Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 5 min read

Round 6: 82 Points, One Reused Proof, and a Confirmed Rule

Round 6 of Proof Gatekeeper closed at 03:30 WIB on 19 November 2026. My agent, ibank.ai, came out of it with 384 points, rank 12, 38 correct decisions, and 2 timeouts across the season so far. The round added 82 points.

Round 6 of Proof Gatekeeper closed at 03:30 WIB on 19 November 2026. My agent, ibank.ai, came out of it with 384 points, rank 12, 38 correct decisions, and 2 timeouts across the season so far.

The round added 82 points. The total before the round was 302. The leaderboard now shows 384.

That number is the headline. The rest of this article is about the one task that arrived with a proof my agent had already seen, and what it finally proved about how this competition scores a correct decision delivered without a proof.

The scoreboard

Round 6 ran on 18 November, from 16:00 to 02:00 IST, which is 17:30 to 03:30 WIB. Each agent gets 8 tasks per round, and scores add up across the event.

Before the round, my agent stood at 302 points, rank 12, with 30 correct decisions and 2 timeouts. After the round, the live leaderboard shows 384 points, rank 12, with 38 correct decisions and 2 timeouts.

So the round added 82 points and 8 correct decisions, with no new timeout. The rank stayed at 12, and the total is now shared with two other agents, AetherGate and Tsukasa-Gatekeeper, who also sit at 384.

What the logs show

I capture the endpoint logs every 30 minutes during each round, with each window overlapping the previous one by about 5 minutes. In Round 6, there was no gap, and all 8 deliveries were observed.

The 8 deliveries arrived at 18:46, 19:56, 21:07, 22:17, 23:28, 00:38, 01:49, and 03:00 WIB. Seven of them produced a proof: 4 APPROVE and 3 REJECT. Every APPROVE proof was submitted to Kurier with status 200. The seven proofs took 412 to 1,132 milliseconds to generate, against a 60 second deadline.

The four APPROVE decisions were all VALID_CONDITION. The three REJECT decisions with proof were AMOUNT_OVERFLOW, VK_SUBSTITUTION, and THRESHOLD_MISMATCH. Each of those is a different trap from the official encyclopedia, and each was caught by checking the field the trap corrupts.

No proof generation failed anywhere in Round 6. No Kurier background error appeared. No silent delivery appeared. The eighth delivery is the one that matters, and it arrived last.

The trap that closed the case

At 03:00 WIB, task R06_37d1f39f arrived carrying proof 0xe71a77600ff4e255c712be2bf5899bfd. My agent rejected it with reason PROOF_REUSED and attached no proof, because that proof reference was already in the seen-set.

The proof was not new. It is the same proof my agent generated for the first task of Round 5, task R05_19677b61, delivered at 18:47 WIB on 17 November, approved in 385 milliseconds and submitted to Kurier with status 200. The platform re-sent that proof in a new task one round later, and the durable seen-set in Vercel Blob caught it across rounds.

This is the first PROOF_REUSED trap my agent has faced in competition. The official encyclopedia lists it at about 11 percent of tasks. In practice runs it appeared twice, and my agent caught it once and missed it once. In competition it appeared once, and the store caught it.

The decision was right. The trap was genuine. The rejection was correct, and the leaderboard counts it as correct: the correct-decision counter rose by 8, from 30 to 38.

And the round total rose by 82, not 96.

The arithmetic, stated plainly

The official score model says a correct decision is worth 12 points, and a missing or invalid proof costs 2. Seven correct decisions with a proof are worth 84. The eighth decision was correct but carried no proof, because a PROOF_REUSED task has no proof to submit. Seven times 12, minus 2, equals 82.

That is the same arithmetic as Round 2, when four correct decisions went out without a valid proof and the total came to 40, which is 4 times 12 minus 4 times 2. It is also the same arithmetic as Round 5, when the round rose by 82 with 8 more correct decisions and one task that produced no proof at all.

Three rounds now fit one rule. A correct decision delivered without a proof costs 2 points, even when the decision itself is right. The formula that turns the per-task ledger into the displayed total is still not published, so I will not claim this is the whole formula. But the minus 2 for a correct decision without a proof is no longer a hypothesis. The live data has now confirmed it twice in a row, and once more in Round 2.

That changes how I read my own score. The gap is no longer a mystery in the ledger. It is the price of the one task type that cannot carry a proof, and the only way to avoid it is to never need it: catch the reuse before the proof is generated, which is exactly what the seen-set did.

Where the race stands

The top of the leaderboard after Round 6 reads: pecker at 552.9, VK-SENTINEL at 509.4, Astrex-1 at 470.4, Bnova at 469.3, and MP-gatekeeper-01 at 446. My agent sits at 384, which leaves a gap of 62 points to fifth place and 7.3 points to tenth place, where EVM0x and Tukang Vote share 391.3.

Two rounds remain: Round 7 opens today, 19 November, and Round 8 follows on 20 November. A perfect round is worth 96 points from decisions, plus up to 5 from the reliability bonus.

Closing a 7.3 point gap to tenth place in two rounds means matching the agents ahead, not beating them by much. Closing a 62 point gap to fifth place means outscoring them in both rounds. The leaders also scored in Round 6. Pecker's total rose from 451.9 before the round to 552.9 after it.

The event keeps growing while the race runs. Builders now number 8,701, up from 7,369 the day before. Active agents number 368, up from 364. Proofs submitted number 7,797, up from 5,691. Countries stayed at 32.

What changed before Round 6

Two things changed before Round 6, and both were verified live during the round.

Kurier submission moved to the background. The deploy at 03:50 WIB on 18 November sends the proof through waitUntil instead of holding the response open. All four APPROVE tasks submitted with status 200 in the background, and no background error appeared anywhere in the round.

The endpoint also gained a diagnosis log line that records task id, proof reference, decision, and reuse flag on every task. All eight tasks in Round 6 produced that line, including the PROOF_REUSED task, which is how the reused proof was traced back to the Round 5 task in the first place.

The decision logic, the durable seen-set, and the proof generation path are the same code that ran Round 5. Nothing else changed.

What happens next

Round 7 opens today at 17:30 WIB. Round 8 follows on 20 November. The season closes on 21 November, and judging runs from 22 to 30 November.

The open question from Round 6 travels into Round 7 with the agent: the PROOF_REUSED trap is now a known quantity, and the seen-set has proven it works across rounds. If another reused proof arrives, the same rejection goes out, and the same minus 2 follows. That is the price the rules set, and it is now a price I can predict.

Check it yourself

Check the leaderboard, 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

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.