Goodguy AI Labs · Open weekly experiment
The Fantasy Fruit Fly
We’re testing whether a decision circuit built from a fruit-fly connectome can learn fantasy football. We give it public NFL facts. It decides which patterns matter and builds the football stat lines behind the rankings.
The software keeps the records straight: stable identities, dates, roles, and source receipts. It does not supply a football answer. The Fly’s learned routes decide what matters and how each stat line takes shape; a holdout controls how much that route can influence the final board. We do not import analyst rankings, another model’s picks, or a fantasy answer key.
Follow the Fly through the season: weekly calls, game-by-game grades, and what it learns next.
- Board
- Checking release…
- Scoring
- PPR · Half PPR · D/ST
- Answer key
- None
Rankings stay blank until a complete release passes the page’s checks.
The Fly’s player rankings
Week 1 rankings by position.
The Fly makes the football call. It weighs the public NFL record and chooses the range of attempts, yards, touchdowns, carries, targets, receptions, and defensive events. Only then does the scoring bank turn that Fly-chosen stat line into PPR, Half PPR, or D/ST points. Open a row to see the range and role reference.
Loading the released board…
| Rank | Player | Matchup | Projected points | Likely range | Actual | Points off | Actual rank | Role reference |
|---|---|---|---|---|---|---|---|---|
| The board appears after JavaScript verifies the release contract. The Fly gets the public record; it does not get somebody else’s answer. | ||||||||
The released board scrolls inside this frame. Search or filter first, then scroll to inspect more players.
No released players match those filters.
Week results
How close were the Fly’s calls? These grades use the saved rankings, the NFL results, and the scoring format you’ve selected.
- Status
- Results appear after the weekly record is verified.
One weekly receipt
Wednesday report card
Decision groups · points and range
How did the Fly do on the players it put near the top? We check its top 12 QBs, 24 RBs, 36 WRs, 12 TEs, and all 32 defenses separately from the full pool. Raw NFL outcomes are the Fly’s lessons; analyst rankings are comparisons only.
| Position | Graded | Average points off | Scores in range |
|---|
Same-player comparison
FLEX · Half PPR
Only players ranked by all three are compared.
Original public board, corrected replay, and FantasyPros ECR are checked against the same players and actual results. Lower rank error is better; higher percentages are better.
| Ranker | Point error | Rank error | Pairwise order | Top-range hits |
|---|
Point error is average fantasy points off; ECR publishes ranks only. Rank error is average places off. Pairwise order is the share of player pairs ordered correctly. Top-range hits shows precision / recall: how many selected players finished inside the scoring cutoff / how many top finishers were selected.
Download same-player scorecard CSV ↓Coverage and starter range
| Ranker | Imported rankings | Fly entries compared | Fly: places off | Ranker: places off |
|---|---|---|---|---|
| External comparisons appear after the weekly data is verified. | ||||
Position-by-position comparisons
| Ranker | Position | Fly entries compared | Fly: places off | Ranker: places off |
|---|
No rank in archived feeds
Captured expert feeds can rank FLEX pools, while dedicated position feeds have their own scope. A player without a rank in the archived feed was not assigned a last-place rank.
The experts’ complete ranking lists
We also check each expert’s entire imported list against NFL results, including players the Fly never ranked. “Graded / imported” shows the coverage; an unresolved NFL identity is not silently scored as zero. This is a separate report: the Fly has no saved forecast for those extra players, so they cannot enter a head-to-head score. A longer list also changes the size of the ranking pool.
| Ranker | Position | Graded / imported | Ranker: places off |
|---|
How do we grade the Fly?
Think of a teacher checking a worksheet. The Fly has already written its answers. The games give us something to check them against.
Keep the original calls.
We grade the exact board people saw, not a new set of predictions. Each call stays attached to its player, team, game, and week.
Check what happened on the field.
We match the recorded NFL stats to stable player IDs, then apply the same scoring rules used for the forecast. PPR awards one point per reception; Half PPR awards half a point. D/ST uses the defensive rules listed in the lab notes.
Count the misses.
The points-off column is projected minus actual. A call of 20 points against a score of 15 is +5: five points high. A call of 10 against 15 is −5: five points low. Those two calls have an average miss of five points, even though their high/low lean balances to zero.
Compare the whole board.
Every Fly entry is checked against the full imported expert rankings. The expert table measures rank order, not projected points. For entries with both forecasts, the Fly, each ranker, and the actual results are reordered within the same matched player pool and position. The whole-board average pools those misses across all positions; you can also inspect each position. PPR is compared with PPR; Half PPR with Half PPR. Ties share an average rank. Fourth versus tenth is a six-place miss. An absent rank stays absent, not a made-up last-place pick.
Give the Fly its next lesson.
The Fly learns from raw NFL stats: workload, yards, touchdowns, turnovers, and defensive events. Expert rankings are only a comparison on this page. They do not become the Fly’s homework answers. One weekly receipt records the graded outcomes and learning state for that week.
- Average points off
- The size of the miss, averaged across the graded entries. Researchers call this mean absolute error, or MAE. Mean means average; absolute means a miss counts in either direction. Lower is better for point accuracy.
- High / low lean
- The average signed miss, also called bias. A positive value means the Fly ran high overall; a negative value means it ran low. A value near zero can still hide large misses in opposite directions.
- Scores in range
- How often the actual score landed inside that entry’s likely range. The range runs from the 10th to the 90th percentile: the middle 80% of simulated scores. Coverage of 65% means about 65 of every 100 graded scores landed in their own predicted ranges.
- Places off
- The average distance between a predicted rank and the actual rank, measured in ranking places. It tells us whether the order was useful. A good order and a close point projection are two different tests.
- Professional analysts & ECR
- FantasyPros ECR is its Expert Consensus Rankings. Individual rankers appear as numbered professional analysts; the numbers are labels, not a leaderboard. Each row uses its own shared player pool, so compare the Fly and ranker within that row.
- Provisional results
- The games are graded, but official stat corrections can still change the result. A correction gets a new dated results record; it does not rewrite the Fly’s saved predictions or give the same week a second cumulative reward.
Not numerically ranked
Unrated players
Players without enough applicable history are kept out of the numerical board instead of receiving a made-up estimate.
View unrated players
Human vs. model
Are you smarter than a fruit fly?
Choose one player from each same-position pair before the first listed kickoff. Pairs come only from the current release. This browser-only practice round has no verified leaderboard and is not personal roster advice.
Waiting for a current board before setting a lock time.
Five pairs will appear only when the current released rankings contain enough same-position comparisons.
Make a call first. Final outcome records are needed before this round can be scored.
Honor system: we do not send or verify your picks. If browser storage is unavailable, the page will say so rather than pretending your calls were saved.
Past practice rounds saved in this browser
ELI5: the method
Put the Fly at the controls.
Here is the marble-machine version. Software puts clean NFL facts in the hopper and keeps every player and team attached to the right record. The Fly decides which signals matter and what football stat line to project. The scoring bank only adds up the fantasy points. That division is the point of the experiment.
- 01
We load the obvious stuff.
Schedules, stable IDs, prior workload, team and opponent context, scoring rules, and a cutoff-safe depth-chart role reference enter the chute. It is the public NFL record—the stuff any fantasy player can find with a little digging. We do not hand the Fly a ranking.
- 02
The Fly gets to work.
The factual companion and measured BANC path learn from the same untouched history. Bzzz. This is where the Fly begins to learn and make decisions; neither route receives somebody else’s fantasy answer.
- 03
The Fly earns its say.
Unseen weeks decide how much the Fly-derived route can influence each part of the projection. No one dials up a favorite.
- 04
The Fly builds the stat line.
First it calls the football: volume, yards, touchdowns, turnovers, sacks, and points allowed. The scorekeeper translates that stat line into PPR, Half PPR, or D/ST points.
- 05
A pretty normal player pool.
One QB, two RBs, three WRs, and two TEs per team, plus every D/ST. The depth chart keeps the right people in the room; the Fly decides what to do with them.
- 06
How much of the call is Fly?
The Fly-derived network works alongside a companion model trained on the same NFL history. Performance on unseen games determines their mix. That gives us something to track each week: where the Fly helps, and how much its contribution grows.
Data notes
Bzzz. Inside the Fantasy Frankenfly.
A quick lab note on the records, wiring, and rules behind this week’s calls. The page reads one release file and shows its work instead of filling the board with stand-in rankings.
Release information
- Status
- Checking for a release…
Primary sources & attribution
- Sources appear after the current release passes its contract.
This release’s files
- Sources will appear with a released dataset.
Limitations
- Limitations will appear with a released dataset.
Evaluation context
How this release was checked
These are release-supplied retrospective checks, not a promise about an individual player or week.
Fair read: compare the primary model with every declared control and retain negative results.