Run any Skill in Manus with one click

heater-appb-fantasy-math-invariants

Stars2

Forks0

UpdatedJuly 6, 2026 at 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.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

hicklax13

hicklax13/HEATER_v1.0.0

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

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

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.

More from this repository

same repository

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.