Strategy reference

This page mirrors STRATEGY.md, the repo's live spec, in plain language. Values shown are the current shipped defaults from config.toml — several were deliberately tightened after live book review, so where the spec's original bracketed default differs, this page shows what actually runs. Everything here is hot-reloadable via Settings / config.toml.

Philosophy: capital preservation first. One-sided SOL bid-ask below price as the default entry, mechanical exits instead of conviction-holding, PnL denominated in SOL. Wherever a human would "ask the group," the bot is strictly more defensive.

1. Scanning

Runs every 60s:

1. Pool sweep — GET dlmm.datapi.meteora.ag/pools, non-blacklisted pools with TVL > $5,000, sorted by 30-minute fee/TVL, first 3 pages.
2. Dedupe to canonical token — group pools by token mint. If multiple tokens share a symbol (copycats), only the one with the highest 24h volume is considered; the rest are ignored for 24h.
3. Best pool per token — highest fee_tvl_ratio_24h among that token's pools that pass the pool gates. One pool per token.
4. Output: scored candidates → vetting → entry queue (best score first, ties broken by younger pool).

Optional discovery enrichment: GMGN trending and smart-money feeds (needs GMGN_API_KEY; auto-off without one — how to get a key). These only add score bonuses (capped at +10 total) — never a substitute for vetting.

2. Pool gates (hard)

A candidate pool must pass every gate:

Gate	Current default	Why
TVL	≥ $5,000 and ≤ $2,000,000	Below: no routing, arb-only. Above: fees too diluted for meme mode
Market cap	≥ $100,000 ($100–200k routes to the micro sleeve)	Hard floor
Fee/TVL 24h (or lifetime if pool < 24h old)	≥ 20%/day	The meme-pool heat threshold
Fee/TVL 30m, annualized to daily	≥ 10%/day	Catches pools that were hot but died
Volume 30m	≥ $25,000	Fees need flow now, not this morning
Volume trend vol_1h / (vol_24h/24)	≥ 0.8	Current hour at least ~80% of the daily average — not in freefall
Base fee	0.2%–5%	> 5% = arb-only pools
Bin step	≥ 80 for tokens < 7 days old	Wide coverage with fewer bins
Fee collection	prefer_quote — quote-only (SOL) pools get a score bonus; both-token pools stay eligible	Quote-only pools pay fees pre-converted to SOL. Hard modes quote_only / both_only / any exist in config
Quote token	SOL required	Accumulate-SOL thesis; stable/USDC pairs are out of scope
Pool price vs Jupiter quote	divergence ≤ 2%	The oracle-glitch / empty-pool trap
Pool share	our position ≤ 20% of pool TVL	Otherwise we'd be the exit liquidity

3. Token vetting gates (hard)

Computed fresh at entry time from RPC + the RugCheck free API:

• Mint authority revoked, freeze authority revoked.
• Token program: SPL Token, or Token-2022 with no extensions beyond metadata (no transfer fees or hooks).
• Single holder ≤ 15% of supply; top-10 holders ≤ 40% — both excluding labeled pool vaults, lockers, and burn addresses.
• Insider/funding clusters ≤ 10% of supply (same-funder wallet clustering + launch-slot snipers).
• Creator has zero rugged tokens in our DB or RugCheck's creatorTokens. One strike = permanent creator blacklist.
• RugCheck score_normalised < 41 (their "Danger" line) — used as a veto only, never as approval.
• Token age ≥ 45 min (survive the instant-rug window) and ≤ 14 days in meme mode. Age is the mint's age via RugCheck, not the pool's.
• Not on our blacklist (§8 below).

Each vetting check has a master on/off switch in Settings (age min/max, insider gate, holder gate, RugCheck veto, creator-rug gate, GMGN honeypot/sell-tax gate). Off = skip that hard fail; thresholds still apply when on.

4. Timing filter (soft — feeds the score)

These never override a hard fail; they only tilt the score (how big, how soon):

• Price not in freefall: 15m return ≥ −20%.
• Not top-blasting: price within 3% of ATH plus stoch-RSI-style overextension on 5m candles → penalty. The bot enters after a retrace leg, not into a vertical.
• Buy/sell imbalance from OHLCV volume: heavy net selling in the last 5m → penalty.
• Volume ignition: last 5m candle ≥ 3× the trailing-hour average → bonus.

5. Opportunity score (0–100)

Weighted blend, drives sizing and queue priority:

Component	Weight
Fee/TVL momentum (30m vs 24h)	30%
Volume/TVL turnover	20%
Vetting softness (holder distribution quality, holder growth, maker diversity)	25%
Timing (§4)	15%
Pool structure (bin-step fit, fee tier vs competition)	10%

GMGN trending bonuses (sustained +8, emerging +4, fading +3) and smart-money flow bonuses stack on top but are capped at +10 combined; heavy smart-money net selling (> $5,000) applies a −8 penalty.

6. Entry execution

Default shape — one-sided SOL, BidAsk, below current price:

1. Compute swing high/low from 5m OHLCV over the pool's life (max 24h lookback).
2. Range top = active bin. Range bottom = the shallower of: fib 0.786 retracement of the swing, or −50% from current price. Floor of −40% minimum depth — never a thin sliver. (The planner also derives a cap from the P0 crash threshold, so no bins are planned below where P0 would already have fired.)
3. Translate to bin IDs. A DLMM position account spans ≤ 69 bins; if the range needs more, split into at most 2 position accounts.
4. Bin-array rent (~0.075 SOL per array, non-refundable when arrays are new): soft budget one array — shrink the range first. Hard budget two arrays (0.15 SOL) only when score ≥ 80. The bot quotes actual uninitialized arrays on-chain; RPC quote failure falls back to the worst-case estimate (fail closed). Never more than two arrays.
5. Open with StrategyType.BidAsk, SOL side only (totalXAmount = 0).
6. Tx policy: active-bin liquidity slippage 5% (≈5 bins at bin step 100 — the earlier 1% setting produced 100% of live open failures as ExceededBinSlippageTolerance). Priority fee auto from recent fees; a failed program simulation never resends the same tx; 3 network retries, then abandon and re-quote. On bin-slippage rejection the live path rebuilds the position rather than spamming the same bad transaction.
7. Second tranche: for score ≥ 85, an additional BidAsk pocket below the primary, sized at 50% of the primary, down toward −70% (clamped by the P0 safety floor). Skipped on the micro sleeve, when the primary already fills the floor, or when slots/size floors block it. Tranches count toward max open positions.

Exit swap path: on close, the token side converts to SOL via the Meteora Zap SDK (Jupiter V6), with a lite-Jupiter fallback. Normal exits use 50 bps swap slippage; P0 safety exits use 1000 bps (speed over price).

7. The P0–P5 exit ladder

Each open position is polled every 15s. Rules are checked in strict priority order — first match wins:

Priority	Name	Trigger	Action
P0	Safety exit	Any of: pool TVL −40% in 10 min · single wallet sells >3% of supply in one tx · tracked insider cluster distributing · new whale >10% · metadata changed or RugCheck flips to Danger · price −60% from entry (at any age — there is deliberately no time window)	Remove 100% + close immediately, market-dump token→SOL at 10% slippage, blacklist token + creator, log incident
P1	Stop loss	Mark-to-market in SOL (both sides + unclaimed fees) < entry × 0.75	Close, swap token→SOL, realize loss, 24h re-entry cooldown. No conviction override — ever
P2	Rotation	Fee rate (30m annualized) < 5%/day for 3 consecutive polls, or 30m volume < $5,000; or position age > 48h (then it stays only if it would qualify as a fresh entry today)	Close — dead weight; capital rotates to the queue
P3	Above range	Price > range top by 5%, sustained	Close via zap, recover rent, record PnL. See win vs missed below
P4	In range	Earning	Claim/bank fees; escape hatch; profit lock (below)
P5	Below range	Price under the bottom bin — position is 100% token	Hold 15 min grace (wick tolerance), then close, swap all token→SOL, realize the loss, 24h cooldown. Coinciding safety signal escalates to P0

P3: win vs missed

Two cases that look the same but mean opposite things:

• Win — price dipped into the range, then recovered above it. Every bin it climbed back through sold our tokens above acquisition price; the position ends ~100% SOL with round-trip profit + fees. Exit after 10 min sustained above range (short, so follow mode can arm).
• Missed — price pumped without ever entering the range. The SOL was never touched: no profit, no loss, just idle capital. Exit after 45 min sustained (a shorter timer was churning ~0.002 SOL rent per round trip on slots the bot wasn't refilling).

Re-entry after P3 (anti-chasing)

• The token goes back through the full pipeline as a fresh candidate — including the timing filter — so re-entry happens after a retrace signal, never into a vertical pump.
• Ladder decay: each successive re-entry on the same token within 24h sizes at 0.75× the previous; max 2 re-entries. A re-entry must still clear a 0.2 SOL viability floor.
• Rate limits: ≤ 2 rebalances per position per 6h; skip entirely if projected rent + tx cost > 25% of fees earned so far.
• House-money rule is off (it banked notional profit with no release path; deployable only ever fell).

P3-F: Follow mode

Optional overtime after any P3 close (win or missed), designed not to chase — the gate set below was the only configuration at/above breakeven in simulation over the recorded post-exit price paths:

• A P3 close on a main position arms a chain.
• A leg opens only when all hold: pool 30m volume ≥ $100,000 (4× the normal entry floor) · current 30m and 1h fee rates annualized ≥ the 24h gate (bypassing the stale 24h average) · price retraced 15% from the post-exit/post-high peak · fresh token vetting passes.
• Leg shape: one-sided SOL BidAsk, 30% deep (tighter than the 40% default), top at current price, escape hatch disabled.
• Up-only: after a leg closes up-and-out, the chain re-arms only once price makes a new chain high. This condition alone separated +EV from −EV in the sim.
• Chain ends on: any non-P3 leg close · 3 legs · cumulative chain PnL ≤ −0.075 SOL · 3 consecutive polls under the normal volume floor · 12h age · blacklist or vet fail. A failed leg open cools the chain 300s.
• Legs are fixed at 0.25 SOL, exempt from the re-entry ladder, and excluded from the main Kelly ledger — the mode must earn bigger sizing with its own closed-leg evidence.
• While a chain is live, the normal pipeline skips that token: one owner of re-entry timing per token.

8. P4 detail: fees, escape hatch, profit lock

• Claim when unclaimed fees ≥ max(0.05 SOL, 20× estimated tx cost), or every 4h, whichever first. When price first drops below range, fees are claimed immediately (min 0.005 SOL) so token-side fees convert to SOL at the top of a dump, not the bottom.
• Fee destination bank: token-side fees swap to SOL at claim time and bank to the wallet. (A compound mode exists in config for pools scoring ≥ 70 but banking is the shipped behavior; majors also bank only.)
• Escape hatch: if price fell through > 60% of the range depth and then recovers to the upper 25% of the range → close and re-enter. This realizes fees and resets near the average acquisition price instead of round-tripping the whole dip. Disabled on follow legs and majors.
• Profit lock: if mark-to-market ≥ entry × 1.30 while still in range → withdraw 30% of liquidity (position stays open and earning) and zap it to SOL. Fires at most once per position. A floor under strong runners without giving up the fee stream.

9. Majors sleeve (separate playbook)

SOL-quoted majors/alts on a symbol allowlist (PUMP, ANSEM, JTO, BONK, WIF, RAY, JUP), discovered by their own sweep. Key differences from meme:

Aspect	Majors behavior
Shape	Spot (uniform bins), 12% below / 6% above price — SOL biased to the downside
Timing entry	RSI(14) ≤ 45, or price in the bottom 40% of the 24h swing; never above 75% of the swing
Size	Fixed 0.75 SOL, max 1.5 SOL per position, max 1 slot, ≤40% of wallet
Stop loss	entry × 0.60 (wider than meme — spot holds inventory)
Below-range grace	120 min (vs meme 15)
Escape hatch / profit lock	Off — majors hold through dips
Rotation	Much slower: 20 consecutive polls under a 0.05%/day fee floor; max age 168h; above-range sustains 240/480 min
Fees	Bank only, claim min 0.02 SOL

Majors run after meme entries and only when ≥ 2 slots stay free for memes. Token age must be ≥ 7 days.

10. Blacklist & skip rules

Kept in the blacklist table with reason + expiry:

Kind	What lands there
Permanent	Creators with any rug in history; tokens that triggered P0; Token-2022 with transfer-fee/hook extensions
24h cooldown	Copycat losers of symbol dedupe; tokens that hard-failed vetting; tokens exited at a loss
Structural skips (never candidates, not blacklisted)	Base fee > 5% pools; non-SOL quotes (incl. SOL-USDC); pools where our position would exceed 20% of TVL

Every skip is logged with the failing gate in the decisions table — the tuning dataset. Skipped candidates get outcome backfill (did the ones we passed on moon or rug?) so gates can be tightened or loosened with evidence, not vibes.

11. Accounting

• SOL is the unit of account (USD stored as a snapshot column for readability).
• Per position: fees claimed + fees unclaimed + (exit value − entry value) − rent − tx costs, all in SOL. Realized PnL is measured wallet delta — what actually left and returned — never our own intent.
• Reconciliation: on startup, the chain wins. The bot enumerates the wallet's actual DLMM positions, diffs against the DB, repairs, and logs discrepancies.

How it worksRisk & sizingConfigurationSTRATEGY.md