How Wave Works

Wave is a 24/7 scoring engine that watches the top ~50 crypto perp pairs by volume, scores each candidate across ~8 components, then opens paper trades when the score clears both a regime gate and an execution gate. The trades are real-time against real Binance prices, but the money is paper — no capital deployed until the system proves a 90-day positive edge over BTC buy-and-hold.

Single bankroll, not two. Longs, shorts, and the moonbag accumulation sleeve all share one $100K pool (Spec 2589 architecture). Anything that talks about "$200K combined" is a bug — the integrity audit catches that string in Telegram and report files automatically.

Right now (live from lib/crypto/equity.js):

• Equity: $96415.89 (-3.58% MTM / 0.91% realized)
• Closed trades all-time: 219 — 122 in the v3 cohort (post-rebuild)
• Open positions: 1+ longs · 0 shorts (see /wave for live counts)
• Cash available: $50906 · Invested: $50000 (50% deployed)

Pillar status — live from <pillar>_pillar_status settings keys, written by the integrity audit every 30 min:

Per-component breakdown:

Strip-dead-features setting. When score_strip_dead_features=1 (current state), the Smart Money and Narrative Alert contributions are forced to zero in total_score regardless of what the per-symbol scorer returned. Re-enable cleanly by flipping their individual *_enabled flags AND the strip-dead setting back to 0.

onchain_score is null for non-DeFi tokens as of Spec 2599 Phase 1#3 — distinguishes "no DefiLlama mapping" (N/A) from "mapped but scored 0" (real signal). Eliminates phantom-zero dilution in total_score.

Several capture-only systems exist purely to tell us whether Wave is making the right calls. None of them touch a live score. They record what happened, what we skipped, and what we are watching, so a decision to add weight later is backed by data instead of a hunch.

The discipline is deliberate: a new signal records its read with zero influence, we compare it against what actually happened, and only then do we consider giving it weight. The first review window has passed and every captured source still sits at weight 0 — none has yet earned a live weight.

Every candidate passes through two independent gates. A signal can be EMITTED (score ≥ threshold) but still BLOCKED from execution by the regime floor. The two were coupled until Spec 2598 — shorts now have their own execution floor so a chart-only short setup can fire in flat tape without lowering the long-side defensive floor.

Gate 1 — score-emit threshold: shorts_min_score_after_3_trades=3 (default 4 pre-shorts-pilot). Anything below this never even enters as a signal candidate.

Gate 2 — regime execution floor: long-side reads min_score_neutral=5, short-side reads shorts_min_execution_score=4 (Spec 2598 decoupling). Both havestrict_floor_enforcement kill-switch.

Plus per-trade vetoes: cluster cap, day-of-week filter, funding-APR veto, onchain-zero veto, gap-through guard. Any one of these short-circuits an otherwise-qualifying signal.

Every rule below has a settings-key killswitch. Flipping it to 0 takes effect on the next scan tick — no redeploy.

The system catches its own bugs before David eyeballs them. Each audit run cross-checks every live API endpoint, every snapshot table, every score-feature variance against lib/crypto/equity.js(the canonical truth). Dedup state machine prevents repeat alerts inside a 4-hour cooldown.

Run-history table: every wave/crypto cron writes a row to crypto_cron_runs via the shared tripwire helper. Status, duration, output_summary_json, error_message. 30-day retention. Replaces pm2-log-grepping with a queryable history.

P&L canonical source rule (Spec 2599 Phase 8): every $/% number that gets shown or spoken to David comes from lib/crypto/equity.js. scripts/check-pnl-canonical.sh +.git/hooks/pre-commit fail any commit that introduces inline equity math outside the canonical writers.

Tier-1 auto-fix: snapshot drift > $1K vs equity.js auto-repairs the row; stale crons with last_success_at > 2×expected_interval auto-restart via pm2. Cooldown 3/24h per fix-type, 10/24h global cap. Killswitch integrity_audit_autofix_enabled.

crypto_paper_trades.cohort column added by Spec 2599 Phase 9. All trades before the rebuild ship are tagged pre_rebuild; new trades opened by either scanner now writecohort='v3'. This is NOT data deletion — pre-rebuild trades are still queried for all-time lifetime numbers, but they don't feed Phase 2 weight tuning.

Phase 2 trigger: when cohort='v3' AND status='closed' count reaches n ≥ 30 AND every scoring feature has been alive for the full window, the daily audit auto-proposes re-weights via crypto_audit_proposals (Tier 2 — Silas review, no auto-apply).

Recap audio explicitly distinguishes the two scopes: "Lifetime, all trades: -$1939 over 71 trades" (canonical, equity.js) AND "V3-era cohort: -$X over 122 trades" (subset, labeled). No more pre-rebuild numbers wearing the word "Lifetime."

Current v3 cohort count: 122 closed trades (target n ≥ 30 to unlock Phase 2 tuning).

Telegram cadence: morning health DM at 08:00 PT is the canonical "is Wave OK?" digest. Integrity audit Telegram currently muted per operator flag (integrity_audit_telegram_enabled=0) — runs silently, writes reports, alerts when flipped back on. Audit MP3s land at /recaps/<date>-audit.mp3 regardless.

The original rebuild on 2026-05-17 added 13 rails over one day. Specs #2826 and #2827 layered another ten on top through May 21 / May 22 — portfolio drawdown breaker, BTC flash breaker, counterfactual shadow replay, wiring drift audit, BTC tape filter, counter-trend guard, daily loss cap, funding entry-gate, news weight set to zero, and the inverse-side shadow loop. Each row shows what it does, why it exists (with backing data where the case was empirical), and the exact DB key to flip if a rail needs to be disabled. Cohort marking is active throughout — pre-rebuild and v3 trades are tagged separately so post-rebuild metrics don't get muddied by legacy data.

Watchdog retune (May 31): liveness, near-miss, and dead-scoring monitoring. A quiet day no longer triggers false alarms now that the score floor is 6; alerts fire only on a stalled scanner or genuinely dead scoring.

All detectors live in lib/crypto/patterns.js. The aggregators (detectBullishPatterns, detectBearishPatterns) return the top confidence match plus a 0–3 score that scales into total_score. Killswitches: bullish_pattern_enabled=1, bearish_pattern_enabled=1.

Cohort marking is active — pre-rebuild and v3 trades are tracked separately so the detectors' post-rebuild contribution can be measured without legacy contamination.

The diagnostic on the first 116 trades showed that exits work (trail stop is +$6,985) but the entry direction is the leak. The May 22 batch attacks both sides of that problem: new data feeds (BTC dominance, liquidations, funding extremes, event calendar) give the scanners context they didn't have before; new shadow strategies (liquidation fade, Kelly sizing, bankroll compounding, slippage-modelled backtest, pre-listing watchlist) run alongside the live system without affecting real PnL until the data says they earned the upgrade.

Data feeds (always on)

Shadow strategies + sizing helpers

The full strategy plan that motivated this batch lives at /root/cc-summaries/2026-05-22-boostio-10x-strategy-plan.md. Promotion criterion across the board: ≥30 closed shadow trades with WR + payoff that beats the live cohort by more than the round-trip cost. Until then, the live book runs on the existing engine and the new strategies record themselves into crypto_shadow_trades + their per-strategy log tables.

Research engine, honest outcome. We stood up a full historical-cycle research pass and an analog engine that finds past look-alikes for the current tape, framed around a "Master Trader OS" idea of keeping signal research strictly separate from execution. The honest result: these are context and risk tools, not a proven buy/sell edge. They help us reason about where we are in a cycle; they do not place trades.

We did not wire a fake edge. We swept a portfolio simulator plus position-sizing, exit, veto and entry variants, each tested out-of-sample against a plain fixed-fractional baseline. None beat the baseline out-of-sample, so we wired exactly none of them into live trading. The discipline is the point — we do not ship a strategy change just because it looked good in-sample.

What we actually built — all capture-only research data quality:

What did not change. No trading, scoring, sizing, entry, exit, leverage, order or real-money setting was touched. The delisting/dead-coin events table stays forward-only, and anything merely medium-confidence, conflicting, not-found, unverifiable or tokenized is not promoted — only contract-proven crypto graduates into the lifecycle.

Wave ships in phases. Below is everything that someone might reasonably assume is part of a "real trader" stack but is deliberately deferred — with the reason for each, so nothing surprises you later.

No fixed dates. The roadmap moves when the data does — n ≥ 30 v3 trades unlocks Phase 2, not a calendar. If a phase needs more time, it gets more time. If it ships earlier, it ships earlier. Either way the morning health DM tells David where we are.

Wave's trade book is built for momentum: it enters on a score, manages a stop, and exits. That is the wrong tool for a different job, which is accumulating coins you believe in near a cycle bottom and holding them for the move up. The moonbag sleeve is that second tool. It is part of Wave, funded from the same one $100K bankroll, but its bags follow different rules: no hard stop, no take-profit ladder, no time-based exit, and a deeper drawdown tolerance. The momentum manager skips them, so the two styles share one bankroll without interfering. Every bag is a real, tagged Wave long, so it counts in the bankroll's equity and shows up, marked, in the Open Trades table.

How the picks are chosen

The basket is the top Scout candidates by score: deep below all-time high, scored on value, quality, structure, narrative and entry. Every position is tagged with why it is held: the moonbag hold-type (longer hold, higher drawdown tolerance), the signal source (Scout), the score at entry, the live current score, and the percent below ATH. So you can always see at a glance which holdings are patient bags and exactly why each was picked.

The list stays current, not frozen. Scores refresh every run, and a candidate only stays eligible if it still meets the moonbag score bar AND was re-scored recently. A high score that has gone stale (not re-scanned in days) is excluded, so the sleeve holds today's high-scorers, not last week's. Picks that are not on Binance are priced live via CoinGecko, so the best Scout signals are not left out just because they sit off our main exchange feed.

How it accumulates

Built for a no-leverage, hold-the-cycle style: no tight stop, no time-based exit. It deploys a first tranche now, then adds on weakness (pre-planned dip tranches below the entry reference) so being early lowers the average cost, and adds one more tranche when Bitcoin reclaims its rising 50-day average (the trend-turn confirmation). It caps total deployment to keep dry powder in reserve and posts every add to the crypto channel. Nothing is auto-sold: a bag that decays below the bar is flagged for review, and the decision to take profit stays manual.

Paper only, real money off, killswitched (accumulation_enabled). The bags are first-class Wave longs in crypto_paper_trades tagged cohort=moonbag, so they count in the one bankroll's equity and appear, tagged, in the Open Trades table under the Moonbag filter. They sit alongside the on-chain bottom score that reads real MVRV, SOPR, NUPL, realized price and exchange reserves plus a weighted read of Blockchain Backer's posts.