Data & Model Limitations

Known blind spots, coverage gaps, data sparsity warnings, model caveats, and operational assumptions.

This page is required reading for any institutional or compliance use case. Understanding the system's blind spots is essential for responsible use of scores and recommendations.

Scope

Projects without registry presence

Voluntary projects not listed on Verra, Gold Standard, or other catalogued registries are not ingested and will not appear in scores or recommendations.

Unverified PDD claims

Scores are derived from structured registry data, not from manual review of Project Design Documents. Co-benefit claims are taken at face value from source data.

Post-issuance project changes

Ownership transfers, methodology reversals, or buffer drawdowns may not be reflected until the next successful ingest run.

Retirement data lag

Retirement records may lag 24–72 hours behind the registry public record depending on connector refresh frequency.

Coverage

DimensionCurrentNotes

Registry coverage6 registriesVerra VCS, Gold Standard, CAD Trust, RegenNetwork, Carbonmark, CarbonPlan. Others in catalog-only (not implemented).

Field completeness~78% avgGeometry, document URLs, and co-benefit detail are missing for a significant portion of projects.

Price coverage~45%Only projects listed on Carbonmark or other integrated venues have live price data.

Governance data vintageCPI 2023Country risk scores are updated annually. Mid-cycle changes are not reflected.

Sparsity

Missing fieldScore impact

batches.vintageEndVintage penalty defaults to 0.2 (unknown) — conservative but inaccurate for real projects without batch data.

identifiersCross-registry concordance scores neutral (0.5) when only one registry source exists.

qualitySignals.sourceMetricsSource reliability falls back to tier-based default when connector hasn't run with full metrics.

project.countryGovernance risk defaults to global average (0.45) when country is missing or unrecognised.

Assumptions

The scoring recipe is deterministic — same inputs always produce the same output. Randomness is not used.

All weights and risk tables are static until the recipe version is bumped. There is no online learning.

Source tier assignments (T1/T2/T3) are set manually in the connector catalog and not inferred from data quality.

The permanence proxy is type-based only — it does not account for project-specific buffer pools or insurance arrangements.

Mapping confidence is a weighted sum of feature scores, not a probabilistic model. It does not generalise beyond the configured feature set.

This system does not provide investment advice. Scores are data-quality indicators, not credit ratings.