How we keep a number defensible.
Our products are built so that, when an auditor asks "how do you know?", there's an answer — at the plot, the transaction, or the source row. This page explains how we handle provenance, how our evidence records work, and exactly which public datasets we rely on, what each is good for, and where each one stops.
We design around keeping the trace intact.
Generic dashboards aggregate numbers and present them. We do the opposite: every value is meant to be walked back to where it came from. Three habits make that possible.
Captured once, with its metadata.
Plot polygons, transactions, and questionnaire responses are collected with the provenance they need from the start — coordinates, dates, and sources — not reverse-engineered from a PDF later.
An evidence record, not just a result.
Where a result matters, we keep the inputs behind it: a per-record hash so you can prove a record hasn't changed, satellite snapshots dated to the day of assessment, and a citation for every source value.
One record, many outputs.
The same dataset feeds compliance filings, funder reports, and internal decisions — without re-keying, reformatting, or quietly rounding to make a chart work.
Verifiable, dated, and cited.
- Source citation Every value carries which dataset it came from and when it was pulled — no anonymous numbers.
- Per-record hash A cryptographic hash over each record lets anyone confirm later that the record they're holding is the one we assessed — unchanged.
- Dated snapshots Satellite and map layers are pinned to the date of assessment, so a 2026 decision can be re-checked against the imagery it actually used.
- Retention Evidence records are retained for audit (Verdex keeps per-plot evidence ≈5 years); the underlying client data stays under the client's control.
The public datasets behind the products — and their limits.
We build on the same sources regulators, funders, and auditors already use. None is perfect; each is good for some things and not others, and we say which.
Market & production data — Commodity
International Cocoa Organization
Use: headline cocoa production, grindings, stocks, and indicator prices. Limit: country/global aggregates published periodically — not farm- or shipment-level.
UN Food & Agriculture Organization
Use: national production, area, and yield context across crops and years. Limit: national aggregates with reporting lags; methods vary by country.
Ghana Cocoa Board
Use: Ghana's official cocoa purchases and producer prices. Limit: Ghana-only, official figures, published on a periodic schedule.
Commodity price & indicators
Use: long-run price benchmarks ("Pink Sheet") and development indicators. Limit: monthly price aggregates, not transaction prices.
Geospatial & EUDR — Verdex
UMD Global Forest Change
Use: flag tree-cover loss on a plot after the EUDR 2020 cutoff (Landsat-derived, ~30 m, annual). Limit: tree-cover loss is not by itself legal deforestation — it includes harvest and plantation cycles — and smallholder mosaics can be over- or under-flagged.
EU Joint Research Centre
Use: forest cover and degradation context (e.g. the Tropical Moist Forest dataset). Limit: resolution and timing constraints; forest definitions differ between datasets.
World Database on Protected Areas
Use: flag plots in or near protected and conserved areas. Limit: boundary accuracy and currency vary, and a designation doesn't capture every local legal nuance.
Claim decomposition is the intellectual property.
Most reporting in this sector quietly presents evidence for one claim as proof of another, and most buyers cannot see it happening. Take a shade-tree planting claim: each rung needs its own evidence.
- "100,000 seedlings were dispatched"Stock leaving the nursery or distribution point.
- "…were received by eligible farmers"Receipt, quantities, recipient identity, and eligibility against a dated register.
- "…were planted"Evidence of planting. Not dispatch. Not receipt.
- "…survived"Observation at an appropriate later period, using an agreed survival measurement method.
The same ladder runs through everything. A farmer who attended one session is not a farmer who completed training. A mapped farm is not an allocated farm. A screened polygon is not an owned polygon. The client is paying us to stop evidence for one statement being presented as proof of another.
A requirements and tests register.
Before a check is written, record which requirement is tested and where it comes from — source, version, test identifier, expected evidence, treatment of exceptions.
A workflow an outsider can rerun.
Input files preserved, rules and code version recorded, excluded records explained, detail reconciled to the reported totals. Legitimate repeat attendance is tested separately from duplication.
The evidence is never overwritten.
The supplier's original submission, our working copy, and any authorised corrected version — always kept separate.
Is the number even possible?
Area-to-volume mass balance from public benchmarks — FAO, ICCO, COCOBOD, national statistics — not from our own data. A supplier can argue with our numbers; it is harder to argue with FAO's.
Declared before testing.
Risk-based or random, size justified, what can and cannot be inferred about the population — and the magnitude of error that matters, agreed with the buyer first.
Someone who is not the preparer signs off.
A named external reviewer checks evidence, sampling, conclusions and report language before issue. When something looks like fraud, we report and hand off — we do not investigate.
Evidence to inform a decision, not the decision itself.
Surface risk, with the evidence.
A Verdex result is a risk rating — pass, watch, or fail — with the dated layers and hashes behind it, to inform legal review.
Issue certification or an audit opinion.
We don't certify a plot as legally compliant, nor file your Due Diligence Statement for you. A verification review is second-party work commissioned by the buyer — not certification, not an independent audit opinion, and not accreditation under ISO/IEC 17029, whose principles inform our method. The final call stays with your compliance or legal team.
Hide the limits.
Where a source is coarse, lagged, or ambiguous, we say so on the record rather than smoothing it into a cleaner-looking number.