Active Experiments & Gates
Curated list of what we're testing. Status badge + hypothesis + next-check trigger. Cohort cards pull live ROI from backtest_analyze over settled positions (full history); gate cards (logOnly) link to their Studio replay page since they have no positive cohort.
Live ROI is full-settled-history for the targeted segment — it is NOT a forward post-change cut. Read it as "is this segment still red/green," then drill into the Studio page or a dated backtest for windowed detail.
Live config overrides (server-stored)
Auto-pulled from user-config_get — these are the server-stored overrides only. Userscript-baked defaults (e.g. the MLB live max-2 leg cap and trajectory.gateMode) are enforced in-script and do not appear here, so an empty section ≠ "no gate." Server-side caps/blocks below never go stale relative to the curated cards above.
Leak re-test watchlist (live leak_scan, 45d · all tiers)
Scans 45d at tier:"all" (candidate + watchlist) so the graduation pipeline actually shows — a short window + candidate-only scan reads "clean" while real red pockets sit one tier down. The holdout column is the discriminator, not ROI: persists− = negative in BOTH the Mar–Apr reference window and recent = a real leak that survived the fixed holdout (graduation candidate); flip = sign-flipped across windows = variance, do NOT act (e.g. the MLB run-line dog pocket); recent = too few in-window rows to validate (forward-only, watch). Deep-red ROI with flip/recent is a concentration, not a confirmed leak. persists here is split-sample within the scan window only. Always cross-check with a dated backtest before scoping a block.
My observations (local to this browser)
Persists only in this browser's localStorage. Use it to capture patterns between weekly re-tests.
How to read this
- Cohort cards (applied/dormant blocks & caps): the live ROI tile is the targeted segment over all settled history. A red segment with the block applied = working as intended (we stopped adding to it). A red segment dormant/unapplied = decide whether to pull the trigger.
- Gate cards (logOnly): no positive cohort to score — these are observe-only gates. Use the linked Studio page to see rejection counts / replay impact before promoting
logOnly → enforce.
- Leak watchlist: the graduation pipeline. Deep-red ROI is necessary but NOT sufficient — the
holdout column is the discriminator: only persists− (negative in both the Mar–Apr reference and recent) with adequate n is a real graduation candidate. flip = variance, recent/sparse = unvalidated (forward-only or history-only). persists is split-sample within the scan window only, not the holdout. Always cross-check with a dated/forward backtest before scoping a block.
- Never loosen: nothing here auto-applies. Caps/blocks are tightening-only; promotion to enforce is a manual, reviewed config write.