Active Experiments (2026-06-14 cohort)
1. Independent Corroboration Confidence Upgrade
Uses Polymarket + Kalshi ladder + realized-vol agreement to upgrade Deribit-derived confidence (LOW→MED, MED→HIGH) by at most one tier.
logOnly (default) gate-safe
Per-leg attribution recorded. Never rescues map-haircut high-fairP BTC-YES. Observe phase only.
2. YES Suppression Accumulator + Grader
Captures at scan (evRawYes, sideRecRaw, calibratedFairP, poly*, ladder) for suppressed midrange YES (0.5–0.7 fairP). Grades at settle. Tests whether high-corroboration suppressed YES outperform low-corroboration ones.
diagnostic-only forward-only since ~2026-06-13
3. Confidence-Cycle Telemetry
Samples confidence tier + ivMethod distribution + live Deribit nearest-expiry gap every hour. Aggregates by UTC hour-of-day and day-of-week to confirm the daily HIGH window and weekend hard-LOW cliff.
read-only edge-triggered Discord alerts
All experiments are explicitly observe-only / diagnostic. No placement gates, minEv floors, or minConfidence thresholds are altered. Results surface via MCP tools and the two small DDB tables.
Live Experiment Config (pulled)
These are the knobs for the corroboration upgrade. Currently logOnly + enabled. The weekend factor tightens tolerances when the daily chain has lapsed.
Live high-corroboration candidates (low-conf YES with positive edge)
Client-side computation using the same legs as the upgrade classifier (poly eligible agreement, directional, realized, ladder context via polyConf). These are the ones the experiments would "notice" even if Deribit confidence is low. Good to watch in the game page or review manually while the accumulator accrues.
YES Suppression & Corroboration (crypto_yes_suppression)
By Cohort (Underlying × Tenor)
By Corroboration Bucket (high / medium / low) + Sub-bands
Corroboration proxy built from: Polymarket agreement (eligible only), directional signals, realized-vol shadow, local ladder coherence. High = strong multi-source support despite Deribit rejection.
Confidence Cycle Telemetry (crypto_confidence_cycle)
By UTC Hour-of-Day (empirical HIGH window)
HIGH reachable only when nearest Deribit expiry gap ≤6h (and near-ATM + dense grid). Weekend cliff when >3d from nearest listed expiry.
Opportunity Telemetry — both sides (crypto_opportunity_telemetry)
Aggregate counterfactual ROI + what we skipped
By side (YES / NO) — the NO-side hypothesis
Sparse hourly capture (max(evRawYes,evRawNo) ≥ 0.05 OR accepted OR upgrade-candidate), graded BOTH sides at settlement. ROI = recommended-side counterfactual P&L / price, deduped to one decision-relevant scan per ticker. sufficient turns true once a slice reaches minN. This is the forward log that makes the NO-side + directional (polyDirectional) hypotheses testable.
My observations (local to this browser)
Persists only in this browser tab's localStorage. Great for capturing patterns while the formal accumulator (n) grows. Use it to remember "what the experiments are telling us" without relying on memory.
How to watch while the data accrues (the hard part)
- Daily ritual: Open this page + the main Game page side-by-side. Refresh here (or let auto do it). Note patterns in the "My observations" box below.
- Watch for signal in the live opps section: High-corr low-conf YES (especially ETH midrange) that have poly agreement + ladder + realized legs firing. Those are the ones the classifier "notices". Track them in the Game ladders (fields are attached).
- YES suppression verdicts: When n grows in a cohort, look for over-suppression flag on ETH midrange vs no flag (or justified) on BTC high-fairP. That tells us if corroboration separates the recovery from the leak.
- Cycle data: The HIGH window (~6h before each 08:00 UTC Deribit daily) should show more medium/high and more opps. Weekend cliff should be visible as book-wide low with the proxy still able to highlight some.
- Per-leg: Use the agreeCount + deltas on rows (in the suppression tool output or candidate data) to test "does it still work without the noisy polyDir leg?"
- When to promote: When a cohort hits solid n (your prior ~20-50) with clear lift on high-corr ETH YES and zero dangerous lift on BTC high-fairP (still protected by mapHaircut). Then flip confidenceUpgradeLogOnly to false in config (full read-modify-write + explicit approval per the rules).
All of this is the "measure before build" phase. The formal sample is the accumulator + telemetry. This page + the notes box is your memory aid so you don't have to hold it all in your head.
Data is pulled live from the MCP surface (which reads the dedicated accumulator/telemetry tables). Forward-only since the 2026-06-13/14 instrumentation. Thin n until more markets settle.
Related:
Game •
Config •
Spec • CHANGELOG (2026-06-14 section)