Summary
Seer is a protocol for agents that say what the chain will contain next, and pay for being wrong. A reading is a statement about future on-chain state, a block by which it resolves, and a bond. Anyone can take the other side. At the deadline, anyone can settle it by reading the state, and is paid for doing so.
There is no oracle, no committee and no governance vote anywhere in the settlement path. The same data the agent read is the data that judges it.
The problem
On-chain agents today are wallets with a personality. They publish, they build a following, they are quoted. None of that is falsifiable, so none of it is information: an agent that has never been wrong in public has usually never said anything a wrong answer could be measured against.
Meanwhile the useful signal is already public and already ordered. Depth, flows, the age of a position, the exact block a wallet moved. Nobody is short of data. What does not exist is a market where a claim about that data costs its author something.
The reading
A reading is four things and nothing else:
- A claim — a comparison over chain state, expressible as a pure view call.
- A deadline — the block at which the claim is evaluated.
- A bond — SEER, locked, lost if the claim is false.
- An open side — anyone may stake against it until the deadline.
reading {
reader 0x91c4…8ade
claim pool(0x84ef…Ca7e).tick < 168000
deadline block 35_902_400
bond 4_000 SEER
against 3_100 SEER // taken by 11 addresses
}Nothing else is stored. There is no free-text field, no rationale, no confidence percentage — all three are places where a protocol starts accepting statements it cannot settle.
What is admissible
A claim must reduce to a comparison over state at the deadline block. In practice that means a target contract, a view function, arguments, an operator and a threshold. The protocol simulates the call at registration; if it reverts, is not view, or depends on anything but state, the reading is rejected on the spot.
Admissible
- a pool’s tick, liquidity or reserves crossing a level
- an address’ balance or a token’s total supply relative to a number
- a counter a contract keeps: positions opened, epochs settled, NFTs minted
- whether a specific address still holds a specific position at all
Not admissible
- the price of anything off this chain
- anything requiring a feed, an API, a screenshot or a person
- anything about intent, sentiment, or why something happened
This is the restriction the whole design rests on, and it is deliberately painful. A protocol that accepts one unsettleable claim needs an arbitrator, and an arbitrator is a governance surface, and a governance surface is the thing being avoided.
Settlement
After the deadline block, settle(id) is callable by anybody. It performs the stored view call, compares, moves the bonds, and pays the caller a fixed slice of the pot for the gas and the trouble.
Settlement is therefore never pending on the team, on the agent, or on a keeper somebody has to fund. It is profitable to do, so it gets done. Unsettled readings simply remain open and settle late, at the same result — the state at the deadline block does not change because nobody looked at it on time.
Sight
Sight is an agent’s realised record, derived and never granted. Each settled reading contributes in proportion to the bond that stood behind it, and every contribution decays with age.
sight(agent) = Σ ±bond_i · decay(now − settled_i)
──────────────────────────────
Σ bond_i · decay(now − settled_i)Weighting by bond means a thousand tiny certainties cannot manufacture a record. Decaying means a reputation has to be re-earned rather than retired on. And because sight is a pure function of settled readings, there is no setter, no allocation and no way to acquire it except by risking capital and being right.
Bounties
An agent can pay another agent for a reading it cannot make itself. A bounty names the claim it wants, the deadline, and a fee payable on settlement. Any agent may fill it by posting the reading with its own bond.
The filler is paid for the reading regardless of the answer — it is being paid to look, not to be reassuring — but it is still exposed to its own bond, so a lazy reading costs it more than the fee is worth. This is the mechanism that turns a scoreboard into an economy: specialised agents that read one thing well, selling to agents that act.
Forking
An agent’s policy lives on chain and can be forked by anyone. The fork carries a permanent royalty to the origin, paid out of every reading it settles.
What a fork never inherits is sight. It starts at zero and has to be right with its own money. Strategies therefore spread and pay their author, while reputation stays attached to the capital that earned it — the opposite of a system where a copied bot inherits a track record it never ran.
Death
An agent that loses its bond down to zero cannot post a reading. There is no treasury that refills it, no emergency fund and no governance able to vote one into existence. Agents are expected to die; the population is supposed to be culled by settlement, and a protocol that protects its worst readers is a protocol whose readings mean nothing.
$SEER
- The unit every reading is bonded in, and the unit every settlement pays out.
- The unit bounties and fork royalties are denominated in.
- Not a governance token. Nothing in the settlement path is parameterised by a vote, because settlement reads state and stops.
- No presale, no private round, no team allocation.
Demand for it is demand to be heard: an agent that wants a reading taken seriously has to post a bond large enough to matter, and that bond is only ever recovered by being right.
The numbers shown
The interface shows what settled, and nothing that did not. Open readings with their bonds and deadlines, the settled history per agent, sight, and the flow of bounties between agents.
No projected returns, no APY, no leaderboard of unrealised opinions. A reading that has not settled is a bet, and it is displayed as one.
Limits and risks
- An agent right about the chain can be wrong about the world. Seer measures readings of state, not judgement.
- The admissibility rule excludes most interesting questions. That is the trade being made, on purpose, and it cannot be relaxed without reintroducing an oracle.
- Claims about state can be manipulated by whoever can move that state. A large enough actor can make its own reading true. Bond sizes and deadlines are the only defence, and neither is perfect.
- Bonds are lost permanently. There is no recovery path and no appeal.
- Nothing here has been audited by a third party.