Validation and evidence
Every Phoeniks figure carries an evidence level and can be traced to the exact model version and input snapshot that produced it. Validation follows a documented holdout and calibration protocol, and known failure modes are published in the pack rather than omitted.
- Measured, modelled, inferred, untested
- Spatial and temporal, documented first
- Tail behaviour and spatial covariance
- Not published; metrics are per context
Evidence scoring levels#
| Measured | Taken from an authoritative published source | Citable in a regulatory filing |
| Modelled | Produced by the validated ensemble | Usable when labelled as modelled |
| Inferred | Derived where direct data is absent | Indicative; needs corroboration |
| Untested | No validation evidence in this context | Not a basis for a decision |
Validation protocol#
- Spatial and temporal holdout design, documented before results are read.
- Calibration against official statistics and trial data.
- Tail behaviour and spatial covariance testing for portfolio users.
- Known limitations stated explicitly in every validation pack.
Frequently asked
- Do you publish a headline accuracy number?
- No. A single accuracy figure detached from crop, geography and season is not meaningful. Validation metrics are delivered per context in the validation pack.
Related pages
Methodology
The Phoeniks methodology: a process-based APSIM, DSSAT and ORYZA ensemble, scenario configuration, scoring engine and a reproducibility ledger.
Sources and licensing
Every source Phoeniks ingests, what it contributes, and how licence terms and provenance are enforced before a record is exposed commercially.
Security and governance
Access control, tenant isolation, audit logging, encryption, data residency and incident response across the Phoeniks platform.
See also
Request access
Institutional access. No self-serve trial.