Sentinel is a screening layer for ICVCM Core Carbon Principle #5 — Sustainable development benefits and safeguards. Every signal we surface maps to a sub-criterion. Every band has a source. Where we proxy, we name the proxy and the upgrade path.
CCP #5 — Sustainable development benefits and safeguards — requires carbon-crediting programs to put in place robust social and environmental safeguards. In practice, the principle covers four obligations:
— Integrity Council for the Voluntary Carbon Market, Core Carbon Principles, Assessment Framework, March 2023. Wording paraphrased; full text at icvcm.org.
We don't claim to be CCP-certified. We claim to be a defensible first-pass screen for a CCP-aligned DD process. Every signal below maps to a sub-criterion of CCP #5.
UNDRIP Articles 10, 19, 32 · ILO Convention 169
Any overlap → flag. Number of overlapping nations contributes (capped at 2).
Binary detection is the standard for FPIC scoping; severity is left to the analyst.
Native Land Digital (free tier deprecated late 2025; bundled overlay used for sample projects)
Wire Native Land paid API key for full coverage.
ICVCM CCP #5 — 'identify, prevent and mitigate adverse social or environmental impacts'
Article scores as adverse only if title contains BOTH a safeguards-topic keyword AND an adverse-claim keyword. NGO-domain articles +1.
Topic+claim co-occurrence kills the 'Project secures FPIC' false positive that plagued single-keyword scoring.
GDELT 2.0 DOC API + Google News RSS, English-language
LLM-per-article claim classifier; multilingual sources
Sabin Center for Climate Change Law — Climate Case Chart (cited in IPCC AR6 WG3 Ch. 15)
Any case matched by project name or developer → flag (+3 per case).
Sabin Center is the field-standard climate-litigation registry.
Curated subset + auto-promotion from news inference
Severity weight by stage: injunction > pending > dismissed
Carbon Market Watch · Survival International · Forest Peoples Programme · Rainforest Foundation
Any complaint from curated ledger or NGO-domain news → flag (+2 per).
These four NGOs are the orgs most cited by Bloomberg, FT, NYT and Mongabay on carbon-project complaints.
Curated ledger + auto-promotion
Direct NGO-feed integration
World Bank, Forest Area (% of land area) — indicator AG.LND.FRST.ZS · FAO Global Forest Resources Assessment 2020
≤ −0.3 pp/yr → RED · −0.3 to −0.05 → AMBER · > −0.05 → GREEN
FAO characterises annual net loss > 0.3% as 'high deforestation rate.' The amber threshold reflects sustained non-trivial loss.
Country-level, not project-polygon
Project-polygon GLAD alerts via Global Forest Watch
Transparency International, Corruption Perceptions Index 2024
CPI < 30 → RED (TI 2024: 'serious corruption problems') · 30–42 → AMBER (below TI 2024 global avg of 43) · ≥ 43 → GREEN
Every threshold cites a TI-published figure — no invented cut-offs.
Country-level enabling-environment proxy, not project-level read
Add World Bank Worldwide Governance Indicators (WGI) Control-of-Corruption percentile
UNDRIP Articles 10/19/32 · ILO Convention 169 · UN-REDD FPIC Guidelines 2013 · Cancun Safeguards (UNFCCC 1/CP.16, Appendix I)
4 checks: Consultation · Consent · Grievance mechanism · Withdrawal protections. ≥2 fails → RED · 1 fail → AMBER · only insufficiencies → AMBER · clean → GREEN.
Mirrors the UN-REDD operational FPIC checklist. A screen defaulting to 'insufficient' is honest — public signals don't replace project-document review.
Derives from already-gathered evidence (territory, news, litigation, NGO)
Direct evidence ingestion: PDD, monitoring reports, FPIC consultation records
The composite is a transparent heuristic — an analyst can read the score breakdown line-by-line in services/score.py. We chose rule-based over a learned model because rule-based scores are litigation-defensible. A model's “0.73 risk score” is not.
The three breakpoints below are heuristic — they balance signal severity against false-positive rate. They are not claimed to be industry-standard; they are claimed to be transparent and tunable.
Material safeguards concerns surfaced. Deep DD before any commitment.
Some signal. Targeted follow-up on the specific flag.
No adverse signal detected in the public-evidence layer.
We are not ICVCM-certified or label-eligible.
We do not replace the project workpaper or the IC memo.
We do not claim project-polygon precision today — country-level proxies are labelled as such on every panel.
We do not adjudicate FPIC — the procedural checklist flags where a DD analyst should pull project documents (PDD, monitoring reports, consultation records).
| Gap | Today | v0.4 plan |
|---|---|---|
| Indigenous overlap dark for uncurated projects | Bundled overlay only | Native Land paid API key wired |
| News scoring keyword-based | Topic + claim co-occurrence (v0.3) | LLM-per-article claim classifier; multilingual sources |
| Environmental + governance country-level | World Bank + TI CPI | GLAD alerts at project polygon + WGI percentiles |
| Litigation severity not weighted | Flat +3 per case | Injunction > pending > dismissed |
| No recency decay on news | Uniform 24-month window | Exponential decay exp(−days/365) |
| No developer-level aggregation | Per-project only | Developer portfolio rollup |