CAPABILITIES
& CONFESSIONS
Everything the robot can do, everything it can't, and the receipts for both. This page is generated from the same backtest data as the picks — it cannot flatter itself.
← BACK TO THIS WEEK'S POLL
🤖 WHAT I AM
A logistic regression over 8 named signals, trained walk-forward on FBS-vs-FBS games since 2015 (CollegeFootballData). Picks lock and are never edited. Every probability decomposes into signal contributions that sum to a points-style lean — the receipt IS the model, not a summary of it.
home_field rating_gap rest_diff opp_pass_def_trend rush_def_trend pass_off_trend lookahead_spot travel_km
📜 THE RECORD — EVERY SEASON, PREDICTED BLIND
Walk-forward: the model only ever saw seasons before the one it was picking.
SEASONGAMESSTRAIGHT-UPVS THE LINEATS %BRIERML ROI
2016 76070.3%358–382–2048.4%.190 no ML data
2017 77670.4%390–365–2151.7%.185 no ML data
2018 77272.4%391–361–2052.0%.175 no ML data
2019 77473.1%377–384–1349.5%.176 no ML data
2020 (COVID) 53471.2%274–251–952.2%.187 no ML data
2021 77069.5%383–378–950.3%.187 -4.0%
2022 77668.7%389–374–1351.0%.199 -4.6%
2023 79272.3%395–381–1650.9%.185 -1.2%
2024 79869.2%391–388–1950.2%.193 -5.4%
2025 80872.5%396–394–1850.1%.184 +1.0%
🎯 CALIBRATION — THE THING I'M ACTUALLY GOOD AT
Career held-out calibration, 2021+. When the robot says a number, this is how often the number was right.
40–50%
47.1% n=444 50–60%
52.6% n=521 59–69%
65.0% n=549 70–80%
72.4% n=557 80–90%
85.6% n=536 90–100%
93.0% n=359
Every reliable bucket within ±7 points of its midpoint. Career Brier .190 vs .191 for the Elo-only baseline.
💸 THE MONEYLINE LEDGER — STRAIGHT PICKS VS THE BOOK
Flat $1 on every qualifying pick at real posted moneylines, 2021-2025 (3,771 games). Historical books: Bovada/DraftKings/ESPN Bet — FanDuel pricing sits within a tick.
STRATEGYBETSWINSFLAT ROI
bet every straight-up pick3,7712,619 -2.81%
any modeled edge1,537895 -0.62%
edge > 2 points1,278709 -0.35%
edge > 5 points968514 +2.91%
robot backs the favorite3,3012,415 -4.83%
robot backs an underdog470204 +11.38%
⚠ The underdog number is the interesting one and the dangerous one: 470 bets is where luck still writes headlines, and six strategies were tested to find it. It runs as a paper-ledger experiment until live results confirm or kill it.
⛏ THE FULL AUDITION RECORD
Every candidate signal, its verdict, and why. Graduation requires: improves held-out Brier, stable coefficient sign, holds post-2021.
SIGNALVERDICTEVIDENCE
talent_gap AUDITIONING passes 2/3 criteria — needs more seasons of evidence
returning_production AUDITIONING passes 2/3 criteria — needs more seasons of evidence
rush_def_trend GRADUATED improves held-out Brier (mean +0.00003), stable sign, holds 2021+
pass_off_trend GRADUATED improves held-out Brier (mean +0.00003), stable sign, holds 2021+
rush_off_trend CUT no lift on held-out seasons (mean Brier delta -0.00003)
turnover_luck CUT no lift on held-out seasons (mean Brier delta -0.00002)
letdown_spot CUT no lift on held-out seasons (mean Brier delta -0.00014)
lookahead_spot GRADUATED improves held-out Brier (mean +0.00020), stable sign, holds 2021+
travel_km GRADUATED improves held-out Brier (mean +0.00020), stable sign, holds 2021+
body_clock CUT no lift on held-out seasons (mean Brier delta -0.00014)
line_movement ADVISORY evaluated but cannot graduate on backtest evidence — open→close movement is not observable at Tuesday lock time
🩻 KNOWN WEAKNESSES — THE CONFESSIONS

I don't know the quarterback is hurt. No injury reports, no practice news, no depth charts. My ratings only learn about a missing starter after he misses games. The betting market knows Tuesday; I find out Saturday. Mid-week line movement against a locked pick is usually this weakness wearing a trench coat — it gets flagged as 'steamed' in the ledger, and those flags measure exactly how much my news-blindness costs.

I don't beat the spread. Career 49.5% against the closing line; break-even is 52.4%. The closing number is sharper than I am, which is why the bet card is paper and the real-money gate stays shut.

My cover probabilities are on probation. Win probabilities are proven calibrated (see above). The COVER probabilities on the bet card are derived through a margin model that has not earned the same trust — when the card says 76% to cover, treat it as the model's opinion, not a measurement. Validating them is what the paper ledger is for.

Elo does most of my work. The other seven signals together improve career Brier by less than half a point of Brier×1000. They're real (they survived the audition), but the receipts should be read knowing one clue carries the band.

Small samples flatter me. The +11% underdog moneyline ROI comes from 470 bets across five seasons, found while testing six strategies. That's exactly the shape luck takes. It stays an experiment until live paper results agree.

My historical odds are a proxy. Backtests price against Bovada/DraftKings/consensus (CFBD), not FanDuel directly — close, not identical. Live pricing uses actual FanDuel via The Odds API.

📏 HOUSE RULES

• Picks lock Tuesday and are never edited. The scorecard grades the locked JSON, not a memory of it.

• The model never sees in-week news. If the line moves ≥3 points against a locked bet, the ledger flags it 'steamed' — graded anyway, reported separately, so news-blindness gets measured instead of hidden.

• Better numbers are taken, worse numbers are never chased: if the available line beats the locked number on a side we already bet, the ledger may record a line-shopping top-up. A line that moves against us is information, not a discount.

• Real-money mode requires positive held-out backtest ROI AND positive live paper CLV over at least half a season, then a human typing the unlock phrase. The robot cannot flip its own switch.

• 2020 is reported, not counted. You know why.

ROBOT POLL · this page regenerates from the backtest with every deploy · it cannot flatter itself