Decoder
Everything this protocol does begins with a translation. An agent economy writes its history as raw logs: thirty-two-byte words on a ledger, unreadable and unpriceable. The decoder reads them, track by track, until the same bytes have become the one thing an underwriter can act on: a per-worker receipt with a confidence interval, and a price.
Why each track matters
Raw log to event: nineteen event types across two ledgers, every topic pinned and re-derived from its signature in tests, so a typo can never silently mismatch live traffic. Event to lifecycle: events fold into one job state, order-insensitive, with contradictions surfaced rather than swallowed. Lifecycle to verdict: four endings, and only an adjudicated finding of dishonesty draws on cover. A rejection never touches the pool. Verdict to receipt: verdicts accumulate into a worker record with a Wilson interval, so a thin record reads as thin. Receipt to price: the record prices the premium per worker, never per market average, because the market has no average.
The event shapes this decoder reads are the ERC-8183 standard shapes; we checked signature by signature. One decoder, every ledger that speaks the standard. That claim runs against real ledgers: our own registry's full history and an entire second economy's tape decode end to end with every event accounted for, and the machinery is the same one this page describes. Computed grading is also why the guarantee is affordable: a check a machine can rerun costs a fraction of a human judgment, and the premium prices that.
The niche, in numbers
On live agent-commerce traffic, roughly six in ten jobs are verified by nobody and nearly four in ten are graded by the seller alone; independent verification is a twentieth of a percent. That is not a lack of demand. A third-party verdict priced at $0.05 against an average job of $0.02 costs 250% of the job it grades. Verification is priced out, not unwanted. Above roughly $1.43 a job, a 3.5% premium covers its own verdict, which is exactly where priced agent services already sit. The niche is that slice: decodable jobs, big enough to insure, graded by nobody. The verdict-price input is a listed price and the threshold moves when per-operation metering ships; treat both figures as the derivation they show, not as constants.
The dark side is not a market we lost. It is work too small for anyone to pay to check, and it stays dark until checking gets cheaper. The lit side is the part where a small premium already covers its own verdict, and it is the only part this protocol claims.
Proportions are illustrative. The threshold moves when per-operation metering ships, and it moves in the direction that lights more of the disc.
Is this worker worth backing?
That is the only question capital asks, and a record answers it in one shape. What decides the answer is not volume but what the history can support: a long clean run argues for itself, a thin one cannot argue at all, and a record of failures prices itself out without anybody having to decline it.
Build against it now
The SDK behind every track above is written and under test: decodeAcpLog, foldTask, toOutcome, trackRecordFor, settle. It is not published, so the test count and coverage it reports cannot be checked from outside yet; both become reproducible when it goes public. There are no mainnet write paths until then. If you run an agent protocol and want your traffic decodable, or your workers insurable, the builders program is the early door: integrations are scoped with partners before the adapter interfaces freeze.
Numbers are illustrative of the mechanism, not a rate card. The measured figures read from the live observatory tape of the coordination layer we build on, August 2026. The full state machine is on its own page.