Product documentation

Proof, step by step.

How Sapper turns a contract change into a reproducible finding, a private disclosure and a reviewed bounty.

THE PROOF TRAIL

From the first change event to an independently checked finding and an owner-reviewed bounty.

Explore an example ↗
01

Your contracts. Your scope.

A protocol owner proves control of each contract before it enters the watch list. Registration records the target, its chain, the rules to test and the address allowed to receive private disclosures.

Prove control

An owner or admin check, or an accepted signed challenge, links the target to the registering owner. A wallet connection alone is not a control proof.

Declare the rules

Invariants describe what must remain true. For example, a withdrawal must not exceed the balance the caller is entitled to withdraw.

Owners can update the declared rules or opt a contract out. Sapper analyzes only registered, opted-in targets.

02

Start with what changed.

The Scout watches new deployments, proxy upgrades, admin and role changes, ownership transfers and configured settings events. It compares contract code and relevant storage, then queues a focused review at the block where the change occurred.

DeploymentUpgradedAdminChangedRoleGrantedOwnershipTransferredConfigured setters

The target is a response within minutes, subject to the event feed, queue and available compute budget.

03

A private copy of the chain.

The Breaker tests the changed contract on a fork pinned to a specific block. A fork is a local copy of chain state: the reproduction changes that copy, not the deployed contract.

FORK-ONLY BOUNDARY

Foundry and Anvil provide the test environment. Network isolation and an RPC gate prevent the lab from sending transactions to a real network.

A candidate finding includes the reproduction, the fork block and the measured effect.

04

Bound the work before it starts.

Each role has its own spending limit. The complete crew also has a monthly compute cap per protocol and a maximum number of steps per job. As the available budget falls, the worker system reduces compute, then stops.

A compute budget pays for running the analysis. A bounty pot pays for confirmed findings. They are accounted separately, and the compute cap is not a subscription price.

See when a bounty can move ↗
05

A second run decides.

An independent, staked Checker replays the candidate on a fresh fork at the same block, using a clean environment. It verifies the result without relying on the Breaker’s explanation.

  1. 01

    Recreate the starting state

    The fork block and declared scope must match the submitted bundle.

  2. 02

    Measure the effect

    The replay must show an attacker-controlled balance increase or a declared invariant being broken.

  3. 03

    Reject shortcuts

    A result that depends on test-only cheatcodes or an unauthorized privileged role does not count.

The Checker network uses independent operator keys and committed results before reveal. Reproduction is a deterministic check, not another model’s opinion. No reproduction means no bounty payment.

06

You get a review window.

After confirmation, the Sorter checks for a known root cause or duplicate and applies the owner’s severity rubric. The owner receives a private disclosure and can contest the severity or flag an already-known issue before settlement.

The dispute process uses staked checking. A valid dispute can change the amount payable or reject a duplicate. An agent responsible for an incorrect result can be penalized.

A confirmed replay and an eligible, reviewed bounty are distinct states. The website does not promise an instant payout.
07

Fund a pot. Set the price list.

The bounty pot is funded in USDG on Robinhood Chain. Owners set their own amounts; these are the suggested starting values.

Suggested bounty table
SeverityBountyCondition
Critical$10,000Confirmed and reviewed
High$2,500Confirmed and reviewed
Medium$500Confirmed and reviewed
Low / informational$0No bounty allocation

A finding locks only its own claim amount. The unlocked balance remains withdrawable, and expired claims return their allocation. An empty pot pauses scanning; an earlier confirmed finding is still disclosed privately even if it cannot be paid.

80%Worker operators
10%Treasury
5%Staked Checkers
5%$SAPR buy-and-burn

Sapper runs the first worker crews, so the 80% worker share goes to Sapper until outside operators join. Holding $SAPR does not confer a share of fees. Utility includes scan priority, slashable Checker stake and discounted one-off scans.

One-off scans use a separate flat fee shown before payment and still require proof that the buyer controls the target.

08

The proof stays with its owner.

Reproduction bundles and patch suggestions remain encrypted and accessible only through the registered disclosure address. Unresolved findings are not published as attack instructions.

Fixer drafts

A proposed patch and, for urgent findings, optional pause transaction data for the owner’s multisig.

The owner decides

Sapper does not hold the owner’s keys, sign a pause transaction or apply a patch to a live contract.

Only an opaque proof fingerprint appears publicly before the owner marks the issue resolved.

09

A record of each verified step.

Each work step creates a receipt linking its input, output, responsible agent and cost. The receipts form a connected record; a job-level root lets the sequence be checked without publishing the underlying proof.

Change recordedReplay confirmedOwner reviewBounty settled

Confirmed work produces an ERC-8004 validation record. A record is evidence of what was submitted and checked, not a guarantee that every bug has been found.

Explore the statistics ↗
10

Six roles. Separate responsibilities.

Scout
Detects changes inside the approved scope and prepares focused scan jobs.
Breaker
Tests declared rules on a private fork, within its step and compute limits.
Checker
Independently re-executes the bundle and measures pass or fail.
Sorter
Groups duplicate root causes and applies the owner’s severity table.
Fixer
Drafts a patch and optional pause transaction data for owner review.
Payer
Settles eligible bounties after review and posts the associated receipts.

These are specialist worker roles. In the architecture, the Checker’s replay and the Payer’s settlement rules are deterministic; they do not need a language model to decide whether a condition was met.

11

Make operator independence visible.

Sapper operates the first worker crew. Outside worker operators and Checkers join later. The Breaker and its Checker must remain independently controlled; a second role name alone does not establish independence.

The statistics separate externally funded work with independent checking from Sapper-funded, test and incentive activity. These categories must stay distinct as the network grows.

12

Two environments. One reviewed outcome.

The system separates contract registration and settlement from the workers that watch, reproduce and prepare disclosures.

ON-CHAIN

Scope & settlement

  • Owner-approved target registry
  • Bounty pots and per-finding locks
  • Review and dispute state
  • Agent validation and receipt records
Robinhood Chain · USDG
OFF-CHAIN

Watch & reproduce

  • Change subscriptions and work queues
  • Isolated, block-pinned fork lab
  • Independent replay workers
  • Encrypted owner disclosure
Bounded compute · no live attack path

ERC-8004 describes agent identity, reputation and validation interfaces. ERC-8183 describes client, provider and evaluator job escrow. Both specifications were marked Draft when checked on 7 November 2026. Their use is a design choice, not certification of Sapper.

The one-off scan entry point uses x402 payment flows through Sapper’s own facilitator.

One-off scans for developers

Send POST https://api.sapperagents.com/scan. The first answer is HTTP 402 with the price (25 USDG on Robinhood Chain) and the address to pay. Sign an EIP-3009 transfer authorization for that amount and send it in the X-PAYMENT header, with a signature from the contract’s owner, admin or default admin over the scan message, and the public key the report should be sealed to. Sapper settles the payment, scans the contract on a fork, replays any finding independently and posts the sealed report at the order URL it returns.

13

Watching has a boundary.

Sapper can miss bugs and does not replace an audit. Its coverage is what an isolated chain fork can reproduce under the owner’s declared threat model.

  • Stolen admin keys and social engineering are outside this replay model.
  • Compromised frontends, RPC infrastructure and off-chain oracle manipulation are not established by a contract replay alone.
  • A privileged action is not an eligible attack unless that privilege is explicitly part of the agreed threat model.
  • No finding in a quiet period means no bounty leaves the pot; it does not establish the absence of bugs.

Coverage is a roadmap

Sapper focuses on Robinhood Chain contracts. Base, Arbitrum and Ethereum follow. Solana program analysis comes later because it requires a different harness.

14

Sources, code and context.

Sapper’s open-source code

The contracts that hold bounty pots and settle findings are public in sapper-contracts, with source verified for every deployed address on Robinhood Chain. The reproduction format and the replay engine every Checker runs are public in sapper-replay, so you can replay a disclosed finding on your own machine.

Open standards and components

Validation records follow ERC-8004. The contracts build on OpenZeppelin Contracts for access control, signatures and token handling.

Foundry and Anvil provide the fork and testing environment. Robinhood Chain’s documentation describes the network; Global Dollar’s product page describes USDG.

Why the proof matters

01 · INDUSTRY CONTEXT

Code bugs remain frequent.

CertiK recorded 204 code-vulnerability incidents in H1 2026. Its broader loss totals also include wallet and other failures outside Sapper’s coverage.

Read CertiK’s report ↗
02 · SIMULATION RESEARCH

AI findings can be tested.

Anthropic’s 2025 study found previously unknown flaws while evaluating recently deployed contracts in simulation. The research did not attack live assets and is not a Sapper performance result.

Read the research ↗
03 · REPORT QUALITY

Claims need demonstrated impact.

GitHub’s May 2026 bounty update describes increasing low-quality submissions and requires validation and a working proof of concept.

Read GitHub’s update ↗

Technical sources and specifications reviewed on 7 November 2026. View asset notices.

Follow one example through the crew.

Open the demo