VERSIONS
Every rulebook version, scored the same way.
Both versions are scored by plain code on the same stored model replies, so the only difference between the columns is the rule that changed.
RUN 2026-11-11
Farm wallets caught
Wallets on Bitquery's ROBO farm list that Thuli holds or excludes.
99.0%
11,173 of 11,282
99.0%
11,173 of 11,282
Real users held by mistake
Hop's probably-real wallets, a random sample, held or excluded.
2.53%
38 of 1,500
95% interval 1.85% to 3.46%
1.00%
15 of 1,500
95% interval 0.61% to 1.64%
Blocked decisions no longer made
Holds that Reppo's judges blocked. Right means this version no longer makes that hold.
0 of 11
10 of 11
The cost of the change
Wallets Hop removed, a sample on rules A8 and A9, still held. Not a catch rate.
11.2%
56 of 500
3.0%
15 of 500
WHAT CHANGED IN V2
One rule, added after outside feedback.
Reppo's judges read 49 of Thuli's decisions with every address, hash and amount removed. They blocked 11. In each one the reason given for the hold was a funder shared with other wallets.
Rulebook v2 adds: a shared funder alone cannot hold a wallet. How many wallets a funder paid is a count, not a trail. A wallet with a link of its own, or a traced trail to a shared source, is still held.
THE 11 BLOCKED DECISIONS
- contract-holdSCORE 38medium confidenceV1 holdV2 full
- contract-holdSCORE 43high confidenceV1 holdV2 full
- contract-holdSCORE 53high confidenceV1 holdV2 full
- contract-holdSCORE 53high confidenceV1 holdV2 full
- contract-holdSCORE 50high confidenceV1 holdV2 full
- contract-holdSCORE 50medium confidenceV1 holdV2 full
- real-contract-holdSCORE 44medium confidenceV1 holdV2 full
- eliminated-holdSCORE 51medium confidenceV1 holdV2 full
- real-contract-holdSCORE 43medium confidenceV1 holdV2 hold
- real-contract-holdSCORE 49high confidenceV1 holdV2 full
- eliminated-holdSCORE 43medium confidenceV1 holdV2 full
Reppo's routes over all 49 checks: 3 ship, 35 review, 11 block.
CONTROL GROUP
Hop's list, with no model in the loop.
Hop Protocol published the wallets it kept and the wallets it removed from its 2022 airdrop. Thuli's code alone, reading only chain data from before Hop's snapshot:
0.6%
of the wallets Hop kept are held
119 of 20,420, 1 excluded
57.9%
of the wallets Hop removed are held
5,078 of 8,769
Not a catch rate: Hop found its list with a transfer graph, so a graph method agrees with it partly by construction.
ASKED AGAIN
The same wallet, the same answer.
30 wallets from the replay were each screened 5 more times from scratch on 2026-11-11. Every wallet got one outcome across all its runs.
30 of 30
WALLETS WITH ONE OUTCOME EVERY RUN
95% interval 89% to 100%
150 of 150
RUNS THAT MATCHED THE WALLET'S OTHER RUNS
95% interval 98% to 100%
0
RUNS ABOVE THE CODE'S CEILING
The ceiling is set before the model runs.
It also found a flaw in the replay itself. 8 of the 10 sampled wallets that the stored replay holds because of the model's answer got a reward on every fresh run. Their stored hold came from a single answer the checks refused, and a refused answer is usually a one-off. The campaign worker asks again, twice, before it stores a hold of that kind.
Not covered: 30 wallets is a small sample, so the intervals are wide. The same answer twice is not the same as the right answer; the replay and the control sample measure that. This run cost 249 SERV calls.
NOT COVERED
- Each wallet has one stored reply and both columns check that one reply, so neither shows how a model would answer under a different rulebook text.
- Reppo judged decisions made before the shared-funder rule. No verdict here is on a v2 decision.
- Right on a blocked entry means that version no longer makes the decision Reppo blocked, not that Reppo would pass what it does instead.
- Hop's probably-real wallets are not proved real, and the rate comes from a sample.
- A record with no model reply to check again keeps its stored hold in both columns: 3 in ROBO, 9 in the Hop sample, 9 in the eliminated sample.
- One appeal exists and it was written by a test wallet, so it is left out of every number here.
- Dollar figures elsewhere on this site value ROBO at the claim-day price. Nothing was sold or paid.