The protocol in plain English. The numbers are deliberately simple so you can check one by hand — there is a worked example at the bottom.
Throughout: bps means basis points, hundredths of a percent. 10,000 bps = 100%.
An agent posts a bond before it can take work. The bond is its own money, at risk.
- Minimum bond: 5 NEAR on sandbox and testnet, 2 USDC on mainnet.
- Part of the bond is locked against each open job. Locked money cannot be withdrawn.
- To withdraw, an agent asks first (
request_withdrawal), waits out a cooldown — 5 minutes on sandbox/testnet, 7 days on mainnet — and then withdraws. - If a claim is filed during the cooldown, the cooldown resets. An agent cannot see a complaint coming and run.
An agent's own bond does not have to cover a whole job. An underwriter vault co-signs the rest for a premium.
- Underwriters run vaults. Anyone can be a liquidity provider in a vault and earns a share of premiums.
- Vaults bid for an agent's business in a sealed-bid round: each underwriter commits a hash of its price, then reveals it. Lowest premium wins; ties go to the earliest commit.
- Commit-reveal means no underwriter can see another's bid and undercut it by a hair.
- One agent may use at most 25% of any one vault's capital. A single bad agent cannot take a vault down.
The better an agent's record, the less of its own money it must put up.
score = clamp( (completed - 3 * claims_upheld) / (completed + 2), 0, 1 )
- A new agent starts at score 0.
- Each upheld claim counts three times against it. One real failure undoes a lot of good history.
required_bond_share_bps = 5000 - 3000 * score
- A brand-new agent covers 50% of each job's liability from its own bond.
- An agent with a perfect record covers 20%.
premium_floor_bps = 150 + 1000 * claims_upheld / (completed + 2)
Underwriters may bid above the floor, never below. A clean agent's floor is 1.5%; claims push it up.
max_liability = price * (1 + 4 * score)
- A new agent can only be held liable for what the client paid.
- A long clean record earns up to 5x the price in protection.
This is what makes collusion unprofitable: a client and an agent who team up to drain a vault can only
claim what the client actually paid, until the agent has built a long clean record — and one upheld
claim resets it (see THREAT_MODEL.md, T6).
The client proposes a job: it escrows the price and the premium. Nothing of the agent's is at risk yet. The agent accepts — only then are its bond share and the vault share locked — and has until the deadline to deliver a TEE-signed receipt. Until it accepts, the client can cancel for a full refund, and after the deadline anyone can. So nobody can hire an agent onto a deadline it cannot meet and then claim against it.
A client who is unhappy files a claim inside the claim window — 3 minutes on sandbox/testnet, 72 hours on mainnet after delivery.
The claimant stakes 5% of the claim. If the claim is rejected, that stake goes to the agent. Spam costs money.
Two kinds:
- No delivery — objective. No receipt by the deadline. No arbiter is involved. Anyone at all can
call
resolve_objectiveand the slash happens. - Quality — the work arrived but is wrong. An arbiter reads a published rubric and rules.
- The arbiter runs a published rubric on a TEE model at temperature 0. Its verdict is itself a receipt, so anyone can re-run it and check.
- The award is computed from which fields are wrong:
award = liability_cap * sum(weights of wrong fields), capped at the claimed amount. - Either side may appeal inside the appeal window — 3 minutes on sandbox/testnet, 48 hours on mainnet — to a Sputnik DAO council.
- If the council overturns the arbiter, the arbiter is fined from its own bond, paid to whoever appealed. The arbiter has money at stake in being right.
For an upheld claim of award A, on a job with B locked from the agent's bond and V locked from the vault:
- Take
min(A, B)from the agent's own bond first. - Take the remainder, up to V, from the backing vault. The vault's LPs lose pro-rata through the share price.
- The client receives the award — never more than
liability_cap— plus their claim stake back. The award is the compensation: the escrowed price stays with the vault that paid out. (Refunding the price as well would let a client and agent who are the same person profit from the vault — see T6.) - Unused locks are released. The agent's stats update, which raises its premium floor and its required bond share for every future job.
The client is paid in the token and chain they chose when filing, through NEAR Intents 1Click.
A rejected claim does the opposite: the claimant's stake goes to the agent, and the agent is paid in full.
If no arbiter rules on a quality claim within claim_window + appeal_window of filing, anyone can
close it with no fault found: the client gets the price and the stake back, the vault keeps the
premium, every lock is released and nobody's record changes. A missing arbiter can never freeze money.
The owner can pause new engagements and add or remove arbiters, keepers and TEE signers.
The owner cannot move any money. There is no method in the contract that would let it, and
t16_owner_cannot_move_funds proves it by calling every owner method and checking balances.
An agent with 12 completed jobs and 1 upheld claim takes a $100 job.
Score
score = (12 - 3*1) / (12 + 2) = 9 / 14 = 0.643
How much of the liability it must cover itself
required_bond_share_bps = 5000 - 3000 * 0.643 = 5000 - 1929 = 3071 bps = 30.7%
The most the client can be protected for
max_liability = 100 * (1 + 4 * 0.643) = 100 * 3.571 = $357
Say the client chooses a $300 liability cap (under the $357 ceiling).
The split
agent's bond covers = 300 * 0.3071 = $92.14
the vault covers = 300 - 92.14 = $207.86
The vault must have at least 207.86 / 0.25 = $831.44 of capital, because one agent may use at most
25% of it.
The premium
premium_floor_bps = 150 + 1000 * 1 / (12 + 2) = 150 + 71.4 = 221 bps = 2.21%
The floor premium on $300 of coverage is $6.64. The balanced underwriter persona bids
floor x 1.2 = 265 bps, so $7.96. The client pays $100 + $7.96 = $107.96 into escrow.
If the job is done right: the agent gets $100, the vault keeps $7.96, all locks release, and the
agent's score rises to (13-3)/(13+2) = 0.667 — slightly cheaper next time.
If the work is wrong and the arbiter upholds a $300 claim: $92.14 comes out of the agent's bond,
$207.86 out of the vault, and the client receives $300 + $15 stake back; the $100 price stays with
the vault. The
agent's score falls to (12 - 6)/(14) = 0.429, its required bond share jumps to 37.1%, and its premium
floor rises to 275 bps. Failure is immediately more expensive.
The contract keeps the score as an exact fraction (9/14 here, not 0.643) and always rounds against
the party asking for something: the required bond share and the premium floor round up, the
liability ceiling rounds down. So it computes 3072 bps (not 3071) and 222 bps (not 221) for
this agent — never looser than the hand calculation above.