Frankel

Frankel 🏇 — AI Racing Analysis Platform

Owner: Frankel (pm-frankel) Status: Active — pipeline redesign in progress

Overview

UK horse racing analysis product. Subscription model. Revenue engine for the org.

Current State

Project files at /home/winnie/projects/Frankel/
Real DB: form.db (57MB, 13,522 horses)
Python venv active, all deps installed
Pipeline redesign approved by Chris (2026-02-20)

Architecture

Python pipeline (orchestrate_v4.py) processing race data
Betfair integration for odds data
Sonnet/Stage2 AI calls for race analysis
Per-race log files in projects/Frankel/logs/
jCodeMunch pre-indexed: 7,644 symbols

Pipeline Redesign

Spec: /home/winnie/projects/Frankel/research/IMPLEMENTATION_SPEC_v2.md

Phase	Description	Status
Phase 0 — Quick wins	Per-furlong split preservation, enhanced DI output, raw data noise cap	✅ COMPLETE (Stack, 2026-02-20)
Phase 1	TBD	In progress
Phase 2	TBD	In progress
Phase 3	TBD	Pending
Phase 4	TBD	Pending

Phase 0 validation: Dry-run 7 races Lingfield 2026-02-21 ✅

Recent Work Log

2026-02-25 — Betfair Time Matching Fix

orchestrate_v4.py updated to match Betfair race times using 12-hour meeting times with optional BST/GMT drift handling.
Verification: Kempton R01 now matches Betfair 17:05 snapshot; best_value no longer auto-skips there.
Remaining gap: Lingfield R01 still skips best_value due to missing 13:12/13:42 Betfair snapshots.

2026-02-25 — Kempton python-only R02–R07

Executed orchestrate_v4.py --python-only sequentially for Kempton races R02 through R07.
R02 followed expected skip path via race-type filter; R03–R07 completed Stage 0 successfully.
Verification: zero Sonnet/Stage2 calls.

2026-02-25 — Combined Kempton + Lingfield python-only run

Run batch run_report_20260225_174334.csv completed Kempton (R02–R07) and Lingfield (R01–R08).
Outcome: 13 PASS races, 1 expected skip (Kempton R02 excluded by race-type filter).
API usage: sonnet_calls_total=0, stage2_calls_total=0 across all PASS races.
Best-value populated in: Kempton R04 (Bronte Beach), Lingfield R03 (Mayflyer), R05 (Miss Moonshine), R06 (Fitzmaurice), R07 (Back Before Dark).

2026-02-25 — V5 Kempton R01 Phase 1 + Phase 2 test

Ran both V5 tool-assisted orchestration tests for Kempton R01; both exited 0.
Stage 1 fell back in both runs due to merge parse failures.
Stage 2 encountered transient Anthropic 429s but recovered automatically.
Next action: investigate Stage 1 merge parse failures, rerun with improved settings.

Model Policy

All Claude Code invocations for Frankel must use --model opus. Never downgrade to Sonnet for this project.

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude \
  --teammate-mode tmux \
  --dangerously-skip-permissions \
  --model opus \
  --max-budget-usd 5 \
  -p "Frankel task brief..."