Accuracy, transcripts, allowance

Status

Forecast performance

Winner accuracy

How often the issued forecast picked the winning team.

Full scorecard

Live · 2026

70.6%

▼ 29.4 pts 7d

17 games · as issued

Held-out · 2024–25

56.1%

· 0.0 since Sep 22

544 games · backtest 2024–25 (fit ≤2023)

Live − backtest +14.5 ptsDifferent seasons and sample sizes; this difference is not a model improvement.

Served v1.0 (fit ≤2025), as issued before kickoff · from wk 2

Build movement

Last 15 minutes

Read Sep 26 · 1:51 AM MT

Board 215 done · 26 open · plan 127/127 · chain 8/8 · gate CLOSED

Landed (1)

H-193 q-language: line revisions 15% too large, FantasyPros underreacts, the shipped blend is calibrated; no b

Compared with Sep 26 · 1:35 AM MT.

Measured vs live

Forecast quality

Winner ↑ · error ↓

measuredactualgap

Winner correctn 17

56.1%· 0.0 since Sep 22

70.6%▼ 29.4 7d

+14.5

Marginn 17

96.7%· 0.0 since Sep 22

97.0%▲ 23.9 7d

+0.2

Totaln 17

23.3%· 0.0 since Sep 22

25.3%▼ 1.7 7d

+2.0

Team pointsn 34

33.6%· 0.0 since Sep 22

43.0%▲ 15.9 7d

+9.3

Pass yardsn 34

24.9%· 0.0 since Sep 22

18.3%▲ 2.3 7d

−6.6

Rush yardsn 34

32.9%· 0.0 since Sep 22

35.6%▼ 9.7 7d

+2.7

Fantasy pointsn 383

55.8%· 0.0 since Sep 22

65.9%▲ 7.5 7d

+10.1

Winner = % correct. Other rows: %error = mean |pred − actual| ÷ mean |actual|. Held-out: 2024–25 (fit ≤2023). Live: 2026 as issued. Read Sep 26 · 1:51 AM MT.

With vs without

Feature contribution

Green helps

B hypB liveQ hypQ live

Winner correct

0.0

+3.0

0.0

+3.0

Margin

+0.2

+0.1

Total

−0.1

+0.2

−0.1

+0.1

Team points

0.0

+0.1

0.0

+0.1

Pass yards

−0.1

0.0

−0.1

0.0

Rush yards

0.0

Fantasy points

0.0

−1.5

+0.1

+0.3

B = coaching decisions; Q = injury-report text. Test = backtest 2024–25 (fit ≤2023); live = 2026 wk 1–3 refits. News claims are withheld.

Podcast transcripts

T183,768 episodes

downloaded16,930 · 20%

transcribed978 · 1.2%

+670 24h · pace 723/day · done Jan 18, 2027

T2164,222 episodes

downloaded3,280 · 2%

transcribed657 · 0.4%

+53 24h · pace 723/day · done Sep 1, 2027

T3227,173 episodes

downloaded3,343 · 1%

transcribed1,999 · 0.9%

+0 24h · pace 723/day · done Jul 9, 2028

Pace = all tiers, last 24 h. Worked T1, then T2, then T3. Counts Sat Sep 26 1:48 AM MDT.

Research

Findings and adoption

69findings

6adopted

0live in served

18in review

See findings

In served v1.0hypactual

CP-H49 · trailing windows include last game

−0.352 FP MAE

no live A/B

CP-H65 · stop pricing all weeks on wk-1 market

−0.063 FP MAE

no live A/B

CP-H199 · kalman share leak removed

+0.040 FP MAE

no live A/B

CP-H6 · order-independent room sums

−0.006 FP MAE

no live A/B

CP-H43 · pass attempts weighted to recent

−0.071 att MAE

no live A/B

CP-H47 · DEF events from implied total

−0.178 DEF MAE

no live A/B

Adopted findings · live shadow vs servedhypactual

B13 · draft capital on thin history

+0.048 pts RMSE vs BASE (thin)

gain −0.046 MAE, 136 rows, 3 wk [−0.296,+0.011]; backtest −0.011 · not promoted

B41 · draft + contract on thin history

+0.059 pts RMSE vs BASE (thin)

gain −0.031 MAE, 136 rows, 3 wk [−0.210,+0.016]; backtest −0.030 · not promoted

Q-13 · Saturday attention → availability

−0.0030 Brier (P(play))

live shadow scores from week 4; backtest +0.0029 · not promoted

Q-14 · fame adds nothing (a null)

null: −0.005 pts, floor not cleared

nothing to serve

Q-19 · expert file leaks availability (control)

control passed: −0.020 Brier = contaminated

not servable: leak

XD-S07 · empirical null for the batteries

method: null sd 2.22 vs 1.00

not a model feature

Σ hyp −0.38 FP MAE (4 rows) · Σ actual no live A/B

Hypotheses: docs/hypotheses.json · review incl. 1 L-07 · fixes: CHANGES-PROPOSED.md

Capacity

Weekly model allowance

Account-wide · projected from the last 24h

Claude57% remainingProjected out Sep 27 · ~8 PM MTRead Sep 26 · 1:46 AM MTHow estimated

32.0 points used over 23.8h of the last 24h.

Weekly allowance resets Oct 1 · 9:00 PM MT.

Logged scheduled starts, last 24h: 0

Account-wide allowance; includes work outside Playerweek.

Codex58% remainingProjected out Sep 30 · ~1 PM MTRead Sep 26 · 1:51 AM MTHow estimated

13.0 points used over 24.0h of the last 24h.

Weekly allowance resets Oct 1 · 6:26 PM MT.

Logged scheduled starts, last 24h: 0

Account-wide Codex allowance. Direct account-status read; no model request.

Antigravity · Gemini models85.5% remainingExpected to last until Oct 1 resetRead Sep 26 · 1:51 AM MTHow estimated

12.0 points used over 24.0h of the last 24h.

Weekly allowance resets Oct 1 · 7:29 PM MT.

Logged scheduled starts, last 24h: 180

Shared by Gemini Flash and Pro in Antigravity. Local scheduled starts also have a 15-minute cooldown. Gemini app/web usage is unknown here.

Antigravity · Claude and GPT models96.9% remainingExpected to last until Oct 1 resetRead Sep 26 · 1:51 AM MTHow estimated

0.0 points used over 24.0h of the last 24h.

Weekly allowance resets Oct 1 · 11:19 PM MT.

Shared by Claude Opus, Claude Sonnet and GPT-OSS in Antigravity; separate from Claude Code and Codex.