How it works · 2026

Seven stages. One trade.

From pump.fun launch to on-chain memo proof, a Kabal trade moves through seven stages in under one second. Every stage is deterministic. Every stage is logged. No LLM in the hot path.

The pipeline

1
Sense · Bumble + Vision
Watch the chain
Bumble scans every pump.fun launch in real-time, pulling metadata, deployer history, and bundle signatures. Vision reads Twitter, Telegram, Farcaster for narrative heat. They publish enriched CANDIDATE_TOKEN events to the bus.
Websocket feed Enriched metadata ~200ms
2
Filter · Ghost Pass
Five checks, zero mercy
Bundle detection · deployer age · LP-lock status · insider percentage · sniper clustering. Any failure and the candidate gets dropped before traders even see it. ~78% of launches never pass the ghost filter.
5 filters Hard drop on any fail ~80ms
3
Decide · Trader squad
Four specialists, one vote each
Reaver (PUMP_PHASE), Ash (BREAKOUT), Myth (NARRATIVE), Ravage (TRENDING) each score the candidate against their own strategy. A trader only fires if its internal confidence passes threshold. Traders do not coordinate — by design.
4 strategies Independent decisions ~50ms
4
Gate · Sentinel + Aegis
Three tiers of veto
Tier 1 (per-trader caps) · tier 2 (Sentinel cross-position check) · tier 3 (Aegis system-level guard). Any fail → trade dropped, reason logged. Sentinel can veto for correlation, drawdown, cold-start cap. The only three places that can stop a trade mid-flight.
Per-trader Squad-level System-level
5
Execute · KKM bridge
Sign and send
Trade instruction + SPL Memo are bundled into one transaction. Per-tenant wallet signs. Sent to Solana through the KKM broker path. Confirmation in ~400ms on healthy RPC.
Atomic txn Memo attached ~400ms
6
Prove · On-chain memo
Trail of receipts
SPL Memo encodes agent · strategy · token · sizing · ghost result · Sentinel verdict · session. Publicly verifiable. Append-only. If there is no memo, there was no trade. See /proof.
Verifiable ~200 bytes Append-only
7
Learn · Librarian
Every trade teaches
On position close, Librarian reconciles outcome vs. decision features: which strategy called it, which filter flagged it, what time-of-day edge emerged. Feeds priors back into each trader's knowledge base — the squad gets smarter overnight, not every quarter.
Thompson bandit Auto-learning Async
⟜
The event bus — one spine, no god object
Every stage publishes typed events — CANDIDATE_TOKEN, GHOST_PASS, TRADE_SIGNAL, RISK_VETOED, POSITION_OPENED, POSITION_CLOSED — and every other stage subscribes to what it cares about. No agent calls another directly. Swap any stage without touching the rest.

Why this shape

✕ Monolithic LLM bot
  • One prompt, one vibe, one catastrophic failure mode
  • Latency unpredictable (LLM calls in hot path)
  • Hallucinations reach the wallet
  • Retraining = re-running the whole thing
  • No way to isolate blame when something breaks
✓ Kabal multi-agent pipeline
  • Each stage has one job, typed inputs, typed outputs
  • Sub-second deterministic path, no LLM at decision time
  • Failures are localized, not systemic
  • Librarian learns online per-stage, not per-rebuild
  • Memo trail makes every decision independently auditable

Where the LLMs actually live

LLMs aren't absent from Kabal — they're quarantined to the right layers.

Vision uses a small model to classify narrative signals. Librarian summarizes post-mortem learnings. The Council (Q3 2026) adds narrative commentary above the pipeline. None of them can move funds, change caps, or override Sentinel. They advise, annotate, teach — they do not trade.