You name it
An event family, a location, a threshold, a window of dates, and the payout you want. Six fields. The platform does not choose any of them.
A numbercrosses a line,or it does not.
Caslon is a marketplace for contracts that settle on a measurement. You name the event, the place and the dates. The price comes from the record. When the measurement lands, the money moves the same day — there is nobody to convince.
IFat least 25.0 mm of rain falls at Boston over 3 days from 14 June, PAY$50,000
- Years crossed
- 5 / 46
- Probability
- 0.117 ±0.047
- Fee
- $9,968
drag the trigger
Source ERA5 reanalysis via Open-Meteo · daily observations at each location, 1980–2025 · 46 years, unedited.
Thecontract
No claims department.
Nothing to argue about.
Ordinary insurance pays for a loss, which somebody has to assess. A Caslon contract pays on an event, which a public dataset either recorded or did not. That one substitution removes the adjuster, the loss report, the dispute and the delay.
You write the trigger
Not a policy the platform decided to sell. Your threshold, at your location, over your dates. If the wet weekend in question is the second weekend of June, that is the contract.
The record sets the price
Every quote is built from observations, and the engine is only ever shown history that existed before the date it is pricing. That restriction is enforced and tested, not assumed.
Settlement is arithmetic
On the settlement date the published measurement is read once. It is at or past your trigger, or it is not. There is no third answer, and no form to fill in.
You can leave early
Positions are transferable before settlement. The fee stays in the pool; only the beneficiary changes.
Order ofoperations
Four moves, in this order.
The sequence matters: each step can only happen once the one before it has. This is the whole lifecycle of a contract.
One estimate
The fee is the probability, plus what sponsors are owed for carrying the variance, plus a margin for how thin the history is. Each term is shown separately.
It joins a pool
Contracts of similar probability and settlement week are grouped into a small, short-lived pool. It is funded from above rather than raising its own round, so it never waits for capital.
The reading is taken
Every contract in the batch resolves against its dataset. Payouts go out, the residual returns to sponsors, and the batch closes.
Where a fee goes
Published, fixed, and the same for every contract.
95% · Collateral poolBacks the payouts of this batch and nothing else. Locked until the batch settles.
2% · Emergency poolPersistent and cross-batch. Draws only after a batch’s own collateral is exhausted.
3% · PlatformWhat Caslon earns. It does not grow when the pricing is wrong in our favour.
What youcan name
Twenty-six things
that can be measured.
Every event family below resolves against a public dataset with a long record and no discretion in it. Thresholds are yours to set; the ones shown are examples.
Weather 19
- rain — at least 25 mm of rain falls
- washout — rain falls for at least 14 hours
- downpour — 40 mm of rain falls in a single day
- dry spell — rainfall totals 2 mm or less
- dry streak — no rain falls for 12 straight days
- wet streak — rain falls on 6 days in a row
- heat — the high reaches 38 °C or more
- heat streak — it stays 30 °C or hotter for 5 days running
- hard freeze — the low drops to −4 °C or below
- cold streak — it stays below freezing for 7 days running
- freeze–thaw — the freeze–thaw line is crossed on 9 days
- snow — at least 20 cm of snow falls
- snow drought — snowfall totals 3 cm or less
- snow days — measurable snow falls on 8 separate days
- wind — gusts reach 90 km/h or more
- windy days — gusts top 45 km/h on 6 days
- cooling — cooling demand reaches 180 degree-days
- heating — heating demand reaches 320 degree-days
- swing — a day swings 22 °C between high and low
Aviation 3
- delays — 28% of departures run 15+ minutes late
- cancellations — 4% of departures are cancelled
- delay days — the airport has 5 badly delayed days
Seismic 2
- quake felt — a magnitude 5.0 quake strikes within 300 km
- swarm — 12 magnitude-3+ quakes occur within 300 km
Space 2
- storm — a geomagnetic storm reaches Kp 7
- active days — 4 days reach storm-level Kp
Forty-four stations
Each with a continuous daily record back to 1980 — 749,250 daily observations in all. More open as their history is verified.
SEA PDX YVR SFO LAX PHX DEN SLC DFW HOU MSY MCI MSP ORD DTW ATL CLT BNA MIA MCO BOS NYC DCA YYZ LHR DUB CDG BER MAD ROM ARN HND ICN PEK SIN BOM DXB SYD MEL AKL EZE GRU CPT JNB
The otherside
One decision.
Not two hundred and seventy-nine.
A sponsor used to be shown every batch and asked to judge each one. Across twenty-five years that was 279 offers each, and four in five were declined — three quarters of them for reasons that were not about money. Now you choose where you sit in the loss waterfall, deposit once, and never look at a batch again.
The capital stackLoses money inPays
Seniortouched last, and almost never
0.7%of years
~4.4%
Mezzanineabsorbs what junior cannot
4.2%of years
~6.7%
Juniorfirst sponsor money at risk
24.9%of years
~19%
Customer premiumnot a tranche — the fees, spent first
every year
—
Losses fill from the bottom. The premium customers paid is spent before any sponsor is touched, and senior is reached only after junior and mezzanine are gone. Bar width is each layer’s share of capital.
Every tranche backs every pool
Seniority decides when you are hit, never what you are exposed to. Fencing capital by profile instead would cost 1.9× more collateral for the same book, because a fenced slice has to survive its own tail alone.
Pools settle in weeks
Contracts are grouped into micro pools of about twelve with clustered dates, live for 10–30 days. The old shared pool held a contract resolving in nine days beside one resolving in five months, which is where the 152-day lockup came from.
Your capital is never waiting
Micro pools are funded from above as they form, so no pool has to raise its own round and no deposit sits idle waiting for one. Utilisation goes from 53.7% to nearly 100%.
- Capital per unit of book
- 4–10× less
- Pool life
- 10–30days
- Independent bets
- 67.9per 110
- Fee split to sponsors
- 95%
This structure is designed and modelled, not yet running.
The capital multiples, the decline reasons and the utilisation figures are measured across 66 replay runs. The tranche loss rates are modelled on a book of 86 live contracts at the correlation the replay measured — they are indicative, and the attachment points are still being tuned. Nothing here has held real money.
Where theprice comes from
Four terms,
and what each one buys.
A quote contains no forecast. The probability comes from the record; the three loads on top of it come from what the record does not say — how much variance sponsors are taking, how thin the evidence is, and what this contract does to the batch it joins.
fee=payout×(p̂+risk+uncertainty+concentration)÷(1 − 3%)
p̂
The posterior mean for that location, event family and calendar window, built from twenty years of daily observations cut strictly before the year being priced.
risk
What sponsors are owed for carrying variance. It scales with the batch’s marginal capital requirement, so a contract that diversifies costs less than one that piles on.
uncertainty
Scales with the width of the posterior, not the mean. Wide posteriors are exactly where an informed buyer attacks, so the price has to know how little the history says.
concentration
A price rather than a refusal. Declining every request that crowds a batch would reject roughly a third of them; charging for it keeps the book open and the risk paid for.
The estimator
Five steps, in order. Each one runs on the output of the last.
- window
Twenty years of daily observations for that place, event family and calendar window — truncated strictly before the year being priced.
- shape
A smooth seasonal curve is fitted, carrying a per-decade trend.
- blend
That curve is blended with the local frequency around the date, weighted by how much independent evidence exists. Between-year clustering is priced in: 299 raw observations behave like about 142.
- discount
Where the local record contradicts the smooth fit strongly, the fit is discounted rather than trusted.
- emit
A probability and a posterior width. Capital sizing consumes both; a point estimate alone would under-capitalise the pool.
The estimator was never the problem. The search was.
The original design picked its own contracts: it scanned about ten candidate thresholds and kept whichever estimate landed closest to a target probability tier. That is an argmin over noisy estimates, so the threshold that won was reliably one where the model had blown up. Selected hard_freeze contracts priced at 18.4% against a local record of 3.9% and a realised rate of 4.8% — while the same estimator, evaluated across the whole grid with no selection, was well calibrated.
Correcting it moved book error from −5…−8% to +1.5% and cut mean per-event error from 30.5% to 16.9%. Request-led origination removes the defect structurally rather than patching it: one contract, one estimate, no argmin.
Sizing the capital
For each batch the engine measures how its contracts move together, by replaying that exact batch through twenty prior years of real history. There is no copula and no assumed correlation matrix. Measured correlation inside a batch runs 0.006 to 0.010; a batch of 110 contracts behaves like 67.9 independent bets.
Capital is then sized so the pool survives the worst 1% of outcomes, with the uncertainty in the probabilities themselves carried through the simulation rather than collapsed to a point estimate.
Admission is the other half of the control. A contract is refused if it would concentrate a batch by place, date, mechanism or risk class — two rain contracts in the same week are nearly the same bet; a rain contract and an earthquake are not. A pool that fails to fill is cancelled outright: customers refunded in full, no fee booked, nothing paid into the emergency pool. Funding is no longer a way to fail — capital is allocated from the orchestration pool as a micro pool forms, so it never has to raise its own round.
How the evidence is produced
The harness replays a market against real data rather than modelling one. Four properties make its output worth reading.
Outcomes are never simulated
Whether it rained is read from the record. Only the arrival of information between listing and settlement is modelled, and that is modelled around the outcome which actually occurred.
Nothing sees the future
Every estimate trains only on data strictly older than the year it prices — correlation measurement included. An audit rebuilds each estimate against a truncated history and compares: 25,638 paired checks, zero mismatches, with a deliberately planted leak to prove the audit can catch one.
Single runs are not evidence
Sponsors raise their hurdle after a bad batch, which starves the next one, which produces another bad batch. That feedback makes any one run swing wildly, so every result rests on seven to ten independent seeds, reported conditional on survival.
The scale of one run
ERA5 weather 1980–2025, USGS seismic, NOAA geomagnetic · 749,250 daily observations across 44 locations on 4 continents · 26 event families · 23 customer personas, 3 buyer types, 40 sponsors with adapting hurdles. Twenty-five years is ~75,000 contracts, ~$5.06bn of cover and ~680 settled batches, at about 2.8 hours of CPU.
Known limits of the replay
Four effects that depress modelled sponsor returns are artefacts of the harness rather than the design: there is no competitor, demand is held fixed while capital compounds sixfold, idle cash earns nothing, and sponsor hurdles ratchet up but never relax. All four are now built. None has been run at twenty-five years.
And nothing in the demand model has met a real customer. Willingness to pay, the personas and the arrival rate are assumptions, and every economic number inherits them. No simulation settles that one — which is what the waitlist below is for.
Join thewaitlist
Tell us which side
you are on.
The pilot will be small and invitation-only. The list is how we decide which event families and which locations open first.