Science Writer
frames the claim for a non-specialist audience.
Tropical regions with high active-fire counts AND active deforestation alerts are losing forest carbon at rates exceeding bottom-up inventory.
Net biosphere carbon sink collapses to neutral or net source in measured tropical regions within 5-10 years.
This hypothesis is in the forming stage. Captain is accumulating the data stream necessary to detect the SUPPORTS or FALSIFIES condition with statistical significance. The metric — Percentage of active fire alerts within 50km of active deforestation alerts, per tropical region (Amazon, Congo basin, SE Asia), 12-month rolling window — needs to stabilise across the 3 endpoints, and the council has not yet seen enough data to assess proximity to either threshold.
Decision point: when enough data has accumulated to compute the metric with stable confidence intervals, the hypothesis advances to monitoring.
Metric: Percentage of active fire alerts within 50km of active deforestation alerts, per tropical region (Amazon, Congo basin, SE Asia), 12-month rolling window
Status: requires forestwatch historical series for percentile rank
/api/fires
loading
/api/forestwatch
loading
/api/deforestation
loading
Captain reads 3 Earth API endpoints together (/api/fires + /api/forestwatch + /api/deforestation). The hypothesis emerges only at their intersection — none of these streams alone reveals the pattern.
Spatial-overlay: % of active fires within 50km of active deforestation alerts. Baseline ~15%; sustained > 25% confirms co-evolution.
frames the claim for a non-specialist audience.
weighs cross-evidence strength.
Synthesises 2 angles into the formal hypothesis, sets thresholds, schedules revisits when data lands.
Five independent claude-sonnet-4-6 calls, one per persona — Skeptic, Fact-Checker, Researcher, Compliance-Guard, Falsification-Auditor. Each writes its hardest objection from its own seat, paired with the methodological resolution it would accept. Run on the static catalogue spec Jun 3, 2026; a live council for any topic is at /try.
The spatial co-occurrence of VIIRS/MODIS active fire alerts and GLAD deforestation alerts within 50km is most plausibly driven by a single upstream confounder—agricultural frontier expansion—rather than two genuinely co-evolving ecological processes. In the Amazon and Borneo specifically, smallholder and industrial land-conversion follows documented seasonal burning calendars (captured in MAPBIOMAS v7 land-use classifications), meaning both alert streams are triggered by the same clearing event on overlapping timelines. This would inflate the co-occurrence percentage well above the 25% threshold and produce the apparent top-quartile simultaneity even if fires and deforestation had no mutual reinforcing feedback whatsoever, rendering the carbon-sink collapse prediction untestable from the spatial-overlap metric alone.
Apply an 8-day lagged Granger causality test between VIIRS fire radiative power (NASA FIRMS archive) and GLAD weekly alert counts at 0.1° grid cells across each basin, with MAPBIOMAS v7 annual agricultural-expansion area as a panel covariate to partial out shared land-conversion variance. A true co-evolutionary mechanism requires statistically significant bidirectional Granger causality (F-statistic p<0.05, FDR-corrected) in both the fire→deforestation and deforestation→fire directions, with coefficients remaining significant after the MAPBIOMAS control is included; the agricultural-burning alternative predicts only unidirectional causality with deforestation leading fire by one to three 8-day lags and the coefficient on fire→deforestation collapsing to near-zero once agricultural footprint expansion is held constant.
The VIIRS active fire product (VNP14IMG, 375 m) carries cloud-related omission errors of 20–40% in tropical humid regions, while Landsat-based GLAD deforestation alerts (30 m, ~16-day nominal revisit) suffer 30–60% temporal data gaps during monsoon periods in the Congo Basin and SE Asia. Crucially, both omission effects are spatially and temporally correlated: the same rainy-season cloud decks simultaneously suppress fire occurrence, fire detection efficiency, and deforestation alert generation, producing a systematic, season-locked downward bias in the co-occurrence percentage. The proposed 10-percentage-point threshold band (15% baseline to 25% confirmation) falls well within the combined joint-detection uncertainty, estimated at ±5–8 percentage points under these correlated gap conditions.
Restrict analysis to dry-season windows where MODIS MOD35 cloud-mask coverage is below 20% of the study tile, and use only VIIRS detections flagged as "high confidence" (≥75% in the VNP14IMG confidence field) paired with GFW GLAD alerts confirmed by at least two independent Landsat passes (the "confirmed" rather than "potential" alert layer). Widen the SUPPORTS threshold to ≥35% co-occurrence (adding the ±8 pp uncertainty buffer) and require it to persist across at least two consecutive dry-season windows in each basin; bootstrap-resample cloud-gap-imputed pixels to construct 95% confidence intervals on the co-occurrence percentage before any threshold comparison, so that the FALSIFIES criterion (inverse correlation) can only be declared when the upper CI bound remains below the 15% baseline.
The dominant uncontrolled confounder is ENSO-driven regional drought, which simultaneously elevates background fire ignition probability across all vegetation classes—including intact forest far from any clearing activity—and increases deforestation alert detection rates by producing anomalously cloud-free dry-season imagery that exposes fresh clearings otherwise masked by persistent cloud cover. Because both fire alert counts and deforestation alert density spike during El Niño episodes through a shared climatic mechanism rather than through any causal co-evolutionary process, the 50 km proximity overlap metric will routinely breach the 25% threshold in strong ENSO years without any true coupling between logging frontiers and fire regimes. The coarse spatial buffer cannot discriminate anthropogenic clearing fires from drought-stressed intact-forest fires, so the estimated co-occurrence percentage absorbs both signals indistinguishably.
Incorporate ERA5 monthly vapor pressure deficit (VPD) anomaly and the 3-month Standardized Precipitation Index (SPI-3) from the Copernicus Climate Data Store as time-varying covariates, alongside the NOAA Oceanic Niño Index (ONI) as an explicit ENSO state flag, within a region × calendar-month panel regression that includes region and year fixed effects. Restrict the primary co-evolution test to ENSO-neutral or La Niña periods (ONI < +0.5 °C) to remove drought-confounded observations, and apply a within-frontier versus outside-frontier contrast using MapBiomas annual land-use transition layers (available at mapbiomas.org) to confirm that elevated co-occurrence is geographically concentrated along active agricultural expansion edges rather than distributed uniformly across drought-stressed landscapes—thereby isolating the anthropogenic clearing channel from the background climate-fire mechanism.
The primary regulatory exposure sits at the intersection of SEC Rule 10b-5 material-misstatement liability and the Verra Verified Carbon Standard (VCS/REDD+) offset protocols integrated into California's Cap-and-Trade compliance market, where corporate emitters hold tropical-forest carbon credits against enforceable emissions obligations. If the unvalidated sink-collapse prediction embedded in this hypothesis is cited as established fact in SEC climate-risk disclosures—now subject to the Commission's mandatory climate-disclosure rules—or in VCS project-validation documents, issuers and offset developers face enforcement exposure for overstating the permanence and additionality of credits whose baseline integrity presupposes a functioning regional carbon sink. Parallel liability arises under IFRS S2 and EU CSRD (ESRS E1), where companies counting tropical land-sector removals toward net-zero targets must faithfully disclose material uncertainties; treating a falsifiable co-evolution hypothesis as an established finding would corrupt the reliability of those filings and trigger potential greenwashing enforcement by the European Securities and Markets Authority or national competent authorities.
Before this hypothesis may be cited in any SEC climate filing, VCS validation report, SBTi Net-Zero Standard target submission, or CSRD/IFRS S2 disclosure, the spatial-overlap metric must be cross-validated against at least two independent ground-truth data streams—specifically FLUXNET tropical eddy-covariance flux-tower records and IPCC national greenhouse-gas inventory submissions—confirming that the API-joined fire/deforestation co-location signals translate to statistically significant above-baseline biospheric carbon loss beyond measurement uncertainty. The SUPPORTS threshold (both metrics simultaneously in the top quartile of the five-year range, sustained above 25% spatial
The FALSIFIES condition — an inverse correlation across only three macro-regions (Amazon, Congo, SE Asia) — is computed on n=3 data points, meaning the Pearson or Spearman correlation carries confidence intervals that span essentially the full [−1, +1] range under the null, making any sign of correlation reachable by chance alone. Simultaneously, the 50 km spatial buffer is large enough that Monte Carlo random placement of fire points within high-density tropical deforestation zones produces null proximity rates of roughly 30–45% — well above the stated 15% baseline and above the 25% SUPPORTS threshold — so the SUPPORTS condition can be entered by chance while the FALSIFIES condition remains statistically indistinguishable from noise. This means the hypothesis is simultaneously too easy to confirm and too ambiguous to falsify.
Replace the three-region correlation with a country- or biome-level decomposition yielding at least n=15 comparable spatial units, then pre-register a specific inverse-correlation threshold (e.g., r < −0.40, p < 0.05 by permutation) so the FALSIFIES arm has genuine statistical power and degrees of freedom. In parallel, run ≥10,000 Monte Carlo permutations that randomly shuffle fire-alert locations within each region while preserving count and temporal structure, using the 95th percentile of the resulting null proximity distribution as the data-driven SUPPORTS threshold rather than the fixed 25% figure; if that null 95th percentile itself exceeds 25%, the spatial buffer must be tightened or the metric reformulated as excess proximity above the local null rather than an absolute percentage.
Unlike the static stress tests above (synthesised against the frozen catalogue spec), this is what a 3-voice council found in the most recent biweekly review. Refreshed on the 1st and 15th of each month at 09:00 UTC. Each voice runs one bounded web search via Anthropic's web_search_20260209 tool, cites what it finds, and recommends a verdict.
The verdict aligns with the curated catalogue status (forming).
Two of three council voices returned 'revision_needed', driven by two distinct but compounding problems: the hypothesis's causal claim is overstated because fire–deforestation coupling demonstrably decouples under strong governance (Brazil's 36% rainforest-loss drop in 2025 while fire risk persisted, per GFW April 2026), and the core metric is structurally compromised by the MODIS Terra decommission and the newly documented systematic suppression of low-confidence nighttime VIIRS detections (Systematic Absence of Low-Confidence Nighttime Fire Detections in VIIRS Active Fire Product, Oct 2025), which breaks the 5-year rolling baseline and biases the SUPPORTS threshold non-conservatively. Revision should disaggregate the co-evolution claim by governance regime, condition thresholds on ENSO state, and recalibrate fire-count metrics against the VIIRS-only instrument uncertainty budget.
Recent evidence (2025–2026) shows that (1) deforestation and fire alerts demonstrably decouple under strong governance (Brazil 2025), contradicting universal co-evolution; (2) the observed sink weakening is better explained by climate-driven fire amplification across multiple biomes rather than tropical fire-deforestation coupling specifically; and (3) the hypothesis's core metric misses the dominant and growing degradation pathway, making its falsification threshold structurally incomplete and the causal claim as currently stated factually overstated.
Brazil's non-fire primary forest loss fell 41% in 2025 to a record low, driven by policy enforcement — demonstrating that deforestation and fire alerts can decouple when governance improves, directly undermining the hypothesis's assumption of tight co-evolution across all tropical regions and its near-term prediction of sink collapse.
WRI/GFW analysis shows the 2023–2024 sink collapse was driven primarily by extreme climate-driven fires — including in boreal and temperate zones — rather than fire-deforestation co-evolution per se, offering climate amplification as an alternative and more parsimonious explanation for the observed carbon anomaly.
Mongabay's synthesis finds that a growing share of tropical carbon loss in 2025 stems from sub-canopy degradation — selective logging, roads, edge effects — which falls outside standard deforestation and active-fire metrics, suggesting the hypothesis's chosen metric (fire/deforestation alert co-location) structurally under-captures the dominant carbon-loss pathway and may misattribute causality.
The imminent MODIS Terra decommission and the transition to VIIRS-only fire and burned-area products constitute a sensor-change discontinuity that breaks the 5-year rolling baseline used to define 'top quartile' thresholds; simultaneously, the newly documented systematic suppression of low-confidence nighttime VIIRS fire detections means the active-fire-count metric underestimates edge fires most relevant to fire–deforestation co-location, making the current SUPPORTS threshold non-conservative and requiring recalibration against the VIIRS-only instrument uncertainty budget.
Analysis of 21.5 million VIIRS fire detections found zero low-confidence nighttime fire classifications where ~697,000 were statistically expected, indicating systematic undocumented filtering that inflates apparent fire-count certainty. This directly weakens the hypothesis threshold, as active-fire-count metrics used to assess 'top quartile' co-location with deforestation alerts are likely undercounting low-intensity fires—particularly those characteristic of understory burning at forest edges adjacent to deforestation fronts.
GFW confirmed the Terra satellite (MODIS) will be decommissioned, discontinuing MODIS near-real-time deforestation alerts and burned-area products. This is a methodology-breaking event: the 5-year rolling baseline underpinning the hypothesis's 'top quartile' threshold was built in part on MODIS data; transitioning solely to VIIRS introduces a sensor-change discontinuity that makes the historical quartile range non-comparable going forward.
NASA released the new VIIRS burned area standard science product (VNP64A1) as a continuity replacement for MODIS MCD64A1 Collection 6.1, but noted that native VIIRS input-band spatial resolution is coarser than MODIS before resampling to 500 m. Coarser radiometric sensitivity means the new product may undercount small fragmented burns at deforestation edges, shifting the detectability floor and potentially misclassifying co-location events that the threshold was calibrated against.
Recent peer-reviewed literature (2025–2026) strongly converges with the hypothesis: fire and deforestation are measurably co-amplifying tropical carbon losses, with 2023–2024 data showing the biosphere sink near its weakest point in two decades. The one nuance is that the coupling strength is heterogeneous across regions (Amazon >> Congo >> SE Asia) and sensitive to ENSO state, suggesting the 5-year threshold window and per-region metric are directionally correct but may need climate-mode conditioning to avoid false falsification signals in La Niña years.
Peer-reviewed satellite analysis confirms fire-induced degradation has overtaken deforestation as the primary Amazon carbon emission driver in 2024, with 791 Mt CO₂ released — directly validating the fire–deforestation co-evolution mechanism and the claim that carbon losses exceed bottom-up inventory expectations.
GFW/Land & Carbon Lab analysis shows tropical forests absorbed only ~25% of their typical annual carbon uptake in 2023–2024 due to compounding fire and deforestation pressure, consistent with the hypothesis's prediction of net biosphere sink collapse.
2026 GFR report documents a 36% drop in tropical primary forest loss in 2025 linked to reduced fire activity under La Niña, but flags SE Asia (Indonesia +14%) as a diverging region — introducing important cross-regional heterogeneity that nuances the hypothesis's uniform threshold logic.
This is an original cross-correlation hypothesis. The pattern emerges only when 3 Earth API endpoints are read together; no single dataset or existing publication isolates the claim as stated here. Captain proposes it as a testable scientific question.
Captain Landseed. (May 30, 2026). Fires and deforestation are co-evolving in the tropics [Working hypothesis, forming, catalogue v6.3]. Landseed PBC. Retrieved Jun 6, 2026 from https://captain-landseed.pages.dev/h/fire-deforestation-coupling/
@misc{captain_landseed_fire_deforestation_coupling,
author = {Captain Landseed},
title = {Fires and deforestation are co-evolving in the tropics},
year = {May 30 2026},
howpublished = {Working hypothesis, status: forming, catalogue v6.3},
publisher = {Landseed PBC},
url = {https://captain-landseed.pages.dev/h/fire-deforestation-coupling/},
note = {Module: biosphere; Originality: NOVEL; Accessed: Jun 6, 2026}
}
TY - GEN AU - Captain Landseed TI - Fires and deforestation are co-evolving in the tropics PY - May 30 2026 PB - Landseed PBC UR - https://captain-landseed.pages.dev/h/fire-deforestation-coupling/ N1 - Working hypothesis (status: forming); catalogue v6.3; module: biosphere ER -
JSON snapshot with all hypotheses, archived council deliberations, current live-state, and the build-over-build activity log. SHA-256 manifest included. CC-BY-4.0.
Five personas deliberate in real time. Typically ~$0.08, 40-60 seconds. Three free runs, then bring-your-own Anthropic / OpenAI / Gemini.