Punjab, kharif season 2026 · Sentinel-2 · Monad testnet
Punjab's fire count fell 90%. Burnt area fell a third.
Farmers learned to light stubble after the satellites pass at 1:30 pm. Afterburn reads the scar a fire leaves for days, whatever hour it was lit, and pays farmers who don't burn through a pact anyone can recheck.
The pact, live on Monad testnet
24 real fields in three villages, judged on this season's imagery. The farmers and the enrolment are simulated; nobody signed up these fields. A quarter of the payment goes out at sign-up, the rest after a clean verdict, and a village bonus only if every enrolled field stays clean.
What happened on-chain
How it works
Read the scar
For every clear Sentinel-2 pass, each field is compared with its last clear view. A burnt field has no green crop, turns dark, and its Mid-Infrared Burn Index rises to the level of fresh char. A dozen published thresholds, no black box.
Post the verdict
The verifier posts the share of the field scarred and the keccak-256 hash of an evidence file: the field's polygon hash, the exact scene IDs, the rule and the numbers. The challenge window opens.
Anyone can recheck
Refetch the same public scenes, rerun the rule, compare hashes. Different answer? Stake a bond and challenge. The arbiter upholds it, or the bond goes to the pool.
Pay, or don't
Clean field: the remaining 75%. Burnt field: nothing more, and the money is freed. Village bonus only if every field stayed clean. Each 50 m land cell can be enrolled once, so the same field can't be paid twice.
Does the rule find real burns?
We checked it against a different satellite on a different schedule. Inside the 375 m footprint of each VIIRS fire detection between Oct 13 and Oct 18, the rule calls 18.9% of the harvested land newly burnt. On random harvested land: 5.6%. That's a 3.4× lift.
And the other way round: of the 5,585 scars it found in those five days, only 2.9% have a VIIRS detection within 375 m. That's the gap the fire count leaves.
A learned model, trained positive-unlabeled on fields Sentinel-2 caught in flames, did worse (2.5× against 3.7× in the same test). The rule makes the verdicts.
What it can't do yet
- A field burnt and ploughed before the next clear pass can look like plain ploughed soil. The rule leans toward clean, since a false "burnt" takes money from a farmer who did the right thing.
- No field-level accuracy figure yet. The best published model for Punjab (a CNN, Anand et al. 2025) reports a Dice score of 0.64. A pilot needs ground visits on a sample.
- Clouds and smoke haze cost passes: two of six were unusable here.
- The Punjabi notices are machine-written and not yet checked by a native speaker.
Recheck a verdict yourself
Every number on this page can be recomputed from public data in a few minutes.
git clone https://github.com/JannetEkka/afterburn && cd afterburn
pip install -r requirements.txt
python pipeline/recheck.py docs/evidence/plot-19.json --pact <pact address> --plot 19It reads the scenes the evidence lists from the Earth Search catalogue, reruns the rule, and prints the recomputed share next to the file's hash and the chain's. All 24 demo verdicts match.