How a pick actually gets published.
This page walks through the full pipeline — from source scraping to confidence scoring to public grading. If you can read JSON, you can replicate it.
Not a chatbot. Reynard.
Reynard runs a panel of four specialist analysts and a judge. Every pick has to survive their debate. He's been training on the public market since 2017 — every line move, every sharp account, every pro capper.
Reynard's Line Tracker
Trained on historical line moves. Detects steam, reverse-line moves, and where sharp money is hitting before the public catches it.
Reynard's Mirror
Trained on the pick history of verified sharp accounts. Identifies plays that fit their known edges — angle, sport, market.
Reynard's Crowd-Reader
Trained on public betting % and handle splits. Flags games where the public is overcorrecting — the fade opportunity.
Reynard's Judge
Real-time closing-line value model. Vetoes any pick that doesn't project positive CLV at current market price.
The system gets sharper with every pick.
Every graded result feeds back into source reputation, confidence calibration, and agent weighting. As we wire in more odds APIs, historical line data, and sharp tracking, Reynard's panel gets more evidence to reason over. We're early — his models already ship, and they only get better from here.
Source → debate → score → publish.
Source layer
Live feeds — sharp account signals, pro handicapper picks, private betting communities, line-movement streams, public handle splits, and steam-move detectors. Sources rotate in/out based on rolling CLV performance. No source is trusted on faith.
Reynard's panel debates
Every candidate pick passes through Reynard's four specialists. Each scores independently; they then argue the thesis in a structured exchange. His judge weighs CLV, variance, and historical base rate. Consensus is required before a play advances.
Publication layer
Only picks above our confidence threshold ship. The row is written immutably — sport, bet, odds, stake, timestamp, source ID — before the game starts. Grading is automatic and public. We can't rewrite history; the DB doesn't let us.
Here's the exact shape of a published pick.
Every field is visible on /record. Nothing is computed server-side post-hoc; the grade engine only reads settled game results.
{
"id": "a3f91c",
"published": "2026-04-17T00:47:12Z", // ← immutable, before tip-off
"sport": "NBA",
"league": "NBA",
"type": "Spread",
"pick": "Lakers -3.5",
"odds": "-110",
"stake": 2.00, // ← units
"confidence": 8, // ← 1–10 score
"sources": ["#04", "#11"],
"clv": +0.14, // ← closing line value
"graded": "2026-04-17T03:42:09Z",
"result": "WIN",
"units": +1.82
}What we don't do.
Delete losing picks
Ever. Every published row stays in the ledger with its ID. Audit is trivial.
Edit stakes post-hoc
Stake size is written at publication and never changed. Units are computed from that stake.
Publish after tip-off
The published timestamp is compared to the game start. Late rows are flagged and excluded.
Private "VIP" picks
Every subscriber sees every pick. No tiered info. One record.
Parlay inflation
We don't ship parlays to inflate a win rate. Singles and straight props only.
Promoted sportsbooks
We don't take affiliate money. Your bookie is your business.