paper-dynasty-card-creation/CLAUDE.md

# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

This is a baseball card creation system for Paper Dynasty, a sports card simulation game. The system pulls real baseball statistics from FanGraphs and Baseball Reference, processes them through calculation algorithms, and generates statistical cards for players. All generated data is POSTed directly to the Paper Dynasty API, and cards are dynamically generated when accessed via card URLs (cached by nginx gateway).

## ⚠️ Critical Lessons Learned

**MUST READ**: `docs/LESSONS_LEARNED_ASTERISK_REGRESSION.md` before working with player names or card generation.

**Key Points**:
- API parameter for player name is `name`, NOT `p_name`
- Card generation is cached - always use timestamp for cache-busting: `?d={year}-{month}-{day}-{timestamp}`
- S3 keys must NOT include query parameters
- Player names are ONLY in `players` table (not in `battingcards`/`pitchingcards`)
- NEVER append visual indicators (asterisks, hashes, etc.) to stored player names

## Key Architecture Components

### Core Modules
- **batters/**: Batting card creation with rating calculations (calcs_batter.py) and card generation (creation.py)
- **pitchers/**: Pitching card creation with ERA/WHIP calculations (calcs_pitcher.py) and card generation (creation.py) 
- **defenders/**: Defensive rating calculations and fielding card generation (calcs_defense.py, creation.py)
- **db_calls.py**: Paper Dynasty API interface with authentication and CRUD operations
- **creation_helpers.py**: Shared utilities including D20 probability tables, stat normalization, and data sanitization

### Data Flow
1. **Input**: CSV files from FanGraphs/Baseball Reference placed in `data-input/[Year] [Type] Cardset/`
2. **Processing**: Statistics are normalized using league averages and converted to D20-based game mechanics
3. **Output**: Generated card data is POSTed directly to Paper Dynasty API; cards rendered on-demand when URLs accessed

### Entry Points
- **live_series_update.py**: Main script for live season card updates (in-season cards)
- **retrosheet_data.py**: Main script for historical replay cardsets
- **refresh_cards.py**: Updates existing player card images and metadata
- **check_cards.py**: Validates card data and generates test outputs
- **check_cards_and_upload.py**: Fetches card images from API and uploads to AWS S3 with cache-busting URLs
- **scouting_batters.py** / **scouting_pitchers.py**: Generate scouting reports and ratings comparisons

## Common Commands

### Testing
```bash
pytest                    # Run all tests
pytest tests/test_*.py    # Run specific test file
```

### Card Generation
```bash
python live_series_update.py     # Generate live series cards
python retrosheet_data.py        # Generate historical replay cards
python refresh_cards.py          # Update existing card images
python check_cards.py            # Validate card data
```

### Scouting Reports
```bash
python scouting_batters.py       # Generate batting scouting data
python scouting_pitchers.py      # Generate pitching scouting data
```

### AWS S3 Card Upload
```bash
python check_cards_and_upload.py # Fetch cards from API and upload to S3
```

### Analysis and Reporting
```bash
python analyze_cardset_rarity.py # Analyze players by franchise and rarity (batters/pitchers/combined)
python rank_pitching_staffs.py   # Rank teams 1-30 by pitching staff quality
```

### Position Validation
```bash
# Verify position assignments after card generation (recommended after every run)
./scripts/check_positions.sh <cardset_id> [api_url]

# Examples:
./scripts/check_positions.sh 27                                    # Check production
./scripts/check_positions.sh 27 https://pddev.manticorum.com/api  # Check dev

# The script flags:
# - Anomalous DH counts (should be <5 for full-season cards)
# - Missing outfield positions (indicates defensive calculation failures)
# - Mismatches between player positions and cardpositions table
```

### Outfield Arm Ratings (Retrosheet)
```bash
# Generate arm ratings CSV from Retrosheet play-by-play data
python generate_arm_ratings_csv.py --year 2005 --events data-input/retrosheet/retrosheets_events_2005.csv

# Test/validate arm ratings
python test_retrosheet_arms.py

# Output: data-output/retrosheet_arm_ratings_YYYY.csv
```

## Data Input Requirements

### FanGraphs Data (place in data-input/[YEAR] [TYPE] Cardset/)
- **vlhp-basic.csv** / **vlhp-rate.csv**: vs Left-handed Pitching stats
- **vrhp-basic.csv** / **vrhp-rate.csv**: vs Right-handed Pitching stats  
- **vlhh-basic.csv** / **vlhh-rate.csv**: vs Left-handed Hitting stats
- **vrhh-basic.csv** / **vrhh-rate.csv**: vs Right-handed Hitting stats

### Baseball Reference Data
- **running.csv**: Baserunning statistics
- **pitching.csv**: Standard pitching statistics
- **defense_*.csv**: Defensive statistics for each position (c, 1b, 2b, 3b, ss, lf, cf, rf, of, p)

### Retrosheet Play-by-Play Data
- **retrosheet_transformer.py**: Preprocesses new Retrosheet CSV format to legacy format with smart caching
- Place source files in `data-input/retrosheet/` directory
- Transformer automatically checks timestamps and only re-processes if source is newer than cache
- Normalized cache files saved as `*_normalized.csv` for fast subsequent runs
- Performance: ~5 seconds for initial transformation, <1 second for cached loads

### Defense CSV Requirements
All defense files must use underscore naming (`defense_c.csv`, not `defense-c.csv`) and include these standardized column names:
- `key_bbref`: Player identifier (required as index key)
- `Inn_def`: Innings played at position
- `chances`: Total fielding chances
- `E_def`: Errors
- `DP_def`: Double plays
- `fielding_perc`: Fielding percentage
- `tz_runs_total`: Total Zone runs saved
- `tz_runs_field`: Zone runs (fielding only)
- `tz_runs_infield`: Zone runs (infield only)
- `range_factor_per_nine`: Range factor per 9 innings
- `range_factor_per_game`: Range factor per game
- Catchers only: `caught_stealing_perc`, `pickoffs` (not PO)
- Position players: `PO` for putouts (not pickoffs)

### Minimum Playing Time Thresholds
- **Live Series**: 20 PA vs L / 40 PA vs R (batters), 20 TBF vs L / 40 TBF vs R (pitchers)
- **Season Cards**: 50 PA vs L / 75 PA vs R (batters), 50 TBF vs L / 75 TBF vs R (pitchers)

## Configuration

### Database Settings (db_calls.py)
- Production: `https://pd.manticorum.com/api`
- Development: `https://pddev.manticorum.com/api` 
- Change `alt_database` variable to switch environments

### Live Series Settings (live_series_update.py)
- `SEASON`: Current year for live updates
- `CARDSET_NAME`: Target cardset (e.g., "2025 Live")
- `GAMES_PLAYED`: Season progress for live series calculations
- `IGNORE_LIMITS`: Override minimum playing time requirements

### Retrosheet Data Settings (retrosheet_data.py)
Before running retrosheet_data.py, verify these configuration settings:
- `PLAYER_DESCRIPTION`: 'Live' for season cards, or '<Month> PotM' for promotional cards
- `CARDSET_ID`: Correct cardset ID (e.g., 27 for 2005 Live, 28 for 2005 Promos)
- `START_DATE` / `END_DATE`: Date range in YYYYMMDD format matching your Retrosheet data
- `SEASON_PCT`: Percentage of season completed (162/162 for full season)
- `MIN_PA_VL` / `MIN_PA_VR`: Minimum plate appearances (50/75 for full season, 1/1 for promos)
- `DATA_INPUT_FILE_PATH`: Path to data directory (usually `data-input/[Year] [Type] Cardset/`)
- `EVENTS_FILENAME`: Retrosheet CSV filename (e.g., `retrosheets_events_2005.csv`)

**Configuration Checklist Before Running:**
1. Database environment (`alt_database` in db_calls.py)
2. Cardset ID matches intended target
3. Date range matches Retrosheet data year
4. Defense CSV files present and properly named
5. Running/pitching CSV files present

### AWS S3 Upload Settings (check_cards_and_upload.py)
- `CARDSET_NAME`: Target cardset name to fetch players from (e.g., "2005 Live")
- `START_ID`: Optional player_id to start from (useful for resuming uploads)
- `TEST_COUNT`: Limit number of cards to process (set to None for all cards)
- `HTML_CARDS`: Set to True to fetch HTML preview cards instead of PNG
- `UPLOAD_TO_S3`: Enable/disable S3 upload (True for production)
- `UPDATE_PLAYER_URLS`: Enable/disable updating player records with S3 URLs (careful - modifies database)
- `AWS_BUCKET_NAME`: S3 bucket name (default: 'paper-dynasty')
- `AWS_REGION`: AWS region (default: 'us-east-1')

**S3 URL Structure**: `cards/cardset-{cardset_id:03d}/player-{player_id}/{batting|pitching}card.png?d={release_date}`
- Uses zero-padded 3-digit cardset ID for consistent sorting
- Includes cache-busting query parameter with date (YYYY-M-D format)
- Uses persistent aiohttp session for efficient connection reuse

**AWS Credentials**: Requires AWS CLI configured with credentials (`~/.aws/credentials`) and appropriate IAM permissions:
- `s3:PutObject`, `s3:GetObject`, `s3:ListBucket` on the target bucket

## Important Notes

- The system uses D20-based probability mechanics where statistics are converted to chances out of 20
- Cards are generated with both basic stats and advanced metrics (OPS, WHIP, etc.)
- Defensive ratings use zone-based fielding statistics from Baseball Reference
- All player data flows through Paper Dynasty's API with bearer token authentication
- Cards are dynamically rendered when accessed via URL, with nginx caching for performance

### Rarity Assignment System
- **rarity_thresholds.py**: Contains season-aware rarity thresholds (2024 vs 2025+)
- Rarity is calculated from `total_OPS` (batters) or OPS-against (pitchers) in the ratings dataframe
- `post_player_updates()` uses LEFT JOIN to preserve players without ratings (assigns Common/5 rarity + default OPS)
- Players missing ratings will log warnings showing player_id and card_id for troubleshooting
- Default OPS values: 0.612 (batters/Common), 0.702 (pitchers/Common reliever)

### Position Assignment Rules
- **Batters**: Positions assigned from defensive stats, sorted by innings played (most innings = pos_1)
- **DH Rule**: "DH" only appears when a player has NO defensive positions at all
- **Pitchers**: Assigned based on starter_rating (≥4 = SP, <4 = RP) and closer_rating (if present, add CP)
- **Position Updates**: Script updates ALL 8 position slots when patching existing players to clear old data
- Player cards can be viewed as HTML by adding `html=true` to the card URL: `https://pddev.manticorum.com/api/v2/players/{id}/battingcard?d={date}&html=true`

## Common Issues and Solutions

### Multi-Team Players (Traded During Season)
**Problem**: Players traded during season appear multiple times in Baseball Reference data (one row per team + combined total marked as "2TM", "3TM", etc.)

**Solution**: Script automatically filters to keep only combined season totals:
- Detects duplicate `key_bbref` values after merging peripheral/running stats
- Keeps rows where `Tm` column contains "TM" (2TM, 3TM, etc.)
- Removes individual team rows to prevent duplicate player entries

### Dictionary Column Corruption in Ratings
**Problem**: When merging full card DataFrames with ratings DataFrames, pandas corrupts `ratings_vL` and `ratings_vR` dictionary columns, converting them to floats/NaN.

**Solution**: Only merge specific columns needed (`key_bbref`, `player_id`, `battingcard_id`/`pitchingcard_id`) instead of entire DataFrame.

### No Players Found After Successful Run
**Symptoms**: Script completes successfully but API query returns 0 players

**Common Causes**:
1. **Wrong Cardset**: Check logs for actual cardset_id used vs. cardset queried in API
2. **Wrong Database**: Verify `alt_database` setting in db_calls.py (dev vs production)
3. **Date Mismatch**: START_DATE/END_DATE don't match Retrosheet data year
4. **Empty PROMO_INCLUSION_RETRO_IDS**: When PLAYER_DESCRIPTION is a promo name, this list must contain player IDs

**Debugging Steps**:
1. Check logs for actual POST operations and player_id values
2. Verify cardset_id in logs matches API query
3. Check database URL in logs matches intended environment
4. Query API with cardset_id from logs to find players

### String Type Issues with Retrosheet Data
**Problem**: Pandas .str accessor fails on `hit_val`, `hit_location`, `batted_ball_type` columns

**Solution**: retrosheet_transformer.py explicitly converts these to string dtype and maintains type when loading from cache using dtype parameter in pd.read_csv()

### Pitcher OPS Calculation Errors
**Problem**: `min()` function fails with "truth value is ambiguous" error when calculating OB values

**Solution**: Explicitly convert pandas values to Python floats before using `min()`:
```python
ob_vl = float(108 * (df_data['BB_vL'] + df_data['HBP_vL']) / df_data['TBF_vL'])
result = min(ob_vl, 0.8)  # Now works correctly
```

### Outfielders Assigned as DH (Defense Column Mismatch)
**Problem**: All outfielders show `pos_1 = "DH"` instead of LF/CF/RF; cardpositions table has 0 outfield positions

**Root Cause**: Code checks for `bis_runs_outfield` or `tz_runs_outfield` columns in defense CSV files, but Baseball Reference only provides `tz_runs_total`

**Symptoms**:
- 50+ players with DH as pos_1 (should be <5 for full season)
- No LF/CF/RF positions in player records
- Log errors: "Outfield position failed: 'tz_runs_outfield'"

**Solution** (retrosheet_data.py lines 889, 926, 947):
```python
# Wrong - checks batter stats row instead of defense dataframe columns
if 'tz_runs_total' in row:  # ❌

# Correct - checks defense dataframe for actual column
if 'bis_runs_total' in pos_df.columns:  # ✅

# Wrong - column doesn't exist in CSV
of_run_rating = 'bis_runs_outfield' if 'bis_runs_outfield' in pos_df else 'tz_runs_outfield'  # ❌

# Correct - fallback to column that exists
of_run_rating = 'bis_runs_outfield' if 'bis_runs_outfield' in pos_df.columns else 'tz_runs_total'  # ✅
```

**Verification**: Run `./scripts/check_positions.sh <cardset_id>` after card generation to catch this issue

**Additional Fix**: Modified `post_positions()` to DELETE all existing cardpositions for the cardset before posting new ones. This prevents stale DH positions from remaining in the database when players gain defensive positions after bug fixes.

## Outfield Arm Ratings from Retrosheet Data

### Overview
For historical seasons where Baseball Reference's `bis_runs_outfield` is unavailable, we calculate OF arm ratings directly from Retrosheet play-by-play event data using assist rates and quality indicators.

### System Architecture

**Location:** `defenders/retrosheet_arm_calculator.py`

**Key Components:**
1. **Calculation Engine** - Analyzes play-by-play events to measure arm strength
2. **CSV Persistence** - Saves calculated ratings for reuse
3. **Load/Lookup Functions** - Easy integration with card creation scripts

### Formula (Rate-Dominant)

```python
raw_score = (
    (assist_rate * 300) +        # PRIMARY: Assists per ball fielded
    (home_throws * 1.0) +        # Quality: Throwing runners out at home
    (batter_extra_outs * 1.0) +  # Quality: Preventing extra bases
    (total_assists * 0.1)        # Minimal volume bonus
)
```

**Design Philosophy:**
- **Assist rate is king** - 300x weight (primary driver)
- **Quality indicators** - Home throws and batter extra outs add context
- **No throwout rate** - Assists already imply outs (redundant)
- **Minimal volume bonus** - Raw count provides tiebreaker only

### Rating Scale (-6 to +5)

Ratings follow a calibrated distribution (peak at 0 = ~45-50%):

| Rating | Description | Z-Score | Approx % |
|--------|-------------|---------|----------|
| -6 | Elite cannon | > 2.5 | ~1% |
| -5 | Outstanding | 2.0-2.5 | ~2% |
| -4 | Excellent | 1.5-2.0 | ~3% |
| -3 | Very Good | 1.0-1.5 | ~5% |
| -2 | Above Average | 0.5-1.0 | ~10% |
| -1 | Slightly Above | 0.0-0.5 | ~15% |
| 0 | Average | -0.15-0.0 | ~45% |
| +1 | Slightly Below | -0.5--0.15 | ~10% |
| +2 | Below Average | -0.9--0.5 | ~5% |
| +3 | Poor | -1.3--0.9 | ~3% |
| +4 | Very Poor | -1.6--1.3 | ~2% |
| +5 | Very Weak | < -1.6 | ~1% |

**Note:** Thresholds calibrated to actual data distribution after 300x assist_rate weight compressed z-scores.

### Critical Bug Fix: Fielder vs Lineup Columns

**Problem:** Original implementation used wrong columns for fielder positions.

**Wrong Columns (Lineup Order):**
- `l7`, `l8`, `l9` = 7th, 8th, 9th batters in lineup (NOT field positions!)

**Correct Columns (Actual Fielders):**
- `f7`, `f8`, `f9` = Fielders at positions 7 (LF), 8 (CF), 9 (RF)

**Impact:**
- Was measuring arm strength of whoever batted 7th/8th/9th
- Known strong arms (Ichiro, Crawford, Edmonds) didn't show up
- Rankings were based on batting order, not defensive positions

**Fix:** All references updated to use `f7`, `f8`, `f9` fielder columns.

### Data Requirements

**Retrosheet Columns Used:**
- `f7`, `f8`, `f9` - Fielder IDs at LF/CF/RF (CRITICAL: not l7/l8/l9!)
- `a7`, `a8`, `a9` - Assists by position
- `po7`, `po8`, `po9` - Putouts by position
- `brout1`, `brout2`, `brout3`, `brout_b` - Which fielder got the out
- Event descriptions for context

**Minimum Sample Size:** 50 balls fielded per position (adjustable with `season_pct`)

### Generating Arm Ratings

**Command:**
```bash
python generate_arm_ratings_csv.py --year 2005 --events data-input/retrosheet/retrosheets_events_2005.csv
```

**Output:** `data-output/retrosheet_arm_ratings_2005.csv`

**CSV Columns:**
- `player_id` - Baseball Reference ID (key_bbref)
- `position` - LF/CF/RF
- `season` - Year
- `balls_fielded` - Sample size
- `total_assists` - Assist count
- `home_throws` - Throws to home that got outs
- `batter_extra_outs` - Prevented extra bases
- `assist_rate` - Assists / balls fielded
- `raw_score` - Pre-normalization score
- `z_score` - Position-adjusted z-score
- `arm_rating` - Final rating (-6 to +5)

### Using in Card Creation Scripts

**Load pre-calculated ratings:**
```python
from defenders.retrosheet_arm_calculator import load_arm_ratings_from_csv, get_arm_for_player

# At script start
arm_ratings = load_arm_ratings_from_csv(season_year=2005)

# When assigning positions
player_arm = get_arm_for_player(arm_ratings, 'suzui001', default=0)
```

**Calculate on-the-fly:**
```python
from defenders.retrosheet_arm_calculator import calculate_of_arms_from_retrosheet

df_events = pd.read_csv('data-input/retrosheet/events.csv')
arm_ratings = calculate_of_arms_from_retrosheet(df_events, season_pct=1.0)
```

**Integration in retrosheet_data.py:**
```python
# After loading events
from defenders.retrosheet_arm_calculator import load_arm_ratings_from_csv

try:
    retrosheet_arm_ratings = load_arm_ratings_from_csv(SEASON_YEAR)
except FileNotFoundError:
    retrosheet_arm_ratings = {}  # Use defaults if not found

# In create_positions(), replace arm_outfield() call:
from defenders.retrosheet_arm_calculator import get_arm_for_player
arm_rating = get_arm_for_player(retrosheet_arm_ratings, df_data['key_bbref'], default=0)
```

### Documentation

**Detailed guides:**
- `docs/of_arm_rating_improvement_proposal.md` - Full methodology and design
- `docs/HOW_TO_USE_ARM_RATINGS.md` - Integration guide with examples
- `docs/formula_weight_comparison.md` - Before/after comparison
- `docs/CRITICAL_BUG_FIX_fielder_columns.md` - Fielder column bug fix details
- `docs/arm_rating_scale_reference.md` - Quick reference for rating scale

### Key Advantages

1. **Historical Availability** - Works for any season with Retrosheet data (1921+)
2. **Rate-Based** - Prioritizes assist rate over volume (no platoon penalty)
3. **Position-Adjusted** - Normalized within LF/CF/RF for fair comparison
4. **Quality-Aware** - Credits high-value throws (home, preventing extra bases)
5. **Persistent** - CSV output allows consistent ratings across runs
6. **Transparent** - Clear formula allows tuning and debugging

### Validation

**Test script:** `python test_retrosheet_arms.py`

**2005 Results:**
- 300 qualified outfielders
- Distribution: ~1% elite (-6), ~45% average (0), ~1% very weak (+5)
- Known strong arms (Ichiro, Guerrero) properly identified after bug fix
- Assist rate correctly dominates over volume