WIGGLY GATORRACE A FLY / THE DATAWIGGLY GATORRACE A FLY / THE DATAEvery chart on this page is generated straight from actual log files and data - nothing here is illustrative or invented. Scroll down for raw downloads if you want to check our work.
6 seeds (Johnny, Frank, Pickles, Meatball, Ziggy, Beans) each ran 2,500 real, logged attempts on Easy, Medium, and Hard - 45,000 real attempts total, 15,000 per difficulty. Every seed on every tier started from the exact same shared checkpoint, so the only thing that differs between difficulties is the game's own physics, not anything about the search process. Full methodology in the project journal →
Within a single fly's run, the brain genuinely learns - each of the 2,500 attempts uses the same evolving set of synaptic weights, not a fresh one. An eligibility trace tracks which synapses just fired together; the attempt's outcome then scales how much that trace becomes an actual, permanent weight change. That change is verified with a clean, no-exploration probe right after - if it helped, it's kept as the new best; if not, the weights are reverted right back to before that attempt. Good changes stick. Bad ones don't.
That's a real line from a real log: the fly tried something, it didn't pan out, and the change was undone - learning from a mistake by discarding it, not repeating it blind. Exploration itself also decays over the run (epsilon 0.12 → 0.02), so early attempts experiment more and later attempts increasingly exploit what's already been learned - more direct evidence of accumulated learning, not independent fresh tries each time. What resets each attempt is the course itself (position, obstacles) - not the brain.
Every clean (no-exploration) score across all 6 seeds, combined per difficulty. Log-scaled Y axis - the tail is real, just rare.
Comparing the best-ever score per tier isn't the same as comparing what a typical attempt looks like - and mixing them up is misleading. Here's both, done honestly.
| Comparison | Mean A | Mean B | z-score | Verdict |
|---|
Best real score per seed, per difficulty (2,500 attempts each). Ranked by average placement across all three tiers.
| # | Fly | Easy | Medium | Hard | Avg. rank |
|---|
After the 45,000-attempt benchmark above finished, we took just the winning fly on each tier and asked: could dedicated extra training push it even further? This is a separate, follow-up experiment on top of the benchmark - not part of the leaderboard, and not run with the same budget as every fly above.
| Does the "wider brain" trick generalize? | Start | Result | Change |
|---|---|---|---|
| Hard (Pickles) | 9,620 | 14,420 | +50% |
| Medium (Pickles) | 6,880 | 6,880 | +0% |
| Easy (Meatball) | 13,930 | 13,930 | +0% |
Real counts from the MaleCNS reconstruction - the same 2,476-neuron, 11,364-synapse subgraph used everywhere on this site.
The individually-recorded showcase runs (not the bulk benchmark attempts above) - no card design, just the raw table.
| Run ID | Fly | Map | Difficulty | Score | Survival | Result | Real? |
|---|
Loading…
totalRealAttempts (45,000) and distribution.n (also 45,000) - that's every real attempt, boiled down to the mean/median/percentiles you see above. Or open any *-benchmark-stats.json and check attemptsLogged per fly (2,500) - six flies × three difficulty files = the same 45,000. Or just click Load Data & Verify above and let your own browser do that arithmetic against the live files.© 2026 Wiggly Gator. Website by BrightRidge Creative.