FINCHAL's prize money is beside the point; its random-trader baseline is not

Run FINCHAL's scoring formula against its own headline test and the two numbers do not sit where you would expect. Beating the ceiling puts an entrant at p = 0.95, and the score at that point is -log10(0.05), or about 1.30 ^[18]. The published landmarks on that scale are 2.0 for a one-in-100 outcome and 3.0 for one-in-1,000 ^[11]. Clearing luck therefore lands near the bottom of the range VIDRAFT chose to display, and by construction 1,000 of the 20,000 random players clear it in every market ^[19]. Reaching 3.0 means finishing ahead of 19,980 of them ^[20].

The asset-specific null is what makes those thresholds mean anything. The pre-season ceilings run from 86.6% on Bitcoin to 9.2% on gold ^[8], a factor of 9.4 between the two ^[21]. An 80% year in Bitcoin sits inside the range random position-flipping could have produced; 12% in gold sits outside it ^[9]. A single cross-asset winner was tried and dropped after six normalisation methods left the tails still favouring particular markets, so the money splits into four separate $500 awards ^[3].

Nor is the distribution fixed. FINAL-Bench says it rebuilds the random-player returns daily on the market path that has actually occurred, specifically to stop a broad rally lifting every long-biased entrant ^[10]. Beta gets re-priced as the season runs: an agent that was long through a rally is measured against 20,000 random traders who lived through the same rally.

The 1x exposure cap does the equivalent job on the other axis. VIDRAFT said an uncapped simulation returned 48,763% cumulative, which is what a leaderboard looks like when it has quietly become a contest over position size ^[6].

The eight self-tests on the scorer are where the lineage shows ^[13]. The interface withholds future prices from agents ^[12], but the harder leak is on the scoring side, and the test VIDRAFT calls critical checks that an entrant who opens a position on the same bar as a price jump earns nothing from that jump ^[13]. One shifted index would let the scorer read the future while still printing plausible numbers ^[14]. That is the failure FINAL-Bench documented on August 22, when the same drug-prediction task came out 0.211 AUROC apart depending on whether the split was time-based or random ^[15]. Leakage does not announce itself. It just improves the results.

Which is why the pot is the least interesting line in the announcement. Split four ways across a 122-day season, $500 per asset works out at roughly $4.10 per asset-day ^[22]; nobody staffs an agent for that, and runtimewire's own read is that the design work is the draw ^[17]. What is missing is ownership. The announcement runs under a SeaWolf-AI account displayed as VIDRAFT_LAB alongside the FINAL-Bench organization, whose page links to VIDRAFT, and the operational relationship among those four names is not defined in the public materials ^[16]. For a contest whose entire value rests on the credibility of one scorer, the party answerable for that scorer should be nameable.