Paper
Technical whitepaper
A formal model of Quiver as a control plane between probabilistic agents and deterministic financial rails: what the interface makes true, what it only bounds, and what has to be measured against live markets.
A Control Plane for Autonomous On-Chain Agents is the working paper behind the system these docs describe. It treats an agent as a constrained state machine and separates the properties that follow from the interface — confirmation gates, idempotent launches, whitelist-bounded transfers, budget-bounded reasoning, deterministic fee accounting — from the ones that stay empirical: model quality, liquidity, price impact, venue liveness, and whether any of it is profitable. 16 pages, working paper v1.0, 19 September 2026.
How to read a claim
Every statement in the paper carries one of three labels, and they are not interchangeable.
| Label | Meaning | Example |
|---|---|---|
| Invariant | Follows from a correctly implemented interface or state transition. | A transfer destination is in the agent whitelist. |
| Bound | A maximum derived from explicit configuration. | Hosted reasoning spend cannot exceed the run budget plus metering granularity. |
| Criterion | Must be measured against live markets or services. | Quoted output stays above a minimum after fees and slippage. |
A transaction can be safely authorized and still economically poor; a profitable strategy can still hold unsafe authority. The paper argues the two axes should be judged separately.
Contents
- System objective and boundaries — the trust boundary, and why model output is data rather than authority.
- Architecture and state model — the agent tuple, the four capability surfaces (MCP, REST, hosted runtime, automation worker), and the venue abstraction over Pons, V4, and V3.
- Wallet authority and custody — destination reachability under a whitelist, transfer-at-risk, and what a whitelist deliberately does not cap.
- Intent, confirmation, and idempotency — binding a confirmation to a canonical action digest, and duplicate suppression under an Idempotency-Key.
- Execution and routing mathematics — quote decomposition, price impact, the slippage floor, route selection, and quote freshness.
- Pons launches and creator-fee economics — the bonding-curve integral, fee accrual as a function of turnover, and a break-even criterion.
- Autonomous runs and budget bounds — loop semantics, metering, and why a credit budget is not an on-chain notional cap.
- Automations as timed state machines — DCA, limit, and sniper bounds under at-least-once workers.
- Marketplace transfer semantics — atomic sale, credit conservation, snapshot ambiguity, and seller-era authority.
- Reliability, observability, and failure domains — staged success, the identifiers each layer owes, and the reconciliation equation.
- Security model and risk register — threats, controls, and the residual risk each control leaves behind.
- Operating criteria and verification plan — pass conditions and economic quality metrics a deployment should publish.
- Limitations and conclusions, plus notation and public sources.
A few results
- Duplicate suppression. If the key-to-request mapping is durable and checked before side effects, any number of retries with the same idempotency key produces at most one launch.
- Execution floor. If a settlement reverts below
y_min = y_quote(1 - s/10,000), a successful swap returns at leasty_min— bounding slippage against the quote, not against fair value. - Turnover dependence. Under a constant fee rate and share, cumulative creator fees are
F = alpha * tau * V: revenue tracks eligible turnover, not price direction. - Run spend. With pre-debited reservations, hosted reasoning spend stays within the budget; charge after the fact and the bound loosens by one step.
What the paper is not