How BLACKOUT decides

Governments order hundreds of internet shutdowns a year. In July 2024 Bangladesh cut a country of 170 million off for days and froze the income of one of the world's largest freelancer workforces. Insurers exclude political risk, and no single oracle can be trusted to say a government did it while that government denies it. BLACKOUT has GenLayer validators decide, one week at a time.

1. Measure: an exact signal

After a covered week, anyone calls observe. Every validator fetches the same historical window from IODA (Georgia Tech's internet outage detection) and computes, in integer arithmetic, how many local days had at least six hours below half of normal connectivity, in either BGP routing or active probing. Normal is the upper quartile of the previous seven days. Because the data is historical and the maths has no floats, every validator must get the identical lost-day mask, or the transaction fails.

Six hours is deliberate: Iraq's one-to-two-hour exam-season cuts, which are scheduled and predictable, never count. A week with no lost day is decided as No shutdown right there, with no news fetched and no model asked.

2. Attribute: a judgment no oracle can make

A collapse in connectivity does not say who caused it. Ghana's March 2024 subsea cable cuts look almost exactly like Bangladesh's shutdown in the data. So each validator fetches that week's news from Google News, keeps only headlines from about 95 allowlisted newsrooms and digital-rights monitors (checked in code before any model sees them), strips invisible characters, fences them, and asks its own model one question with three possible answers: government ordered, not the government, or undecided.

Validators must reach exactly the same class. No answer is taken on the leader's word, undecided included: each class has its own outcome (a payout, a final no, or another look later), so a validator accepts one only when its own reading of its own fetch reaches it. A leader can neither invent a payout nor bury a government-ordered shutdown, and a validator whose model gives no usable answer confirms nothing. An undecided week is observed again after six hours, up to three times. The contract stores the measurement and the class, never a leader's headlines or notes.

3. Settle: no claim to file

Once a week is decided, anyone calls settle. Each policy for a government-ordered week pays the daily amount for every lost day, up to its cap, to its beneficiary: the buyer or anyone they chose, such as family abroad who can still transact while the covered person cannot. A keeper can push the payout to the beneficiary's wallet. Every other verdict expires the policy and its premium stays with the pool.

Rules against gaming it

• No buying into a fire: cover is sold at least 72 hours before the week starts.
• No grinding the judge: one stored observation per week; only an undecided one is retried, on a fixed schedule.
• No injected verdicts: no caller text reaches the model, only allowlisted headlines, and a payout also needs the exact signal.
• No bank run: active cover is reserved and cannot leave; LP withdrawals wait 14 days.
• No concentration: one country-week takes at most 20% of the pool, all cover at most 80%, five policies per wallet per week.
• Priced by history: each recorded government-ordered lost day raises that country's premium by 5%, up to double.

What it does not do yet

• Mobile-only and regional shutdowns barely show in national IODA data (Pakistan, February 2024). A separate order-evidence track is planned.
• Verdicts can be appealed through GenLayer's own appeals; an in-app challenge with a bond is planned.
• This is a test network deployment with test GEN. Past weeks marked Replay were opened by the deployer so the whole flow can run on real events; live weeks follow every rule above.

Source, tests and evidence: GitHub. Contract: 0xb89e2dB363CEC3827a9a7dbBd0CCCe7E75Da538a. Buy cover.