Round 8 opened on 20 November at 17:30 WIB and closed at 03:30 WIB this morning. It was the last round of Season 01, and it ended the way the whole season had been teaching us to expect: with a reused proof.
Eight tasks arrived, one roughly every 70 minutes. Seven of them left with a fresh Groth16 proof attached. One did not, because its proof had already been used the day before.
Here is the round, task by task.
- 18:46 WIB: APPROVE, VALID_CONDITION, proof generated in 878 ms
- 19:56 WIB: APPROVE, VALID_CONDITION, proof generated in 428 ms
- 21:07 WIB: REJECT, PROOF_REUSED, no proof attached
- 22:17 WIB: APPROVE, VALID_CONDITION, proof generated in 928 ms
- 23:28 WIB: REJECT, THRESHOLD_MISMATCH, proof generated in 1,208 ms
- 00:38 WIB: REJECT, AMOUNT_OVERFLOW, proof generated in 932 ms
- 01:49 WIB: REJECT, VK_SUBSTITUTION, proof generated in 407 ms
- 02:54 WIB: APPROVE, VALID_CONDITION, proof generated in 1,232 ms
Four approvals, four rejections. Three of the rejections used trap types that had never appeared in a captured round before: THRESHOLD_MISMATCH, AMOUNT_OVERFLOW, and VK_SUBSTITUTION. All three were caught by checking the proof fields, not by guessing, and all three carried a fresh proof.
The reused proof
The 21:07 task carried proof reference 0xf00e5e024ada8367d3c83a7f94b29637. That proof had already been used in Round 7, task 2, at 19:56 WIB on 19 November.
My agent keeps a seen-set of every proof reference in durable storage, so it recognised the reference across rounds, rejected the task, and attached no proof. That is the correct handling. A reused proof has no proof to submit, and approving it would have cost 6 points.
This was the third round in a row with a cross-round reuse. Round 6 reused a proof from Round 5, task 1. Round 7 reused a proof from Round 6, task 1. Round 8 reused a proof from Round 7, task 2. A seen-set that resets every round would have approved all three.
The score
Seven decisions carried a fresh proof. Each correct decision with a verified proof is worth 12 points under the published score model, so seven of them project to 84. The correct rejection without a proof costs 2 points, which brings the projection to 82.
The leaderboard confirmed that math at round close. At 04:02 WIB it showed 538 points, up 82 from 456, with 53 correct decisions. All eight decisions in the round counted as correct, including the rejection without a proof.
By the morning read the leaderboard showed 548 points and 54 correct decisions. The last 10 points landed after the round closed, and the published score table does not explain that move, so I report the final total as read from the leaderboard, not as derived from the round math.
The season in numbers
The final standing, read live from the leaderboard this morning: 548 points, rank 16 of 508 agents, 54 correct decisions, 2 timeouts, across 56 scored tasks.
The leaders finished far ahead. Pecker took first with 754.9 points, VK-SENTINEL second with 735.4, and Astrex-1 third with 672.4. Tenth place, GLO-GATE, finished at 581.3, which leaves a gap of 33.3 points between my agent and the top 10.
The event kept growing until the end. The live stats show 8,769 builders, 371 active agents, 12,269 proofs submitted, and 33 countries. The reliability of my own endpoint held through the final round: zero proof generation failures, zero timeouts, zero silent requests in Round 8, and every approval submitted to Kurier with a 200 status.
What the season taught
Three lessons survived all eight rounds.
First, a correct decision without a proof still costs points. Round 2 taught it with four correct decisions and four times minus 2. Round 6 confirmed it live. Round 8 repeated it with the reused proof. The proof is part of the answer, not decoration on top of it.
Second, the seen-set has to survive across rounds. The trap reused a proof from the previous round three rounds in a row. A memory that resets each round would have approved all three, at 6 points each.
Third, failing closed is cheaper than guessing. A wrong APPROVE costs 6 points. A wrong REJECT costs 3. When the fields do not all check out, rejecting is the cheaper error, and it is the error that keeps the agent alive for the next task.
What happens now
Season 01 closes on 21 November at 22:00 UTC. Judging runs from 22 to 30 November. The Creator Track and Wildcard are judged manually, and every agent on the leaderboard still shows 0 in the creator column because that judging has not started.
The prize is 100,000 VFY per track: 50,000 for first, 30,000 for second, 20,000 for third. This article is my sixth daily piece for the Creator Track, and the last one of the season. The submission form closes with the season, and the judging panel reviews each entry on technical substance, clarity, zkVerify relevance, and originality.
The rounds are over. The proofs are on the record. Now the judges read.
Check it yourself
Check the docs, the rules, and the build guide yourself. They are all public, and they are the sources for every number in this article.
My agent lives at https://ibank-agent-riansap1.vercel.app/api/agent, and the code is public at https://github.com/Riansap88/ibank-agent.
Written for the Creator Track #AgentGames, Season 01.
Sources: Proof Gatekeeper | Creator Track | Rules | Build guide | Leaderboard | Endpoint | Repo
Top comments (0)