원클릭으로 Manus에서 모든 스킬 실행

heater-appb-fantasy-math-invariants

스타2

포크0

업데이트2026년 7월 6일 19:01

Use when implementing or reviewing any fantasy-baseball category math in HEATER — stat aggregation, SGP/VORP, win probability, Monte Carlo sims, projections, urgency/scarcity weighting — or when symptoms appear like 1-IP relievers dominating lineups, phantom pitchers ranked high, inverse categories (L/ERA/WHIP) moving the wrong way, or probabilities that don't behave.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

hicklax13

hicklax13/HEATER_v1.0.0

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Fantasy-Baseball Math Invariants (H2H Categories) — the domain physics

12-team Yahoo H2H: hitting R/HR/RBI/SB/AVG/OBP; pitching W/L/SV/K/ERA/WHIP; inverse = {L, ERA, WHIP} (lower wins). Category lists ALWAYS derive from LeagueConfig (src/valuation.py) — hardcoded lists fail structural guards.

Aggregation & weighting (the top bug source)

Rate stats aggregate weighted, never averaged: AVG=Σh/Σab; OBP=Σ(h+bb+hbp)/Σ(ab+bb+hbp+sf); ERA=Σer·9/Σip; WHIP=Σ(bb_allowed+h_allowed)/Σip. Zero-denominator → "—"/None (a walk-only ab=0 game once 500'd the player card until the guards were TDD-pinned).
Volume-weight rate impacts: LP weights ERA/WHIP by IP or 1-IP relievers dominate starters; a per-unit WHIP impact was once scaled 9× (the "WHIP 9→1" live-accuracy fix). AVG/OBP marginals blend by AB/PA — omitting that once priced a .215 hitter at market_price +54.84 to the league's best-AVG teams (Bug C 2026-05-23; guard test_opponent_market_value_scale.py).
L is an inverse COUNTING stat: sign-flips in SGP, but NO IP-weighting in LP. Inverse sets must be {L, ERA, WHIP} — guards reject {ERA, WHIP}.
Weekly scaling: counting ÷ 26 (WEEKS_IN_SEASON=26.0, guard-pinned); rates unchanged. Weeks/dates are ET (src/et_today.py).

Canonical calculators — never reinvent

SGP only via SGPCalculator.totals_sgp / marginal_sgp — consolidates 7 prior local reinventions; marginal_sgp owns inverse signs + rate-volume blending. The scar: an inline value += (stat/denom)*weight scored a default-bad unknown pitcher +38.8 instead of −38.8 (PR #99).
Win probability only via api/services/outcome_probs.seeded_outcome_probs — wraps the Gaussian-copula engine with a CRC32-seeded RNG (same matchup = same number on Team/Matchup/Bubba) and normalizes the UPPERCASE→lowercase key mismatch that otherwise silently degrades the copula to a 0.5 coin-flip. Never derive win% as cats-won/12 — that fabricated the "66.7%" incident.
Scarcity via compute_positional_scarcity_factor — applied SYMMETRICALLY to adds and drops (PR #96).
Urgency: sigmoid k-values read from CONSTANTS_REGISTRY at call time.

Simulation discipline

Paired MC: identical seed for before/after arms — always. True antithetic variates = negate the uniforms (u→1−u, z→−z; ~25% variance reduction, trade_simulator.py). Identical rosters must produce exact-zero deltas.
Convergence is checked, not assumed (test_mc_convergence_wired.py); API sims fix seeds for determinism (playoff n_sims=2000).
Grade authority: Phase-1 weighted SGP owns grade/verdict/confidence; MC overlays rename to mc_* so result.update() can't clobber (guard test_phase1_is_grade_authority.py — pins the live "grade=B with surplus −1.58" regression).

Distributions & uncertainty

Gaussian is fine for rates; low-count cats (SB/SV) are right-skewed → Skellam for weekly H2H counting-cat win-prob (_category_win_prob_skellam, guard test_skellam_win_prob.py).
Per-category posteriors: src/player_model/ heteroscedastic {mean, σ², τ²} with NB/beta-binomial margins + availability survival — calibrated on real game logs (406 players/3837 weeks), the σ/τ seeds fixed a real ERA/WHIP under-confidence.
Under-sampled rosters carry MORE variance: Kalman shrinkage pulls favorable win-probs toward 0.5 (test_kalman_variance_wiring.py).
Intervals: split-conformal (src/engine/projections/conformal.py, E10a) — distribution-free coverage; prefer it over ad-hoc ±bands.
Dominance: before presenting option A over B, the stochastic-dominance screen (src/engine/output/dominance.py, E4) must not show A dominated — a point-estimate win can hide a dominated distribution.

Structure gates that ride along

Punt detection needs BOTH gainable==0 AND rank≥10; punted cats weight 0.05× (not 0 — divide-by-zero safety); trades count toward the 10/week cap; can't-drop restricts DROPS not trades; IL players keep _il_weight_from_status-weighted projections (zeroing them made dropping an IL ace look free — the Crochet/Kirk incident, PR #90).

When NOT to apply

Points/Roto formats have their own standalone engines (src/points_scoring.py, Roto valuation) — don't force H2H machinery on them. Display rounding/formatting is a service-mapper concern.

이 저장소의 다른 Skills

같은 저장소

heater-appb-architecture

hicklax13/HEATER_v1.0.0

Use when starting any task on HEATER's live product — the api/ FastAPI backend (Railway service "celebrated-respect") or web/ Next.js frontend (heater-v1-0-1.vercel.app) — or when unsure where code lives, how a request flows, which database owns what, why CORS/Yahoo-writes/billing seem "missing", or which app (Streamlit vs React) is the real one. Load this first on App B work.

2026-07-062

heater-appb-auth-billing

hicklax13/HEATER_v1.0.0

Use when working on HEATER App B auth or billing — Clerk JWT verification, unexpected 401/402, HEATER_API_WRITE_TOKEN, Stripe checkout/webhooks/subscription tiers, require_pro gating, tier-based AI spend caps — or when auth "works locally but not live" (or vice versa), or a page bounces to /sign-in on hard load.

2026-07-062

heater-appb-bubba-ai

hicklax13/HEATER_v1.0.0

Use when working on Bubba, HEATER's AI assistant — /api/chat/* endpoints, provider streaming (src/ai/providers.py), token metering and spend caps, AI provider keys, attachments/page-context, saved prompts, or the chat composer in web/src/components/bubba/.

2026-07-062

heater-appb-calibration-backtest

hicklax13/HEATER_v1.0.0

Use when proving a HEATER algorithm works — calibrating constants, backtesting projections or recommendations, validating win probabilities, adding uncertainty intervals, checking calibration drift — or before letting any new math drive user-facing recommendations. Also when asked "is this signal real?" or when a calibration run finds no clear winner.

2026-07-062

heater-appb-contracts-openapi

hicklax13/HEATER_v1.0.0

Use when editing api/contracts/*, when api/openapi.json or web/src/lib/api/generated.ts drift (test_openapi_contract or the CI openapi-ts-sync job failing), when JSON responses 500 or render "nan"/"+nan" badges, or when adding/renaming fields on a response the live frontend already consumes.

2026-07-062

heater-appb-debugging

hicklax13/HEATER_v1.0.0

Use when App B misbehaves and the cause isn't obvious — works locally but not on Railway/Vercel (or vice versa), a live page renders empty/zero/wrong data, an endpoint 500s or times out, or tests see different data than the app. Run this playbook before changing code.

2026-07-062

name	heater-appb-fantasy-math-invariants
description	Use when implementing or reviewing any fantasy-baseball category math in HEATER — stat aggregation, SGP/VORP, win probability, Monte Carlo sims, projections, urgency/scarcity weighting — or when symptoms appear like 1-IP relievers dominating lineups, phantom pitchers ranked high, inverse categories (L/ERA/WHIP) moving the wrong way, or probabilities that don't behave.

Fantasy-Baseball Math Invariants (H2H Categories) — the domain physics

Aggregation & weighting (the top bug source)

Rate stats aggregate weighted, never averaged: AVG=Σh/Σab; OBP=Σ(h+bb+hbp)/Σ(ab+bb+hbp+sf); ERA=Σer·9/Σip; WHIP=Σ(bb_allowed+h_allowed)/Σip. Zero-denominator → "—"/None (a walk-only ab=0 game once 500'd the player card until the guards were TDD-pinned).
Volume-weight rate impacts: LP weights ERA/WHIP by IP or 1-IP relievers dominate starters; a per-unit WHIP impact was once scaled 9× (the "WHIP 9→1" live-accuracy fix). AVG/OBP marginals blend by AB/PA — omitting that once priced a .215 hitter at market_price +54.84 to the league's best-AVG teams (Bug C 2026-05-23; guard test_opponent_market_value_scale.py).
L is an inverse COUNTING stat: sign-flips in SGP, but NO IP-weighting in LP. Inverse sets must be {L, ERA, WHIP} — guards reject {ERA, WHIP}.
Weekly scaling: counting ÷ 26 (WEEKS_IN_SEASON=26.0, guard-pinned); rates unchanged. Weeks/dates are ET (src/et_today.py).

Canonical calculators — never reinvent

SGP only via SGPCalculator.totals_sgp / marginal_sgp — consolidates 7 prior local reinventions; marginal_sgp owns inverse signs + rate-volume blending. The scar: an inline value += (stat/denom)*weight scored a default-bad unknown pitcher +38.8 instead of −38.8 (PR #99).
Win probability only via api/services/outcome_probs.seeded_outcome_probs — wraps the Gaussian-copula engine with a CRC32-seeded RNG (same matchup = same number on Team/Matchup/Bubba) and normalizes the UPPERCASE→lowercase key mismatch that otherwise silently degrades the copula to a 0.5 coin-flip. Never derive win% as cats-won/12 — that fabricated the "66.7%" incident.
Scarcity via compute_positional_scarcity_factor — applied SYMMETRICALLY to adds and drops (PR #96).
Urgency: sigmoid k-values read from CONSTANTS_REGISTRY at call time.

Simulation discipline

Paired MC: identical seed for before/after arms — always. True antithetic variates = negate the uniforms (u→1−u, z→−z; ~25% variance reduction, trade_simulator.py). Identical rosters must produce exact-zero deltas.
Convergence is checked, not assumed (test_mc_convergence_wired.py); API sims fix seeds for determinism (playoff n_sims=2000).
Grade authority: Phase-1 weighted SGP owns grade/verdict/confidence; MC overlays rename to mc_* so result.update() can't clobber (guard test_phase1_is_grade_authority.py — pins the live "grade=B with surplus −1.58" regression).

Distributions & uncertainty

Gaussian is fine for rates; low-count cats (SB/SV) are right-skewed → Skellam for weekly H2H counting-cat win-prob (_category_win_prob_skellam, guard test_skellam_win_prob.py).
Per-category posteriors: src/player_model/ heteroscedastic {mean, σ², τ²} with NB/beta-binomial margins + availability survival — calibrated on real game logs (406 players/3837 weeks), the σ/τ seeds fixed a real ERA/WHIP under-confidence.
Under-sampled rosters carry MORE variance: Kalman shrinkage pulls favorable win-probs toward 0.5 (test_kalman_variance_wiring.py).
Intervals: split-conformal (src/engine/projections/conformal.py, E10a) — distribution-free coverage; prefer it over ad-hoc ±bands.
Dominance: before presenting option A over B, the stochastic-dominance screen (src/engine/output/dominance.py, E4) must not show A dominated — a point-estimate win can hide a dominated distribution.

Structure gates that ride along

When NOT to apply

Points/Roto formats have their own standalone engines (src/points_scoring.py, Roto valuation) — don't force H2H machinery on them. Display rounding/formatting is a service-mapper concern.