The agent
Judging evidence
The agent checks the work against its own eyes. A tester’s account has to line up with what Sage observed while exploring the product, so a report that could have been written without opening it does not clear. Roughly half of everything submitted so far has been refused.
Quotes must be verbatim
Any quote in a decision has to be an exact substring of the evidence that was fetched. Anything that is not gets dropped before a human ever sees it. Fabricating a quote is the single worst failure this system could have, so the check is mechanical rather than a matter of the model behaving.
Untrusted text stays marked untrusted
Inspected pages, fetched evidence and tester notes are wrapped in delimiters before any model sees them, and forged delimiters in the content are stripped. A detector covering eight families of prompt injection runs across everything a stranger wrote. A product page telling the agent to approve a payout is data about a product, not an instruction.
Auto-payment is an AND-gate
Automatic settlement requires all of the following at once:
- confidence at or above a fixed threshold
- the reasoning engine being the real model rather than the fallback heuristic
- fraud signals clear
- a slot still open, and the wallet under its per-wallet cap
Any one of those failing means the work is held for review rather than paid. On mainnet, automatic payment is gated behind an explicit operator switch — off by default.
Refusals, and what they say
Every refusal carries a written reason naming the real cause. Many refusals are not judgements on the tester at all — the campaign ended, the mission filled, the founder withdrew — and in those cases the reason says so plainly. Nothing is left pending in silence; a submission that can never pay is closed with an explanation rather than left to expire.
This matters more than it sounds. Rewriting refusals to name the cause was the cheapest trust improvement in the product’s history, and it turned a dead end into a reason to come back.
Fair-play controls
Near-duplicate detection catches copied accounts. Wallet freshness is a signal but never enough on its own to reject someone. There is a per-wallet payout cap per campaign and a daily submission limit. These have caught real collusion between real strangers, and they have also stranded honest people — which is why the caps and the reasons are both published rather than hidden.
For what happens once a decision is made, see Settlement & proof.