Methodology

How a Take becomes evidence.

Inventory, transcript validation, extraction, review, and outcome scoring are separate layers. Future results enter only after the Take ledger is fixed.

Evidence boundary

  1. Corpus. 947 episodes from 2021–2026 YTD; 869 have validated transcript pairs.
  2. Extraction. Candidate claims were extracted without future outcome data.
  3. Approval. WP6 approved 67 records in two supported families. Speaker attribution is unknown.
  4. Scoring. Versioned rules attach nflverse outcomes for completed 2021–2025 seasons. Scores bind exact Take, rule, candidate, transcript, and outcome hashes. No 2026 YTD accuracy result is published.
  5. Audit. Agent preflight is complete. Human approval of the 25-record preseason sample is still required.

Spotify transcript coverage is era-dependent. Local Whisper fallback filled approved gaps, and occasional ASR name errors were corrected through the reviewed entity-override layer without changing raw sources. No transcript supplies speaker labels.

preseason-conviction-v2

Sample n=55; eligible n=49; excluded n=6. Categories: aligned 21, availability limited 6, contradicted 13, mixed 15.

Baseline: market adp (unavailable). Outcome dataset: nflverse-active-games-2021-2025-v2; formula: ppr-v1.

Uncertainty: wilson_95, estimate 43%, interval 30–57%.

buy-sell-v2

Sample n=12; eligible n=5; excluded n=7. Categories: improved 2, inconclusive 3, unscored 7.

Baseline: self prior four active games (available_for_complete_windows). Outcome dataset: nflverse-active-games-2021-2025-v2; formula: ppr-v1.

No trade accuracy estimate or interval is published; the 12-record family remains descriptive.

What is not claimed

  • No blended overall accuracy score.
  • No market-outperformance claim without a verified historical baseline.
  • No trade accuracy claim from a 12-record family.
  • No host-level attribution while speakers remain unknown.
  • No inference from missing transcripts.

Outcome sources are nflverse schedule, weekly roster, and player-stat release files, each bound by URL and SHA-256 in release r1-2021-2025-v2.

Excluded families and corrections

Weekly lineup, waiver, and directional prediction failed their WP5 family gates. They receive no release records or outcome charts. A correction creates a new immutable scoring and presentation release; it does not rewrite approved v2 records.

Release
r2-2021-2026ytd-deploy-v1
Generator
presentation-v3
Commit
3c5d9a3
Build time
2026-08-31T09:36:36+10:00
Review
local_review_candidate_pending_human_wp7_audit

Download the compact release JSON and input hashes