2023-07-22 to 2026-07-18
79 rebalances, each reading a trailing 21-day news window
Judges the news gathering, not portfolio gains: completeness of the historical pull and whether the calendar has gaps, upstream of the curator and free of PageRank-mooning. The pool the curator actually receives is the ranked top-100 (salience + authority + per-wave weighting) — the shape that resembles live WebSearch results, not a raw dump. Each article carries a lede: first a look-ahead-clean Wayback snapshot (≤ as_of); where Wayback has no capture, a live-fallback fetches today’s page (LOOK-AHEAD-BIASED, tagged separately). Plot 6 tracks both: the clean join-rate and the clean+live coverage. 79 rebalances over 21-day windows.
Raw data (inspect any file): data/curator_runs/gkg-3yr-geosplit/ — per-window <date>-pool.json (ranked article list; lede = clean Wayback join, lede_live = biased live-fallback, lede_sources = the clean/live/none split), plus _corpus/ (full gathering, one file per day).
The user-set investor_profile.md knobs relevant to news retrieval (the optimizer knobs are in the Curator Backtest + Sweeps DBs).
| Backtest window | 2023-07-22 → 2026-07-18 |
| Rebalance cadence | biweekly |
| News lookback | 14 days |
| Max articles / pool | 100 |
How many of each biweekly pool's articles the curator read with a clean, look-ahead-safe Wayback lede, a look-ahead-biased live-fallback lede, or title only (no lede). The clean-Wayback share is what a look-ahead-honest backtest relies on.
The full per-wave search-term lists behind plot 8 live in retrieval_config.json (wave_keywords), which is gitignored, so this link resolves only when you are viewing a local copy of the dashboard.
52% of the 7,792 pooled articles (4,071) have a known author.