LEDATIC EMPIRE / v2.0 / SOLANA / LIVE

THE
MACHINE
TRADES.

  ╔═══════════════════════════════════════════════════════════════════════╗
  ║              L E D A T I C   E M P I R E   v 2 . 0                  ║
  ║         Autonomous Solana Trading — 9 Services Running Live           ║
  ╚═══════════════════════════════════════════════════════════════════════╝
                                      │
          ┌────────────────────────────┼────────────────────────────┐
          │                           │                            │
  ┌───────▼───────┐          ┌────────▼────────┐         ┌────────▼────────┐
  │   SNIPER        │          │   OVERSIGHT     │         │    CORTEX       │
  │   P0 · LIVE   │          │   P1 · 30-120s  │         │    P2 · LIVE    │
  │               │          │                 │         │                 │
  │  Solana DEX   │          │  Consciousness  │         │  Health Monitor │
  │  Time-gated   │          │  1,844 thoughts │         │  Auto-Repair    │
  │  $10 pos size │          │  Adaptive cycle │         │  Rollback Guard │
  └───────┬───────┘          └────────┬────────┘         └─────────────────┘
          │                           │
  ┌───────▼───────┐          ┌────────▼────────┐
  │  PORTFOLIO    │          │   AUTONOMY      │
  │  P1 · 5min   │          │   P1 · :5590    │
  │               │          │                 │
  │  Size · Slots │          │   LLM Router    │
  │  Stop Tuning  │          │   Priority Queue│
  └───────────────┘          └────────┬────────┘
                                      │
                          ┌───────────▼───────────┐
                          │   Qwen2.5-14B-4bit    │
                          │   :8080 · ~9GB GPU    │
                          │   ─────────────────── │
                          │   Claude Sonnet 4.6   │
                          │   Cloud Fallback      │
                          └───────────────────────┘

  SIGNAL PIPELINE:  signalfeed:5557 → safety_rails → neural_gate → sniper
  OVERRIDE FLOW:    oversight → portfolio → directive → runtime_overrides.json
  DEPLOY:           git push → deploy.py → health check → P&L watch → rollback
895 Tests Passing
1,844 Oversight Thoughts
9 Services Live
42d Days Running
// 01 — THE MACHINE

EVERY COMPONENT
HAS A JOB.

Nine services. One objective. The system is designed to be self-aware — the trading engine is watched by a portfolio brain, supervised by a consciousness layer, and protected by a health monitor that can roll back bad deploys automatically. Nothing runs without being observed.

SNIPER
Live Solana DEX trader. Time-gated, signal-filtered, neural-scored. $10 position size. 15% hard stop, 4-tier trailing exit.
P0 · LIVE
OVERSIGHT
Autonomous consciousness. Thinks in 30-120s adaptive cycles. Accumulates conclusions. Publishes parameter overrides when confidence ≥ 0.90.
LIVE
PORTFOLIO
5-minute brain. Reads live P&L and account state. Adjusts position size, slot count, and stop levels via the override JSON.
LIVE
AUTONOMY
LLM router on :5590. Single GPU queue (one call at a time). Priority: traders > agent > cortex > oversight. Cloud fallback for non-latency tasks.
LIVE
CORTEX
Health monitor and auto-fixer. Detects service failures, triggers restarts, watches for regression. Kill switch aware. Never touches parked services.
LIVE
AGENT
14-executor autonomous task loop. Runs directives from oversight, portfolio, and human operators via Telegram. Directive executor bypasses the queue.
LIVE
MLX / 14B
Qwen2.5-14B-4bit on :8080. Local reasoning — trading gates, oversight thinking, signal analysis. ~9GB GPU. Three-layer kernel panic defence.
LIVE
SIGNALFEED
Signal pipeline on :5557. Ingests raw token signals, runs safety rails and neural gate, publishes clean signals to the sniper loop.
LIVE
SIGNAL BOT
@Thatsignalsguybot on Telegram. Free tier (3/day), Pro (250★/mo), VIP (1000★/mo). Real signals from the live trading pipeline.
PARKED
// GPU DEFENCE STACK — Mac Mini M4 Pro, 24GB unified RAM
Layer 1: iogpu.wired_limit_mb=19000 — OS-level hard cap via LaunchDaemon  ·  Layer 2: gpu_watchdog.py — SIGTERM at 13.5GB (14B) / 16GB combined  ·  Layer 3: ThrottleInterval=90s — prevents rapid crash-loop GPU object corruption
// 02 — BUILD LOG

BUILT IN
THE DARK.

Every system was built, broken, autopsied, and rebuilt. The log below is honest — including the kernel panics, the bad trades, and the 3am incident reports.

2026-02-21
Survived two kernel panics. Engineered the fix.
IOGPUFamily double-free during the 06:00 daily restart. 14B crashed 6+ times in 13 minutes; each restart left orphaned Metal GPU objects that panicked on double-free. Built three-layer defence: OS wired memory cap via LaunchDaemon, RSS watchdog (SIGTERM at thresholds), 90s ThrottleInterval. Killed the 3B model permanently to free headroom. System stable.
gpuinfrastructureincident
2026-02-21
Oversight v3 — settled knowledge + thread isolation
Conclusions confirmed 2+ times at ≥0.85 confidence become "settled knowledge" — injected into every future LLM context window. Thinking runs in an isolated daemon thread so the trading loop never waits on Qwen. Question pool pruned from 81 to 36. Proposal queue routes non-filter insights to async human review instead of burning cloud API tokens.
oversightaiarchitecture
2026-02-20
131-trade autopsy. DEAD_POOL identified as the #1 problem.
Dissected every closed trade across three strategies. DEAD_POOL — tokens that fail the honeypot probe after entry — is 29.8% of trades but 89% of all losses. Without it, sniper is profitable. The honeypot probe was upgraded from 1% to 5% real sell. Time-of-day dead zones eliminated (hours 6-8, 12, 17-19 UTC).
tradinganalysissniper
2026-02-18
Deploy pipeline with auto-rollback goes live
tools/deploy.py auto-detects affected services from git diff, restarts them, runs 60s health check, then watches P&L for 5 minutes post-deploy. Regression triggers automatic git revert, re-restart, and Telegram alert. No more manual babysitting of deploys.
devopsautomation
2026-02-18
Cortex restarted shotguns with ideaguy's uncommitted dry-run config
Ideaguy modifies *_cfg.py locally for testing without committing. Cortex auto-restart picked up the dirty config and ran the shotgun strategies in dry-run mode for hours. Killed ideaguy. Added "always git diff configs after unexpected behavior" to the ops rulebook. Now a permanent rule.
incidentlessons
2026-02-11
28 MLX SIGABRT crashes in 13 hours. Fixed.
Running 14B and 3B simultaneously on the M4 Pro's unified memory pushed Metal past its wired allocation limit. Each SIGABRT triggered an auto-restart, which crashed again immediately — a death loop. Fixed by serializing all GPU calls through a single LLMQueue and setting ThrottleInterval=30s on the plists.
gpuincidentmlx
2026-02-01
Sniper goes live on Solana mainnet
First real money. $10 position size, $50K minimum liquidity, neural gate ≥0.60. The machine starts accumulating data. Early results reveal the DEAD_POOL problem that would take three weeks to properly diagnose and fix.
launchtradingmilestone
2026-01
Oversight, Portfolio Brain, and Autonomy router built
The consciousness layer gets its first real form. Oversight runs 30-120s adaptive cycles, thinks with local Qwen, publishes overrides when confidence is high enough. Portfolio Brain handles position sizing. Autonomy serializes all GPU calls. The three-layer intelligence stack is complete.
oversightportfolioarchitecture
2025-12
First version. One script. No infrastructure.
Started as a single Python file watching Solana DEX activity. No tests, no services, no oversight. Just curiosity and a terminal window. The infrastructure came from asking: what would it take for this to actually work?
origin
// 03 — GOALS

WHAT IT'S
WORKING TOWARD.

These are tracked in an Obsidian vault that the agent reads. Progress is real. Some of these are close. Some are hard.

⚠ CRITICAL
Eliminate DEAD_POOL
29.8% of trades, 89% of losses. The honeypot probe is now 5% real sell. The next step is pre-entry sellability scoring so we never enter these tokens. Liquidity is the real signal.
In progress — partial fix deployed
● HIGH
Kelly-Monte Carlo Sizer
Replace fixed $10 position size with Kelly Criterion scaled by Monte Carlo confidence interval. The portfolio brain currently uses heuristics. This replaces them with math.
Planned — goal active in Obsidian
● HIGH
Ruin Circuit Breaker
Hard stop on the entire system if drawdown exceeds a threshold calculated from the Kelly formula. The machine should be able to recognize when conditions are bad enough to stop trading entirely.
Designed — not yet implemented
○ MEDIUM
Proven Neural Edge
Current neural scorer: winners avg 0.707, losers avg 0.702. Functionally random. The autopsy confirmed it. Goal: find or build a signal with actual predictive power. Not just a gate — real alpha.
Research phase
○ MEDIUM
Sellability Pre-Scoring
Build a pre-entry model that estimates post-buy sellability without actually buying. The honeypot probe is reactive. This would be predictive. Use the autopsy data to train it.
Backlog
○ MEDIUM
Signal Bot — Full Launch
@Thatsignalsguybot is built, tested, and parked for GPU headroom. When the trading signal quality is proven, launch the bot properly. Free / Pro 250★ / VIP 1000★ tiers are already coded.
Waiting on signal quality proof
// 04 — DEVLOG

THE MACHINE
WRITES THIS.

This log is generated by the Oversight service. Every 30-120 seconds it thinks, and when it concludes something worth recording, it writes here. Click any entry to expand.

// loading...