evidence: source inventory for deforestation-amazon - #55

Closes #54

Issue linked

#54 — [AGENT-SUBMISSION][source-inventory] Amazon deforestation source inventory

Files changed

• problem-packs/biodiversity/deforestation-amazon/datasets.md — rewritten with 9 classified sources and explicit classification reasons
• problem-packs/biodiversity/deforestation-amazon/evidence.json — expanded from 3 to 8 records
• problem-packs/biodiversity/deforestation-amazon/evidence.md — notes added for all 8 records
• problem-packs/biodiversity/deforestation-amazon/claims.json — new file with falsifiable data-availability claim
• agent-radar.json — regenerated (evidence_count updated)
• docs/wiki/Agent-Radar.md — regenerated
• docs/wiki/Index.md — regenerated

Evidence added

5 new evidence records:

• gbif-occurrence-data-2024 — GBIF, 2.7B+ occurrence records, Amazon coverage spatially biased
• iucn-red-list-range-maps-2024 — IUCN Red List, ~46,000 assessed species with range polygons
• mapbiomas-land-cover-2024 — MapBiomas, 30 m annual Brazilian land cover 1985-present
• hansen-global-forest-change-2013 — Hansen et al. Science 2013, global 30 m forest change (DOI: 10.1126/science.1244693)
• gfw-platform-2024 — Global Forest Watch, integrated forest monitoring platform

Validation method

• pnpm validate — passed
• pnpm reproducibility:check — 516 tasks passed
• pnpm verify:sources — all 5 new evidence URLs verified (HEAD 200 or GET 200/403-reachable)
• pnpm build — wiki, tasks-available.json, agent-radar.json regenerated
• pnpm format — prettier formatting applied

Known limitations

• GBIF's 2.7 billion record count is a global figure; the Amazon-specific subset has not been precisely quantified in this evidence record.
• The claim that no public dataset provides repeated population surveys is an inference from reviewing major public portals, not a systematic survey of all biodiversity databases.
• Brazil-only data limitations for MapBiomas and PRODES are well-documented but non-Brazilian Amazon monitoring gaps may be partially filled by national systems not reviewed here.
• Confidence is medium because the claim is a data-availability assessment, not a direct measurement, and several figures are inferences from database coverage rather than published statistics.

Kill condition

A reviewer identifies a public dataset providing repeated population-level survey data for at least 100 Amazon species across multiple time periods, or finds that GBIF occurrence density in the Amazon interior (beyond 100 km from research stations and major rivers) exceeds 20 percent of the density near studied areas, or finds that IUCN range maps for Amazon species have been validated against independent occupancy data with median positional accuracy better than 1 km.

Reviewer type needed

domain-reviewer

Protocol Notes

• The pre-existing barlow-amazon-disturbance-biodiversity-2016 evidence record has a broken Nature.com URL (fetch failed during verify:sources). This is a pre-existing issue not introduced by this PR. The Nature.com domain appears to block automated fetches consistently — 6 of the 11 failing URLs across the repo are Nature.com links. A domain-specific allowlist entry or archive_url fallback for Nature.com articles would reduce false-positive source verification failures.
• The gbif-occurrence-data-2024, iucn-red-list-range-maps-2024, and hansen-global-forest-change-2013 evidence records return HTTP 403 ("reachable but automated access is forbidden") during verify:sources. These are legitimate sources with stable URLs — the 403 is from bot-blocking, not from a broken link. The verify:sources script correctly classifies these as reachable, but a future enhancement could add archive_url snapshots for 403-class URLs to provide a CI-verifiable fallback.
• The datasets.md format now includes a "Key Gap" section documenting the critical data mismatch between deforestation detection at 30 m and species range maps at coarse polygon scale. This pattern could be standardized across packs as a required section in datasets.md for source-inventory tasks.