-debt-stock revaluation. The first absorbs; the second often does not. The component sits in fundamentals because the debt structure is durable - it does not move with monthly market noise the way spreads do, and it is not directly captured by the rating agencies' headline letter.

CFA franc treatment. Domestic debt denominated in CFA franc by WAEMU and CEMAC sovereigns is treated as domestic for this component. The CFA peg to the euro has held since 1948, the regional central banks (BCEAO and BEAC) provide effective monetary backing, and CFA-denominated debt is not subject to the same idiosyncratic depreciation risk as a free-floating local currency. The simplification applies uniformly across the 14 CFA-zone sovereigns and is disclosed on each affected country page where relevant. A future v1.3 refinement may downgrade the CFA-as-domestic treatment if regional monetary stress indicators move; the current treatment reflects the May 2026 baseline.

The distress-prevention lens. Reads programme status alongside the IMF debt-sustainability assessment. Multilateral pipeline below uses programme status alone, without the debt-sustainability reading, as a deal-flow lens.

Programme status as a catalytic instrument. An active IMF programme creates the macro framework that other multilateral lenders condition disbursements on. The debt-sustainability reading is excluded here; it sits in the IMF backstop sub-score.

The dimension composite is the weighted mean of the four scored sub-scores (rating 0.30, IMF backstop 0.30, spread 0.25, fundamentals 0.15), re-weighted over whichever are populated, rounded to one decimal. Multilateral pipeline is shown but excluded from the composite, and fundamentals is used only where enough of its components are present. Where a sub-score is not computed - typically because no liquid Eurobond exists or no recent IMF consultation is available - the composite is computed on the remaining scored sub-scores and a "computed on N of 4 scored sub-scores" note is exposed at the country level. Each country also carries a provenance flag: market-validated where a rating or spread is present, IMF/DSA-anchored where neither is.

Where a country has no market signal at all - unrated by every major agency and with no liquid Eurobond spread, which applies to a handful including Eritrea, Somalia, and South Sudan - the score is anchored to the IMF backstop distress reading and capped at half a point above it, so a sovereign the markets will not price cannot read as mid-tier on the strength of the softer sub-scores alone. Those records carry an IMF/DSA-anchored provenance flag. Where neither a market signal nor an IMF reading exists, the dimension abstains rather than inventing a number.

Sub-scores are stored discretely so that a future twin-rating framework - a Political Risk grade alongside a Sovereign Default grade per country - can be surfaced as a render change rather than a rescoring exercise. C-CARBON's sovereign risk dimension reads directly from this C-RISK dimension as of C-CARBON v1.1 (May 2026), as a 50/50 average with the governance dimension above.

Each dimension is rated 1 to 5 against the published rubric below, where 5 is the most favourable posture and 1 the most adverse. Those dimension scores feed the five slices, the two gates, and context. The readiness score is the weighted mean of the populated slices, renormalized so a slice with no data is left out rather than scored zero, the same null-handling the dimensions have always used. The four-band readiness call above is the published headline; the reading below interprets the underlying 1-to-5 number.

A composite that double-counts is a composite that lies. The readiness score draws on five slices, and we tested whether they measure overlapping signals. They do not, and two of the pairings run in directions worth explaining.

Pairwise Pearson correlations across the current 54-country cohort, computed on the published slice scores. Pairs are complete-case: where either slice abstains the country is left out of that pair, so n varies from 44 to 54.

The strongest pairing is deal environment and conservation asset at 0.62. Expected, and the one to watch. Both are downstream of state capacity: a government that can run a functioning capital account and a carbon-market framework tends also to be a government whose protected areas are funded and staffed. It is the highest correlation in the matrix and the closest the model comes to double-counting. It stays below the level at which the two would be measuring one thing: a country can run good parks behind a closed capital account, and several do.

This answers the question for the country score too. Multiplying by 20 is a linear transform, and a linear transform cannot alter a correlation, so the slice structure of the published country score is the slice structure of the readiness score exactly.

Fiscal capacity and delivery capacity run negative, at -0.23. Weaker sovereigns host slightly more operator capacity, not less. The largest managed estates on the continent sit in exactly the places where state capacity is thinnest and the operating model exists, which is what the delivery slice is built to detect. Delivery is not a proxy for country quality; it sometimes pulls against the fiscal reading rather than with it, and a country that scores well on both is a genuinely different proposition from one that scores well on either.

The readiness score weights five slices: fiscal 0.30, deal environment 0.20, conservation asset 0.20, delivery 0.15, and traction 0.15. The weights are a stated editorial choice, not a fitted parameter, so the question is whether the order a reader sees depends on their precise values. It does not.

Each of the five slice weights is shifted by plus or minus 5 percentage points, the other four renormalised proportionally to keep the set at 1.0, and readiness recomputed for all 54 countries. Across the ten perturbations the minimum Spearman rank correlation against baseline is ρ 0.977 (mean 0.991), and at most 5 of the 54 countries cross a tier boundary under any single shift.

Delivery capacity is the most sensitive weight, and for a legible reason: it is the widest-dispersed slice on the cohort, so moving its weight moves the middle of the table more than the tighter slices do. The two slices whose weight barely matters, deal environment either way, are the ones whose values cluster most closely. None of this makes the weights arbitrary; it means a reader who prefers a different split will read the same country order.

One consequence is specific to the published 20-100 country scale: because the tier edges are absolute rather than percentile-based, a weight change moves a country across a tier boundary only by moving its score past 80, 60 or 40. The tier-mover counts above are counts against those edges, and they are recomputed when the edges move.

The per-country detail recomputes with the data, and the live cell sits in the C-VAL validation matrix, alongside the same test for every other Canopy model on one consistent basis.

A point-in-time score tells you where a country is, not where it is going. C-RISK pairs every score with a forward-looking outlook on the sovereign convention - positive, stable, negative, or under review - displayed as an arrow next to the score.

Outlook is computed deterministically from per-dimension trajectory tags. At every scoring pass, each dimension is assigned a trajectory of +1 (improving), 0 (stable), or -1 (deteriorating), driven by published triggers:

Trajectory tags are summed across the dimensions, then mapped to a label from positive through stable to negative. An election within six months in any country whose readiness call is Selective or lower, an active coup, a protected-area degazettement, or active IMF-programme uncertainty overrides to under review.

The discipline: every outlook label is reproducible from the trajectory tags. Two analysts working independently from the same triggers reach the same outlook. That is the sovereign convention.

The 1-to-5 scale published for every dimension. Anyone disagreeing with a country score has a fixed scale to push against.

Composite scores live and die by how they handle missing data. Every country-score component has an explicit missing-data rule, published above and restated here for the reader who wants to know the failure modes before trusting the number.

Abstention, as of v2.0. A country with no rated PA abstains on the conservation-asset slice, and on delivery capacity with it, since delivery reads the operators running those same PAs. The readiness mean renormalises over the slices that carry data. Nine of the 54 countries abstain on both today.

This replaces the v1.2 rule, which imputed the continental P25 (approximately 2.4) so that every country carried a PACE number. Imputation was chosen then over zero-fill, which would have penalised low-coverage countries unfairly, and over the continental median, which produced a perverse incentive where a country with zero rated PAs scored higher than a country with one weakly-rated PA. Abstention is better than either: it puts no manufactured value inside the composite and states plainly that the question has no answer yet for that country. The trade is that a country scored on three slices carries more uncertainty than one scored on five, which the coverage layer discloses.

The P25 imputation still runs in the four-component sub-signal published alongside the score, so the disclosure a reader saw before is intact: an asterisk next to the score in the country table, and a "PACE imputed" pill next to the headline score on the country dossier.

The count of registered carbon projects no longer enters the composite. It was the DEPLOYMENT component until v2.0 and is now published as part of the four-component sub-signal only, where it is still log-scaled so that a country with one tracked project scores meaningfully above zero without claiming parity with a country that has twenty. What a country has actually closed now reaches the score through traction, which reads nature-finance deals and the funding gap rather than carbon-registry listings.

Seychelles, Sao Tome and Principe, Comoros, Cabo Verde, and Mauritius have limited PA cohorts, limited project pipelines, and small absolute capital-deployment footprints. Their country scores should be read with the coverage pip prominently in mind. A Tier 2 Seychelles and a Tier 2 South Africa are both defensible but for different reasons.

The country score applies a conflict cap, as of v2.0. Where the conflict dimension reads at the floor, the country is capped at Tier 3, Selective, early, whatever its score. Seven countries trip that gate; on the current cohort it binds on one, moving the Democratic Republic of Congo down a tier. Somalia and South Sudan trip it too but already reach the capped tier on their own score, so it changes nothing for them. The gate caps above set out the mechanics.

This reverses the position published under v1.x, which was that the weighted formula already produced the right answer for these countries and that a separate override would be redundant. Under the readiness model it is not redundant: a country can score respectably on fiscal capacity, deal environment and delivery while an active war makes the transaction undeployable, and DRC at 62.8 is exactly that case. The gate exists because the score alone was placing active-conflict states in the same tier as functioning markets. What we still reject is the blunter form of the same idea, a hard floor keyed to the composite falling below a cut-off, which would sweep in stressed-but-functioning markets that the score handles correctly.

Not every country can populate every Canopy data layer in equal measure. A country with no Tier-1 protected area at the scale Canopy uses, no keystone in the curated keystone list, or zero registered carbon projects has nothing to score, by reality not by editorial omission. Canopy distinguishes justified-absent (the layer is empty because reality says so, verified against protected-area, carbon-registry, and multilateral climate sources) from gap (Canopy has not yet added data that reality supports). The distinction is encoded in a coverage-flags dataset that marks, per country, the absence of a Tier-1 protected area, of a flagship keystone, or of any registry project. Coverage scoring on the country page counts present and justified-absent equally toward the 5/5 scale. This prevents micro-states and structurally-thin markets from being unfairly penalised against the same yardstick as continental majors.

Slices refresh asynchronously. Fiscal capacity refreshes with the sovereign dataset (rating actions and spreads move continuously, fundamentals on IMF publication cycles), the deal environment on C-RISK review cycles (roughly quarterly), the conservation asset on PA-addition cycles (monthly to quarterly), delivery capacity on operator-registry and PACE-cohort changes, and traction on the nature-finance deals tracker (rolling). The country score is recomputed on every slice update and time-stamped. The number on any dossier is the country score as of the last slice refresh for that country, not a static quarterly number.

A methodology page that shows only the chosen formulation invites the reader to imagine that no alternatives were considered. Here are the ones we considered and why they lost.

Any composite lies in some direction. Here are the directions we know the country score lies in, published so practitioners can adjust.

The country score is not the first index that ranks African countries. It is the first purpose-built for conservation capital. The relevant peers and what they miss:

The country score is designed to be the reference index a conservation-finance practitioner reaches for first. When the country score disagrees with those general governance, country-risk, and corruption indices, the disagreement is itself signal: they were built for different decisions.

C-RISK draft scores are generated by a structured large-language-model pipeline consulting the published source list per dimension and producing dimension scores, sub-scores, trajectory tags, an outlook label, summary copy, and a sources-consulted trail per country. The pipeline writes drafts into a separate file; nothing reaches the live site without an explicit promote step.

An editorial review layer sits between drafter and promote. The architecture supports per-country inspection, edit, or rejection of any field - dimension scores, sovereign sub-scores, trajectory tags, outlook label, summary text - before promotion to live data.

The editorial review layer is not staffed: drafts are promoted to live data after automated parse validation, with no human sign-off in between. The architecture is in place; staffing it is a function of platform stage. Funder partners interested in seconding analyst time to C-RISK editorial review should get in touch.

Corrections are welcomed. Operators, funders, or researchers spotting a factual error should email us.

The country score's source trail. Every C-RISK dimension note carries source URLs, as does every PACE dimension note; the sovereign dataset cites its rating actions and market sources, and the deals tracker cites each transaction. A reader disputing a country-score value can trace it back through the slice to the underlying evidence in two or three clicks. Only part of the score is editorial: fiscal capacity is assembled from ratings, spreads, backstop status and fundamentals, delivery capacity is derived deterministically from the canonical operator registry and the PACE cohort, and traction from the deals tracker and the funding-gap dataset. The country score itself is a pure arithmetic step on the readiness score.

Disputes. Any country or operator disputing its country score, any component score, or any source citation can request a formal review. Canopy will publish the dispute, the review process, the decision, and, if warranted, the correction. Canopy does not suppress disputes.

C-RISK scores are refreshed when underlying data changes. Governance indicators refresh annually. Conflict-event data is continuous. Sovereign rating actions and IMF programme reviews trigger immediate review of the affected dimension. Major political events, PA legal changes, donor cycle announcements, and policy shifts all trigger review of affected dimensions. A scoring date is displayed on each country record, alongside the oldest input feeding the composite. No country score is published older than 12 months without re-review.

How the country score changes over time, and under what circumstances.

The country score recomputes on every refresh of any underlying component. No manual trigger. Each country page shows the last-updated timestamp for its country score, accurate to the day.

Weights, band thresholds, and missing-data rules are treated as version-controlled. Any change to these is published with (a) the old value, (b) the new value, (c) the reason for the change, (d) the effective date. Historical country scores under old methodologies are preserved on request.

The current version is v2.3, effective 15 August 2026. Material revisions bump the version number.

v2.3 - 15 August 2026. The tier labels are now the readiness call strings verbatim: Deal-ready, Ready with structuring, Selective, early and Not yet, monitor. They replace Deployment-ready, Deployable with mitigants, Selective deployment and Monitoring-only. No edge, no score, no rank and no gate-cap logic changes, and no country changes tier. This is a rename of the labels alone. Since v2.2 the tier edges and the call edges have been identical, so tier 2 and the call "Ready with structuring" were already guaranteed to denote the same thing; all that remained was that we printed two different sets of words for it. On the country dossier that was visible: the header read "Tier 2, Deployable with mitigants" directly above "Deal-readiness call: Ready with structuring", the same verdict twice in different words. The country pages now state it once, and show the binding constraint in the space that repetition used to occupy. The call words were kept rather than the tier words because they are actionable: "Ready with structuring" tells a reader what to do, "Deployable with mitigants" does not. Tier 1 to 4 survives as the citation form, so a country reads "Tier 2, Ready with structuring". Note that the stored tier labels in snapshots use the old words, on top of the edge-comparability break disclosed at v2.1 and v2.2.

v2.2 - 15 August 2026. Tier 2 and Tier 3 edges recalibrated from 50.0 and 35.0 to 60.0 and 40.0, completing what v2.1 began. No score, no rank, and no gate-cap logic moves. v2.1 moved the Tier 1 edge on the principle that CCS is the readiness score times 20, so a tier edge and a readiness call edge band the same number and belong in the same place; it left the other two edges, both inherited from the retired four-component blend, where they were. The consequence was that 24 of the 54 countries still carried a tier that contradicted their own published call: 22 read "Tier 2, Deployable with mitigants" against a call of "Selective, early", and two read Tier 3 against "Not yet, monitor". Section 9's claim that the country score and the readiness score move together by construction was true of the number but not of the label. It is now true of both, verified across all 54 countries. The distribution restates from 0 / 39 / 13 / 2 to 0 / 17 / 33 / 4 and 24 countries are relabelled downward. This does not reopen what v2.1 settled. v2.1 declined to re-cut these edges in order to spread the cohort across the tiers, because the shape being fitted was a property of a formula no longer in use. That reasoning stands and this is not that: these edges are taken from the call bands rather than from a target distribution, and the result is bottom-heavy rather than flattened. The mean-of-five compression finding is unchanged and still published, because it describes how the score is distributed and says nothing about where the edges belong. The active-conflict gate now binds on one country rather than three, since Somalia and South Sudan reach the capped tier on their own score. The snapshot comparability break disclosed at v2.1 widens accordingly: stored tier labels cut before these changes do not mean what a tier label means now, though the stored CCS numbers are unaffected.

v2.1 - 15 August 2026. Tier 1 edge recalibrated from 65.0 to 80.0. No score, no rank, and no gate cap moves; nine countries are relabelled downward from Tier 1 to Tier 2 and nothing else changes. Since v2.0 CCS has been the readiness score times 20, so the CCS tiers and the readiness call band the same number, and the call bands (readiness 4.0 / 3.0 / 2.0, equal to CCS 80 / 60 / 40) had already been validated with an empty top band on the reasoning that lowering an edge to populate a band over-claims. The inherited Tier 1 edge of 65.0 sat below that line and re-populated, from the same number, the band that review deliberately left empty. Tier 1 is now empty. The Tier 2 and Tier 3 edges are deliberately NOT moved: the Tier 2 concentration was investigated and attributed to mean-of-five compression rather than to abstention (it is worse, not better, among the 44 countries carrying all five slices), and refitting the edges to the retired blend's distribution was rejected as fitting to an artefact of a formula no longer in use. Percentile bands remain rejected. Consequences disclosed: stored tier labels in snapshots back to 1 May 2026 were cut at 65.0 and do not mean the same thing as tiers cut after this change, though the stored CCS numbers are unaffected; grant-weighted portfolio aggregates band through the same function and so move too; and the sovereign-distress gate cap, which caps at Tier 2, is inert while Tier 1 is empty, having moved no country before this change either. The country-score section above carries the reasoning and the numbers.

v2.0 - 15 August 2026. CCS stops being a weighted blend of four components and becomes the C-RISK deal-readiness score on a 0-100 scale: CCS = readiness × 20, where readiness is the weighted mean of five slices (fiscal capacity, deal environment, conservation asset, delivery capacity, traction) renormalised over the slices that carry data. The country score and the readiness score now move together by construction rather than being two separate constructions of the same country.

What this changes. (1) The retired components: the operator-ecosystem count and the deployed-project count are no longer inputs. Operator presence now reaches the score through delivery capacity, which reads managed-estate scale for state agencies, NGOs and trusts alike rather than counting names off a closed international list; what a country has closed reaches the score through traction, which reads nature-finance deals and the funding gap rather than carbon-registry listings. All four v1 components are still computed and still published as a sub-signal on the country pages. (2) Missing PACE: the conservation-asset slice now abstains where a country has no rated PA, instead of imputing the continental P25 into the composite; nine countries abstain today, and the P25 imputation and its "PACE imputed" flag survive in the sub-signal. (3) Gate caps: two gates now cap the tier. Active conflict caps at Tier 3 and sovereign distress at Tier 2, band-only, never touching the score or the rank order. This reverses the v1.x position that CCS applies no conflict override; at v2.0 it bound on three countries, moving DRC, Somalia and South Sudan from Tier 2 to Tier 3; after the v2.2 edge change it binds on DRC alone. (4) Sections 6 and 7 recomputed against the five slices; the v1.x entropy and CRITIC comparison is retired rather than restated, because it described a formula no longer in use.

Band edges are unchanged at 65 / 50 / 35. The v1 series is preserved in the monthly snapshots and is not restated under v2.0.

v1.5 - 17 July 2026. DEPLOYMENT counts live projects only. The component counted every tracked record per country, including projects the registry had withdrawn, rejected, or suspended, so dead activity scored as deployment. It now reads the live count only. 21 of 269 tracked records are terminal; affected countries lose a little DEPLOYMENT (DRC 18→14 live, roughly -0.5 CCS) and thin cohorts more (Zimbabwe's two projects, the withdrawn Kariba among them, both retired → 0 live → deployment 0). The coverage layer still reads all tracked records, so a country whose projects have all withdrawn holds project data (coverage present) and scores 0 on deployment, which is the truthful reading, not a gap. No weight change and no formula change; only the input to the projects count. The matching change on C-CARBON drops the same 21 from that cohort's average.

v1.4 - 17 July 2026. Cross-model country joins now resolve through ISO country codes rather than a slug derived from whichever spelling arrived, with a single canonical country-identity dataset as the authority, so adding a spelling is a data edit, not a code edit. Fixed two silent join failures: Cote d'Ivoire's accented PACE spelling had dropped its two rated PAs (52.7→54.2), and DRC's projects spelling had counted zero of its registry projects, showing a false coverage gap (41.5→47.9). Unresolvable names now hard-fail the snapshot build instead of falling back to a plausible-looking slug.

v1.3 documentation correction - 17 July 2026. No formula change, no weight change, no rule change; no CCS score moves. DEPLOYMENT's Source line in Section 2 described the projects directory as drawn from several public registries. In fact Canopy ingests a single carbon-project registry and nothing else: nearly all tracked projects (267 of 269) come from that one source. DEPLOYMENT has therefore always counted projects registered under that one registry specifically, which is a narrower thing than the count of registered projects in a country, and it is now described that way. Corrected in the same pass: the coverage-flags note in Section 4 (the flag records the absence of projects in the source registry, not the absence of projects altogether); limitation 2 in Section 8, where single-registry ingest is now named as compounding the Francophone under-count rather than sitting beside it; and Section 11's description of the directory as human-curated, which the automated registry pull superseded. The same false multi-registry claim was corrected on the C-CARBON methodology.

v1.3 documentation update - 18 June 2026. Published the v1.3 weight-perturbation robustness result in Section 7, closing the item flagged in the 14 May 2026 entry. Each of the four weights was perturbed plus or minus 5 percentage points with the remaining three renormalised, and CCS recomputed for all 54 ranked countries: minimum Spearman ρ 0.99 (mean 0.994), at most 7 of 54 countries crossing a tier boundary, OPERATOR the most sensitive input. Per-country detail in the published validation record; the C-VAL robustness cell was refreshed to the current data. No formula change, no weight change, no rule change.

v1.3 - 14 May 2026. Two changes. (1) Component reweight: more weight to C-RISK, less to DEPLOYMENT, with PACE and OPERATOR unchanged. Rationale: aligns the CCS weighting with the sovereign-first methodology lineage that anchors C-RISK. The Spearman rank correlation between CCS and C-RISK rises from ρ 0.70 to 0.85; CCS retains meaningful orthogonal information from PACE, OPERATOR, and DEPLOYMENT. Most material rank movements: Madagascar #7 to #21 (low C-RISK score now dominates), Seychelles #14 to #6 (high C-RISK 4.19 lifts), Botswana #10 to #7, Namibia #19 to #9, Mozambique #11 to #22. (2) OPERATOR detection bug fix: an operator-name-matching bug was corrected. A short substring had been matching unrelated words and falsely crediting five countries (Algeria x2, Burkina Faso, Burundi, Cote d'Ivoire, Madagascar) with an international operator presence they do not have; the longer, specific name patterns are retained. The v1.2 doc-update entry (4 May 2026) stated "Chosen weights stand for v1.3" - that expectation was overturned by this reweight, which was driven by editorial alignment rather than by the data-driven weight comparison's recommendations. Section 7 robustness checks remain in place as documentation of the v1.2 weights; the equivalent v1.3 weight-perturbation analysis is now published in Section 7 (minimum Spearman ρ 0.99 across the eight single-weight +/-5pp sets).

v1.2 documentation update - 4 May 2026. Added Section 7 (Robustness checks) reporting sensitivity to weight perturbation, PCA collinearity check, and a data-driven weight comparison via entropy and CRITIC. Sections 7 to 10 renumbered to 8 to 11. The earlier expectation in Section 6 that data-derived weights would come within 5 percentage points of chosen was retracted: CRITIC sits 5 to 18 points off and entropy up to 42 points off. No formula change, no weight change, no rule change. Chosen weights stand for v1.3.

v1.2 - 29 April 2026. Removed double-counting in the composite. Two changes: (1) POLICY component (carbon-market-policy readiness) was removed - carbon-market-policy status is already inside RISK as its market-policy dimension, so POLICY at the CCS level was double-counting. Replaced at the same weight with OPERATOR ECOSYSTEM, a count of distinct major international NGO operators per country, derived from the PACE dataset. OPERATOR is genuinely orthogonal to RISK's operator-government-relations dimension because the latter scores quality of relationships and OPERATOR scores count of operators present. (2) DEPLOYMENT was simplified from a two-sub-score average (projects + donor) to a projects-only signal. The donor sub-score was identical to RISK's donor-environment dimension and was double-counted, once inside RISK and again inside DEPLOYMENT, materially heavier than intended. Also: PACE imputation rule changed from continental median (~2.86) to continental P25 (~2.4) to close a perverse incentive where a country with zero rated PAs scored higher than a country with one weakly-rated PA. Stale "rejected sovereign-credit-rating" entry in Section 5 updated to reflect that sovereign default is now inside RISK as a ninth dimension. New rejected-alternative entry added documenting the v1.2 removal of POLICY. RISK component description updated to reference the May 2026 RISK rebuild (nine dimensions, three categories, sovereign default added). Limitation 4 (POLICY stepwise movement) replaced with operator-list closure caveat. New limitation 5 added documenting that RISK outlook is surfaced alongside CCS but not yet folded into the composite. Operating principle softened from "CCS more likely wrong than your priors" to a more balanced read.

v1.1 - 25 April 2026. Who's Who removed from CCS composite and repositioned as a separate surface (it was already absent from the v1.0 weighted formula but had been counted in earlier coverage signalling). Coverage flag system formalised as a coverage-flags dataset distinguishing justified-absent from gap. PACE cohort expanded from 38 to 128 protected areas across two batches (Sessions 22 and 23), with operator coverage rebalanced away from a 44% African Parks share down to approximately 17%. Per-PA verified flag retired (it was decorative at one verified entry of 98 and not load-bearing under the published methodology).

v1.0 - 23 April 2026. Initial publication.

Past country-score values are not re-stated under new methodologies. A country's "country score movement" time-series will always compare like-for-like within a single methodology version. Cross-version comparisons are disclosed explicitly.

The country score is a derived composite. Its inputs - C-RISK scores, PACE scores, the operator-ecosystem count, and DEPLOYMENT activity - are now each captured under the time-series capture system documented in their respective methodology pages.

The canonical current-state datasets for C-RISK, PACE, and projects hold current state. Immutable monthly snapshots preserve the inputs that fed each historical country-score computation, with cryptographic hashes and a manifest. Because the country score is computed deterministically from these inputs, any past country-score value can be reproduced exactly from the snapshot of its components.

The first snapshot was captured on 1 May 2026. This is the start of the country score's panel view across the 54 African countries.

This captures changes to country composites over time arising from input updates. It does not yet capture re-scoring of underlying inputs themselves, which is on the broader roadmap.

v2.3 - effective September 2026. Registration B re-locked onto a corrected baseline, disclosed rather than quietly re-pointed. The 1 September snapshot that Registration B registered was replaced upstream the same day, because the original capture had banked the previous month's protected-area scores rather than September's. The registered baseline was therefore recording something that was not true of 1 September. The lock now points at the corrected capture, SHA256 0d800dac...b8273a6c, taken 13:27 UTC. Why this is not the goalpost move the validation section forbids. The window opened on 1 September and the stress-event ledger is empty: not one pre-registered event has been recorded against either registration. There was therefore no outcome data in existence that could have made one baseline look better than the other, and the correction was made upstream for reasons that have nothing to do with this validation. A lock is a commitment not to move the baseline once you can see how it is doing; it is not a commitment to keep a baseline that recorded the wrong month. What moved and what did not: the hash changed, twelve of the fifty-four countries carry a different readiness score, and one crosses a band, so the five registered bands restate from 2 / 11 / 36 / 5 / 0 to 2 / 10 / 36 / 6 / 0. The readiness call bands (4 / 32 / 18 / 0), the tercile edges (2.58 and 3.03) and the tercile cells (18 each) are all unchanged, so the descriptive tables and the reasoning for retiring the band threshold stand exactly as published. The drift guard added at v2.1 caught this automatically and refused to run the calibration against a baseline that no longer matched its registered digest, which is what it was built to do.

v2.2 - effective September 2026. Drafter test-retest measured and published in the third limitation. No score, weight, gate or band edge changes, and no country moves. The methodology already disclosed that the editorially-assisted dimensions are LLM-drafted and that the review layer is unstaffed, but carried no number for how much the drafter varies run to run. Fifteen countries, stratified by readiness call, were re-drafted blind twice in the same hour. Two fresh passes rather than one pass against the published record, because comparing a fresh draft to the August scores would confound drafter variance with a month of world change. Across the seven freely-drafted dimensions (n=97 paired scores) exact agreement is 63.9 percent (95% CI 55.0 to 73.7) and Gwet AC1 is 0.568 (0.438 to 0.678), with within-1 agreement at 100 percent and a largest single gap of 1.0. Carried through the deal-environment slice, a re-draft moves the published country score by 1.3 points on average and 4.0 at the worst observed. Three decisions are disclosed rather than buried: governance is excluded from the pooled figure because the payload pins it to a WGI anchor and it reproduces at 93.3 percent, which would have lifted the headline to 67.9 percent for no reliability reason; a pair where one pass scored and the other abstained is counted as disagreement, not dropped; and the intervals bootstrap over countries rather than observations, since scores within a country are not independent. The per-dimension table is published but explicitly not ranked, because at n=15 the intervals overlap. Gwet AC1 was chosen over Cohen's kappa because these marginals are skewed and kappa collapses toward zero on skewed marginals even under near-total agreement. Harness, run files and per-dimension detail in validation/crisk_test_retest/; all 30 calls ran through the claude -p subscription path, never the metered API.

v2.1 - effective September 2026. Forward-calibration pre-registration restated. No score, no weight, no gate and no band edge changes, and no country moves. Two defects were found in the validation design, both created by the v2.0 restatement rather than present at v1.0. First, the registered t=0 baseline was the 1 May 2026 snapshot scored under risk-v1.0, and v2.0 replaced that composite in place, so the calibration as written would have judged a formula the model no longer runs, and one that tracked the World Bank Worldwide Governance Indicators at about +0.86. Second, the registered hypothesis compared countries scoring 1.5 or below against the cohort average, a band holding 5 countries on the v1.0 baseline and 2 on v2.0, with the five registered bands running 2 / 10 / 36 / 6 / 0 and the top band empty. Neither window could have returned a publishable result. The fixes: the v1.0 registration is retained and will be reported on its original dates, relabelled as a benchmark arm on the retired governance-echo formula, because deleting it is the goalpost move the section forbids; a second registration is added for the live model, locked to the 1 September 2026 snapshot, the first captured under risk-v2.0, SHA256 06e2ebcc...0268ece0, with 12-month and 24-month reports in September 2027 and September 2028. The primary statistic moves from a band threshold to the AUC of t=0 score against at-least-one-event across all 54 countries, published with a bootstrap interval, which removes the dependence on cell sizes entirely. Tercile and readiness-call contingency tables are published alongside as descriptive, both cut at t=0. A power pre-registration is added: the AUC needed to clear 0.50 at plausible event counts is tabulated, the 12-month report is committed in advance to being a first look rather than a verdict, and where fewer than 8 countries record an event in a window no inferential claim is made on it.

v2.0 - effective August 2026. Restated as Conservation Deal Readiness led by the fiscal call. The composite is now the weighted mean of five slices, fiscal 0.30, deal environment 0.20, conservation asset 0.20, delivery 0.15, and traction 0.15, renormalized over the slices that carry data. Governance leaves the additive score for context at zero weight; conflict and legal standing become gates that cap the readiness call; sovereign fiscal capacity moves to the lead and is served from the Sovereign Fiscal Lens. The change roughly halves the score's correlation with the World Bank Worldwide Governance Indicators, from about +0.86 to +0.60, by re-posing the question rather than acquiring new data. Well-governed countries with thin park or delivery footprints score lower than under v1 because the governance echo is gone; this is a restatement of method, and the prior series is preserved in the monthly snapshots. The dimension-level scoring and rubrics below are unchanged and now feed the slices, gates, and context.

v1.6 - effective August 2026. Sovereign default aggregation reworked to a market-anchored model. The dimension composite is now a weighted mean of four scored sub-scores - rating 0.30, IMF backstop 0.30, spread 0.25, and fundamentals 0.15 - re-weighted over whichever are populated, replacing the earlier simple mean of five. Multilateral pipeline is retained on the country page as context and is no longer part of the score, because it reads deal-flow capacity rather than default risk. Fundamentals is used only where enough of its components are present. Where a country has no market signal at all, no agency rating and no liquid Eurobond spread, the score is anchored to the IMF backstop distress reading and capped at half a point above it, so a sovereign the markets will not price cannot read as mid-tier on the softer sub-scores alone; where neither a market signal nor an IMF reading exists, the dimension abstains. Each country carries a provenance flag, market-validated or IMF/DSA-anchored, and the coverage note reads "N of 4 scored sub-scores". No change to dimension weights or category structure.

v1.5 - effective May 2026. Internal weight sensitivity section added between Scoring logic and Outlook. Two parallel tests: category-level (6 perturbations of plus or minus five percentage points across the three category weights, within-category relative weights preserved) and dimension-level (18 perturbations across the nine dimension weights, other eight renormalised). For each of the 24 perturbed weight sets, every country composite is recomputed under the live null-handling rule and Spearman rank-correlated against the baseline ranking. Minimum Spearman rho 0.9970 across all 24 perturbations (operator-government relations dimension, plus-five-percentage-points shock); both levels clear the CCS sensitivity benchmark of 0.989. No change to dimension weights, category structure, or any composite scoring. Validation outputs are preserved.

v1.4 - effective May 2026. Sub-score disagreement reading removed from the country page UI and from the methodology. The reading was a population stdev across populated sub-scores, banded low / moderate / high, surfaced as a pill on the sub-score breakdown header. Removed because the underlying spread is already visible in the sub-score rows that sit immediately below the pill, making the pill a duplicate signal in abstract language. The disagreement payload is retained in the country dataset for analyst use but is no longer rendered. No change to scoring.

v1.3 - effective May 2026. Quantitative governance backbone added on every country record, sitting alongside the editorial governance score without replacing it. Six standard governance-indicator estimates (voice and accountability, political stability, government effectiveness, regulatory quality, rule of law, control of corruption), with global percentile rank computed by Canopy from the raw distribution against the full global cohort each year. The upstream source recently retired its pre-computed percentile-rank series, so percentile rank is now a Canopy computation. An independent corruption-perceptions index added as a triangulation point on the corruption channel, accessed via a stable public mirror. Both fetchers wired into the monthly snapshot job so new annual releases flow through automatically and are captured in the next snapshot. Two countries flag for editorial review where the editorial 1-to-5 score and the computed percentile diverge materially: Cabo Verde and Zimbabwe. No change to dimension weights, category structure, or any composite scoring.

v1.2 - effective May 2026. Several additions to the sovereign default dimension delivered as the v1.2 sprint. Fundamentals adds a seventh component, FX-denominated debt share, capturing structural fragility to currency depreciation that the original six components did not read directly. CFA franc treated as domestic for WAEMU and CEMAC sovereigns. Each country page now renders an implied-rating-in-agency-notches translation of the fundamentals composite alongside the actual rating composite, with the notch difference between them. Spread sub-score now decomposes into total and idiosyncratic components against a Canopy African Composite computed live from the populated spreads. Each country with at least three populated sub-scores carries a disagreement reading (population stdev across sub-scores, banded low / moderate / high) surfacing where the composite hides material divergence between rating, spread, fundamentals, IMF backstop, and multilateral pipeline views. A stress-test panel renders the composite under four IMF debt-sustainability style shocks (commodity price, growth, exchange rate, contingent liability), with shock magnitudes uniform across the universe and per-country calibration on the v1.3 roadmap. Each country with an agency rating gets a peer-comparison panel against rating-band peers (BBB / BB / B / CCC / Distress bands), with rank, band means, and per-sub-score deltas. A coverage panel surfaces front-of-page which sub-scores carry the composite vs which are missing, data-limited, or imputed under fallback methodology. Validation section commits to forward-calibration reports at May 2027 and May 2028 against a pre-registered stress event taxonomy. Sub-score 5 renamed from "catalytic capacity" to "multilateral pipeline" for clearer analyst language; internal keys unchanged. No change to dimension weights or category structure.

v1.1 - effective May 2026. Two additions to the sovereign default dimension, taking it from three sub-scores to five. First, fundamentals added as a primary-data scorecard built from IMF surveillance and country reporting across six components: five fiscal (debt to GDP, debt service to revenue, primary balance, reserves in months of imports, gross financing needs as percent of GDP) and one monetary (headline consumer-price inflation). Each country now carries a divergence-vs-rating signal showing where Canopy's fundamentals view sits relative to the agency rating composite. Second, catalytic capacity added as a four-channel reading of the multilateral capital pipeline - an IMF-programme channel (programme as catalytic instrument, no debt-sustainability component), a development-bank channel (policy-based, catastrophe-deferred, and investment lending), a climate-fund channel (funded-stage projects), and a regional-development-bank channel (active country strategy and sovereign instruments). The original IMF backstop sub-score is preserved alongside catalytic capacity; the two answer different questions from overlapping inputs. IMF backstop is the distress-prevention lens (programme plus debt sustainability); catalytic capacity is the deal-flow lens (programme as one of four channels). A spread velocity render slot is also in place; first reading from snapshot diff. At this version the dimension composite was the simple mean of five sub-scores, superseded by the market-anchored weighting in v1.6. No change to dimension weights or category structure.

v1.0 - effective April 2026. Public release of RISK in current form: nine dimensions across three weighted categories (Stability, Capital Pathway, Site Durability), all 54 African countries scored, A+ to D composite grade, deterministic outlook system computed from per-dimension trajectory tags. Sovereign default risk is included as a ninth dimension, with three discrete sub-scores - rating composite, sovereign Eurobond spread vs US 10-year, and IMF backstop. The September 2025 internal rebuild expanded RISK from an earlier seven-dimension framework to the current nine-dimension structure ahead of public launch; sovereign default and operator-government relations were the dimensions added.

C-RISK began as a snapshot tool. Each country was scored on the day, published, and the score evolved silently in place as new sovereign actions, governance-indicator updates, or political events arrived. The system measured the present; it did not record the past.

From 1 May 2026, every C-RISK record carries provenance fields recording which rubric was applied, when the score was generated, and which underlying source-data versions were used. The canonical published dataset continues to hold the current score for each country. Each month, an immutable snapshot of that dataset is preserved with cryptographic hashes and a manifest. Historical state can be reconstructed exactly.

The first snapshot was captured on 1 May 2026. Snapshots cannot be reconstructed retroactively, so this is the start of Canopy's panel view of African sovereign risk.

This captures changes to scores over time but not retrospective re-scoring across years. Repeat scoring of the same country across multiple years, which would let C-RISK observe regime transitions and rating-cycle effects directly, is on the broader roadmap.

A composite score in isolation is hard to read. South Africa at 3.5 is meaningfully different from South Africa at 3.5 relative to its rating-band peers. The first reading needs the band to be useful; the second is the band reading. Each country page therefore renders the country's sovereign default composite next to the mean of its rating-band peers, plus a per-sub-score delta against that mean.

Bands follow the standard agency rating groups:

Within the band, the country page shows: rank by composite (1 of N is best in band), the band mean for the composite and each populated sub-score, and the delta between the country's value and the band mean for each. A country whose composite is +0.4 above the band mean but whose fundamentals sub-score is -0.6 below the band mean is signalling that something other than fundamentals is propping up the composite. The reader sees that explicitly rather than having to back it out from the unbanded composite.

Countries without an agency rating do not get a peer comparison panel. The eighteen unrated African sovereigns are heterogeneous in credit quality and would not form a coherent peer group; we say so. Bands of size one or two (the BBB and Distress bands at the May 2026 baseline) are flagged with band size on the panel so the reader can weight the comparison accordingly.

Banding is by integer rating sub-score, which is coarser than full agency notch granularity. South Africa (BB-) and Morocco (BB+) sit in the same band even though they are three notches apart on the agency ladder. The simplification is deliberate at v1.2; finer-grained banding using the full notch ladder is on the v1.3 roadmap, where the additional precision is most useful for crossover credits.

Not every country has all four scored sovereign default sub-scores populated. The composite is computed on whichever are present, re-weighted, so readers need to know whether they are looking at a full four-input view or a thinner one, and which sub-scores are absent. Multilateral pipeline is shown alongside as context and does not count toward the four.

Each country page renders a coverage panel below the peer comparison. The panel names the four scored sub-scores and the multilateral-pipeline context reading, and assigns each a status:

The coverage panel does the work that the older "composite computed on N of M sub-scores" line used to do, in front-of-page structured form. A country with two of four scored sub-scores populated still gets a composite, and the reader sees at a glance which are absent and why. For the many African sovereigns that have never issued Eurobonds, the spread sub-score is missing for a structural reason that is not a data gap; the panel says so plainly rather than leaving the reader to infer it from a blank.

Coverage is computed mechanically from existing data flags: data-limited markers at the sub-score level, fallback indicators at the parent dimension level, and missing scores. No new judgement is introduced; the panel surfaces what was already implicit in the data.

Each country page renders a stress-test panel showing how the sovereign default composite moves under four standard IMF Debt Sustainability Analysis shock scenarios. The recompute is mechanical: a defined shock magnitude is applied to the affected sub-scores, each affected sub-score is floored at 1.0, and the composite is recomputed under the same simple-mean aggregation as the unstressed composite. The before-and-after comparison is therefore apples-to-apples; the only thing that changes is the sub-score input.

The four scenarios mirror the standard IMF debt-sustainability stress template:

Shock magnitudes are uniform across all 54 countries. This is the heuristic version. A commodity price shock is in reality not equivalent for Nigeria (oil net exporter, one third of fiscal revenue) and Mauritius (oil net importer, services-led economy); the uniform -0.5 understates the impact for the first and overstates it for the second. The right calibration is country-specific, drawn from the IMF debt-sustainability stress tables published per country, where each shock is already scored in percentage-point impact on debt-to-GDP and primary balance. Per-country calibration of these magnitudes from those stress tables is on the v1.3 roadmap.

In the meantime the panel does what it usefully can. The relative ranking under stress reveals which composites are robust - countries propped up by strong multilateral pipeline and rating sub-scores absorb a fundamentals shock with smaller composite movement than countries already weighted toward fundamentals. Countries where every sub-score is already near the 1.0 floor cannot move further down under stress, which is itself diagnostic. The worst-case scenario is flagged on the country page so the analyst sees at a glance whether the binding fragility is fundamentals-led or spread-led.

The panel does not predict outcomes. It shows sensitivity. A country whose composite drops from 3.5 to 2.7 under FX shock is signalling that its current composite leans on the spread reading; if the FX shock comes, the composite collapses faster than the dimension trajectory would suggest. The shock magnitudes are flagged as heuristic, the v1.3 calibration roadmap acknowledges the gap, and the recompute is mechanical so anyone with the JSON can verify the math.

A risk model that cannot be falsified is not a risk model. C-RISK commits in advance to what counts as model validation, what events the score is being judged against, and what the success and failure conditions look like. The commitment is recorded here, before the data is in, so that future calibration reports can be checked against this page rather than against any retrospective rationalisation.

Why no backward backtest is offered. C-RISK v1.0 went live in April 2026. Reconstructing what C-RISK would have said about Niger in 2022 or Mozambique in 2016 would require us to suppress everything we now know about how those situations resolved, and there is no clean way to do that. Any retrospective scoring would be hindsight-contaminated. We do not publish one. Forward calibration is the validation move available: lock in the model now, observe what happens, report against the lock.

Two registrations, because the model was restated inside the window. The original registration was written against C-RISK v1.0 and locked to the 1 May 2026 snapshot. The v2.0 restatement in August 2026 replaced the nine-dimension composite with the five-slice readiness score, so the registered t=0 value is a number the live model no longer produces: composite_score was rewritten in place and now carries readiness. Re-pointing the original lock at the new model, or quietly dropping it, is the goalpost move this section exists to prevent. Both registrations stand, and both will be reported.

Stands as written. The baseline is immutable and hashed, held at data/snapshots/2026-05-01/risk.json, SHA256 903be479b4b4f486a8b4838b5acecd94dd558f1daf2b6e5a1c748f72529a85c4, captured 3 May 2026 under risk-v1.0. The calibration can therefore still be run against the scores it registered, and it will be, on the dates it committed to. What it tests is the retired formula. That formula tracked the World Bank Worldwide Governance Indicators at about +0.86, so its table reads as a benchmark on how far a governance echo alone anticipates sovereign stress. It is not a test of the model now published. Registration B is that test.

The live model's own lock. t=0 is the 1 September 2026 snapshot, the first captured under risk-v2.0, held at data/snapshots/2026-09-01/risk.json, SHA256 0d800dac4f5a6430d0e6acaa2a45d732352fa32db4ce6b9623a1db66b8273a6c, taken 2026-09-01 at 13:27 UTC. The restatement was fixed in August and the snapshot was taken in September, so no weight, band edge or gate cutoff was chosen with any part of the observation window in view. The manifest hash is what holds the baseline in place.

A stress event is recorded if any of the following happens to a country during the calibration window. The list is fixed; events outside this list are not retro-fitted into the validation set.

The hypothesis is directional and unchanged in substance: the lower a country's t=0 score, the higher its rate of pre-registered stress events across the window. What changes is the statistic testing it, because the banded version could not have produced a result in either direction.

Why the band threshold was retired. The original test compared countries scoring 1.5 or below against the 54-country average. That band holds 5 countries on the v1.0 baseline and 2 on the v2.0 baseline, where the five registered bands run 2 / 10 / 36 / 6 / 0: 36 of 54 in a single cell, and the top band empty. This is the mean-of-five compression set out in the eighth limitation, arriving in the validation design. A conditional event rate computed on a two-country cell is not a finding, and no arrangement of absolute band edges on a cohort this compressed will make it one.

The primary test is rank-based and uses all 54 countries. For each window, take each country's t=0 score and whether it recorded at least one pre-registered stress event, and compute the area under the ROC curve of score against event. AUC 0.50 is no signal. Because a lower score should mean a higher event probability, the model discriminates if AUC falls below 0.50, and the further below, the sharper the gradient. It is published with a bootstrap 95 percent interval. Nothing in it depends on where a band edge sits, so there is no cell that can be cut favourably after the fact.

Two contingency tables are published alongside it, both cut at t=0 and neither re-cut afterwards. Cohort terciles, whose edges on the September baseline are 2.58 and 3.03 and which hold 18 countries each by construction, give the descriptive gradient the balanced cells that absolute bands cannot. The published readiness call bands, holding 4 / 32 / 18 / 0, give the same table on the cut a reader actually uses to decide. Where the two disagree, both are shown.

Pre-registering a hypothesis without pre-registering its power is how an underpowered null gets reported as a finding. Across 54 sovereigns the seven event types plausibly land on 6 to 12 distinct countries in a 12-month window. At those counts the AUC has to be strong before its interval clears 0.50 at all:

The 12-month report is therefore a first look, not a verdict, and the inferential claim sits with the 24-month report where the accumulated event count roughly doubles. Committed in advance: where fewer than 8 distinct countries record an event in a window, no inferential claim is made on that window, the tables are published as descriptive, and the page states that the window was underpowered. That is a pre-registered condition for saying nothing, not a licence to run a second analysis until something clears.

Four reports, two per registration, on a fixed forward schedule:

• Registration A, 12-month - May 2027. Window May 2026 to April 2027 inclusive, against the 1 May 2026 v1.0 baseline. Reported as the benchmark arm.
• Registration A, 24-month - May 2028. Same baseline, window May 2026 to April 2028.
• Registration B, 12-month - September 2027. Window September 2026 to August 2027 inclusive, against the 1 September 2026 v2.0 baseline. The first calibration of the model now published.
• Registration B, 24-month - September 2028. Same baseline, window September 2026 to August 2028, plus a year-1 against year-2 comparison naming which countries changed band and what followed.

The two sets of windows overlap without aligning, which is a consequence of restating the model mid-window rather than a design choice. The registrations are reported as separate tables and are not pooled. Where the 24-month windows allow a like-for-like read on the months they share, it is published as a secondary comparison and labelled as one.

The protocol is short by design, and runs identically for both registrations.

1. Take the registration's t=0 score for each of the 54 countries from its hashed snapshot.
2. For each country, observe the window. Record any pre-registered stress event, with date and source.
3. Compute the primary statistic: AUC of t=0 score against at-least-one-event, with a bootstrap 95 percent interval.
4. Publish both contingency tables, terciles and readiness call bands, with conditional event rates per cell and the 54-country base rate beside them.
5. State the event count. Where it falls below 8, state that the window is underpowered and make no inferential claim on it.
6. Note any data limitations - missing rating data, conflict events ambiguous on the battle-deaths threshold, programme status changes mid-window. Disclose, do not paper over.

The reports are model-evaluation outputs, not advocacy. If the gradient is absent, the page will say so and the model will be re-examined. Both publishable success and publishable failure are acceptable. The unacceptable outcome is silently moving the goalposts, which is why the superseded registration is reported rather than deleted, and why this restatement is recorded in the changelog with the reason for it.

C-RISK scores the country. That answers a real question, but it stops at the border: it rates Kenya, not the specific reserve a funder is about to back, and it cannot tell a well-run, low-exposure site apart from a stressed one in the same country. The site-hazard layer closes that gap. It reads the satellite feeds Canopy already runs over each protected area and turns them into a park-level index of environmental hazard exposure, shown alongside the country composite so a site can be seen sitting below its country's macro risk (de-risked) or above it (a within-country hotspot).

The layer measures exposure, the operating environment, not management. How well a park responds to that environment is the job of PACE. The two are kept apart by construction: the site-hazard index is built only from physical satellite feeds, so ranger presence, staffing, and management effectiveness never enter it. A park can carry high hazard exposure and still score well on PACE, a well-run park in a difficult landscape, which is exactly the distinction a single country number hides.

Five feeds, each already produced per park: forest loss, active fire, vegetation stress, surface-water loss, and rainfall deficit. For forest, fire, vegetation, and water the layer takes the feed's current level against its own baseline, scored 0 when below-normal or normal, 1 when elevated, 2 when high. The low-activity floor is already built into that level: a park that essentially never burns reads "below floor", not a false alarm off a handful of pixels. Rainfall contributes drought exposure from its own seasonal position (dry 1, drought 2). Fire carries one extra check, its live reading is capped by the park's own multi-year fire season, so a site that burns every dry season as a matter of course is not flagged as high hazard for doing what it always does. That cap only ever lowers a reading. Night-lights are produced but held out of this index by design, not for want of data. Radiance measures human pressure, the reach of development, roads and encroachment, which is a different kind of signal from environmental stress, and folding it into an equal-weighted environmental mean would change what the index means and, because it reads normal across almost every park, mostly dilute the real stress signal. It is shown separately as a human-pressure read on each park page rather than scored here.

The index is the mean of the scored feeds rescaled to 0 to 100, where higher is worse. It is deterministic, a fixed function of the feeds, with no model call and no re-derivation of data already held. Four bands read it at a glance: low, moderate, elevated, high. The card leads with the band and the feeds driving it rather than the bare number, so one acute signal is not washed out by calmer feeds around it. A worsening or improving trajectory is shown as a per-feed badge beside the index, which keeps present-state exposure and direction of travel distinct.

Coverage gaps are stated, never filled. Not every park carries every feed with enough history. A feed without a usable reading is dropped from the mean, never counted as zero, because a missing feed is not the same as a calm one. Where fewer than two feeds can be scored, the park is reported unrated rather than handed a low score that would read as "safe" on no evidence.

The site-hazard index is a separate lens, not a new term in either score. It is not folded into the country composite, and it never moves a PACE dimension: the satellite feeds seed this display, they do not overwrite a judged score. The comparison on each park page flips the C-RISK composite (1 to 5, where 5 is safer) onto the same worse-is-higher scale as the hazard index, so a site and its country can be read against each other, and it labels the direction plainly so the two scales are not confused.

C-RISK scores are informational only. They are not investment advice, partnership recommendations, sovereign ratings, or a solicitation for any financial transaction. Canopy makes no warranty as to the accuracy or completeness of the underlying data. All philanthropic or investment decisions should be made on the basis of independent due diligence.