.agents

Imported from DaSilvaDub/outlier (.agents/AGENTS.md). Install upstream with npx skills add DaSilvaDub/outlier --skill .agents. Copyright stays with the author.

Codebase Quirk: normalizer.implied_probability()

When working with outlier_scrapers, note that normalizer.implied_probability(price) returns a percentage (e.g., 52.5 for -110), NOT a decimal probability (e.g., 0.525). If you are passing this value into a field that expects a standard [0, 1] probability (such as those checked by feedback.py), you must divide the result by 100.

Codebase Quirk: Push Probability Adjustments

When working with probabilities in pack.py, game_totals.py, or similar projection modules, consensus probabilities derived from devigged line odds (e.g., devig_decimal or p_over_headline) are conditional on a definitive win/loss (i.e., assuming no push occurs). To determine the true absolute win probability for a given side, you must multiply the conditional probability by (1.0 - push_prob). Ensure this adjustment is made before populating probability output columns (like model_prob or market_consensus_prob) or feeding them into final sizing/edge calculations, otherwise the sizing logic will dangerously overestimate the edge on push-capable lines.

Codebase Quirk: Outlier API JSON Control Characters

When using json.loads() to parse raw responses from the Outlier API (e.g., in api.py), ALWAYS pass strict=False. The upstream API occasionally returns invalid unescaped control characters in string fields, which will crash the standard json.loads() parser if strict mode is enforced.

Codebase Quirk: NFL Consensus Line Selection vs. Alternate Ladders

When extracting player props from data/NFL/normalized/nfl_props_*.json, the Outlier dataset often contains multiple alternate ladder lines (e.g., +750 to -900 odds) and exchange quotes for the same player and market. Never pick the consensus line solely by sum(len(books)), as alternate ladders can aggregate large book counts across non-standard lines. Always enforce a balanced two-way market check: both OVER and UNDER must exist within normal betting juice (-220 <= odds <= +180), minimizing the deviation from -110. For touchdown markets (ANYTIME_TD), select line == 0.5 and position == 'OVER'.

NFL Game Script Calibration Heuristics (Learned from BUF 41 - DET 31 & Slate Re-Basing)

1. Deficit-Risk Discount on Road Underdog RB Rushing Lines: If a team is a road underdog (+4.5 or greater) facing a high-scoring favorite (Team Total >= 28.0), apply a 15% downward volume haircut to the running back's projected rushing attempts and yardage. When an underdog falls behind by two scores, run volume collapses; pivot exposure to Anytime TD or Receiving Props, which remain active in catch-up mode.
2. Two-High Shell Target Divergence in Comeback Mode: When a favorite establishes a multi-score lead, defenses play deep two-high Cover-2/Cover-4 shells. This neutralizes vertical perimeter deep threats (aDOT >= 14.0, e.g. Jameson Williams), while funneling increased target volume (+20%) to intermediate slot receivers (e.g. Amon-Ra St. Brown) and pass-catching tight ends (e.g. Sam LaPorta).
3. Multi-Window Hit Rate Re-Basing & Multi-Book Liquidity: Never select or recommend a prop based solely on an isolated L5 hit rate (avoiding small-sample noise and alternate-ladder bait). A valid selection requires:
   • Hit Rate Convergence: L5 >= 80% and L10 >= 70%–80%, anchored by stable season snap/usage baselines.
   • Balanced Line Movement & Consensus: Quoted across multi-book consensus (minimum 3–5 regulated books) within standard two-way juice (-145 to +115), confirming sharp market validation rather than synthetic bookmaker ladders.
4. Compiled Situational Factors (Injuries, Matchups, Weather): Every prediction must cross-reference:
   • Vacated Usage: Reallocated target/rush share when key personnel are inactive or on IR.
   • Opponent Defensive Efficiency: Matchups against specific coverage shells and defensive front weaknesses (e.g., zone run defense vs. man perimeter).
   • Environmental Calibration: Outdoor wind (>12–15 mph) or precipitation haircuts on vertical passing vs. dome/controlled venue pace upgrades.

Windows PowerShell Python Execution & UTF-8 Console Encoding

1. Avoid Complex Quotes in Terminal One-Liners: In Windows PowerShell, running inline Python scripts (python -c "...") with nested quotes or $() frequently fails with parser errors. Write scratch scripts to <appDataDir>\brain\<conversation-id>/scratch/ instead.
2. Stdout Encoding Constraint: Python on Windows defaults stdout to cp1252, which raises UnicodeEncodeError when printing Unicode characters like ↳ (\u21b3) or em-dashes —. Scripts that print formatted output must call sys.stdout.reconfigure(encoding="utf-8", errors="replace") and favor standard ASCII text.