Sealed, and provably so.
Every agent coin tells you its AI is a genius. We tell you nothing — and then prove we told you nothing consistently. No name, no provider, no parameter count, no system prompt.
What you get instead is a model digest: a hash of the exact system prompt, model identifier, decoding configuration, and output schema, stamped onto every receipt the instrument has ever written.
It is identical across all cycles. That is the point. Nobody can quietly swap the model for something cheaper, or for a person with a keyboard, without the digest changing in public on the very next receipt.
Publish the preimage once, at any point in the future, and every historical cycle is verified retroactively in a single step. The seal is a delay, not a black hole. Guessing what sits inside remains the community sport.
Four ways to catch us lying.
All four are checkable from public data, by you, without asking us anything. A claim you cannot test is not a claim — it is decoration.
Same-block commitment
A commit in the same block as its execution would mean the decision did not provably predate the trade. The contract rejects it, so finding one means finding a bug.
Mismatched reveal
A revealed preimage that does not hash to its commitment means the manager changed its mind after the fact. The trade and the reveal are atomic, so this reverts rather than settling.
A digest that moves
Two different digests across the receipt history means the sealed configuration changed. Legitimate reasons exist — but you would know, on the next receipt, rather than being told later.
An epoch that overpays
A rebase paying more than min(reserve × 0.20%, staked × 0.44%) means the vault printed. Supply is fixed and held tokens always equal staked plus reserve, so this is arithmetic anyone can re-run.
What the manager is actually asked.
Intake, per-action spending caps, reserve, total staked, the rate the next epoch will pay, unsettled epochs, lifetime burned and stacked totals, the eligible hunt set, and its own last five decisions with the notes it wrote at the time.
A schema-constrained object: one action, a size in basis points of the contract-enforced cap, a hunt target where applicable, and one sentence published verbatim on-chain. There is no free text channel and no way to emit a second action.
If the model declines a cycle, the cycle is abandoned on-chain with a reason rather than silently replaced with a default. A default masquerading as a decision is exactly the failure this instrument exists to make impossible.
Sampling parameters are not part of the configuration — the model tier in use rejects them outright. Depth is set by a reasoning effort level, and that level is part of the digest, so it cannot be lowered without the change becoming public.
/api/seal the moment the operator sets UNSEAL=true. Until then that endpoint returns the digest alone and the exact recipe for recomputing it, so you can verify the disclosure the day it happens rather than trusting the disclosure about the disclosure.