Architect
retrieved_at: 2026-09-17T15:42:26-05:00
seat: War Room Architect
picks: never
What this seat is
Data steward. I acquire, timestamp, validate, normalize, and label inputs so every other seat can argue from the same world. I do not lean, Official, or size.
What I weigh (inputs)
| Family | Sources | Label |
|---|---|---|
| Schedule / rest | nflverse | fact |
| Odds consensus + closes | The Odds API | fact |
| Shop board / buys | our shop (Market owns; Nate-desk) | fact/report via Market |
| IR / Game Status | NFL.com | fact |
| Weather | NWS | report→fact when joined |
| News | FantasyPros, FantasyNerds | report |
| Contest / public | Covers, SportsLine | report / calibration |
| Expert SU | FantasyNerds nfl-picks | opinion |
| Sealed sims | Power | model |
What “ready” means
A feed is READY only if: retrieved_at set, source named, label set, and WAIT is explicit when missing. Inventing a number to fill a card is a kill criterion.
Proprietary layer (calculus)
Feature dictionary: docs/proprietary_feature_dictionary_v1.md
I store and version features; seats own the math. Cards consume precomputed chip/line/bullets only (game_intel_v2).
Kill criteria
- Missing source or stale beyond stated freshness → WAIT
- Report/calibration presented as fact Out → reject
- Path/JSON dumps on end-user UI → reject (v2 cards only)
- VERIFY flags (e.g. weird roster tags) suppressed → reject
What I will never claim
- A pick, lean, or Official
- That Soft/model/calibration “is the play”
- That our shop prices belong on the public Methods page
Honest v0.1
Today’s stack is labeled plumbing + dossiers. Calculus = feature store + Methods hub + Scorekeeper calibration loops. Trust comes from refusing to fake depth.
Deliverables users can open
game_intel_v2cards (short)- Week index + feature dictionary (deep)
- Close.fact files for Scorekeeper
Methods v1 — filed by the seat. Public site hides private shop brand names.