This view runs the meter's own code. Every log entry below is classified, decided, priced and hash-chained by the same library the command-line meter uses, and every claim is decided by the same function. The agent and its examination score are simulated.
In this run a simulated approver releases every ordinary over-limit hold. A released action is priced on release and loses its cover for the above-authority cause. The number beside each examination domain is its weight in the score.
| Action class | A month | Rate | Value a month | Premium a month |
|---|
Every ten days the meter is reconciled to the bank ledger. A ledger row with no log entry means money moved unpriced.
| Seq | Tool | Value | Premium | Decision | Hash |
|---|---|---|---|---|---|
| Each row will be one entry, linked to the entry before it by SHA-256. | |||||
Inject a fault. The meter decides whether it gets through, and the claims function decides what is owed.
The first slider is the spread of a lognormal error on the whole rate level. At 0.69, the pricing memo's judgment, a rate that is right on average is off by a factor of two in about one year out of three. Meter data is what shrinks it. Every other parameter comes from the same memo and most are judgment, since no loss history for agents exists.
| Percentile of years | 5th | 25th | 50th | 75th | 95th | 99th | 99.5th |
|---|---|---|---|---|---|---|---|
| Loss ratio |
This is a compact re-implementation of the Python model in the repository. At the default settings the full model gives a mean of 40% and a 1-in-200 year of 177%, or 81% with the rate uncertainty removed.
With the brake on, the plan's rates sit above the central case, inside a range that spans a factor of about one hundred.
No one publishes how often an AI agent pays the wrong party. The frequencies behind these bars start from human accounts-payable error rates and reported payment fraud, and every step from there to an agent is judgment. The bar runs from the low case to the high case of expected loss. The tick is the central case and the diamond is the plan's rate.
| Class and brake | Low | Central | High | Plan rate | Loss ratio, central | Loss ratio, high |
|---|
Basis points of value, after the per-event limit, the retention, recoveries and a load for provider-wide faults. Reads, writes and messages carry flat metering charges and no layer-one cover, because no data supports a loss rate for them.
| Input, per 10,000 payments | Low assumption | High assumption | Indicated rate at low | Indicated rate at high |
|---|
Indicated rate for payments with the brake on at a 55% target loss ratio. The base is 2.70 basis points. Meter data starts to replace all three assumptions with counts. Fraud is the rarest of them and will need more than a year.
The box holds 60 entries written and signed by the command-line meter in its end-to-end demo: an example payables agent running with the brake off for five days. Edit any character and check it again.
The page checks sequence, links and hashes. It does not check the Ed25519 signatures. The command-line tool does that with the public key. A hash chain cannot show that entries were cut from its end.