// ARCHITECTURE LIVE

A closed-loop
intelligence
pipeline.

Forty-one collectors feed eighty-plus processors. Five ML models score what survives. A bot arena evolves the strategies. Eight protection layers guard the capital. Built for one operator, running on one VPS, executing in ~82 seconds per cycle.

41
Collectors · 6 Modules
80+
Processors
5
ML Models
22+
Containers
View Track Record Inspect The Stack
01 / IN SIXTY SECONDS

Three verbs. One loop.

Strip the architecture diagrams away and the whole thing reduces to three actions repeated every hour. Collect the signal. Fuse the signal into an edge. Execute the edge into a position. Then learn what worked.

// 01 · INGEST
Collect.
Forty-one async collectors fire in parallel every hour across six modules — newswire, geopolitical risk, financial flow, cyber + critical-infrastructure threat, sport, and cross-correlation. Source-level identification is reserved for Tier-2 subscribers.
  • ~46s  wall-clock collection window
  • 6 modules  RTM · GR · FFT · CT · SP · CC
  • checkpoint + dedup  at the base class
  • 3000-item  working memory, 72h eviction
// 02 · FUSE
Score.
Convergence engine clusters entity-pair co-occurrences across sources. Signal classifier tiers them T1/T2/T3. Anomaly detector (Welford online) flags statistical breaks. ML models score structural risk and event acceleration. Edge scorer prices each against the live prediction-market book.
  • ~65 clusters  per cycle after quality filter
  • 4-way divergence  Polymarket · Kalshi · Manifold · Metaculus
  • Welford O(1)  online anomaly state
  • SQ scoring  domain-weighted source quality 1-4
// 03 · ACT
Execute.
Portfolio allocator stacks eight intelligence multipliers onto Kelly. Risk manager runs the HEAT system. Watcher daemon polls oracle wallets every 10-45s and copies qualifying E12 entries. Strategy engine routes through five entry models. TP/SL/trailing stops close the loop.
  • conservative  Kelly fraction by default
  • 3 modes  DRY_RUN → PAPER → LIVE
  • HEAT system  WARM · HOT · MELTDOWN
  • emergency cutoff  hard daily floor
02 / STACK INVENTORY

Everything inside the box.

Six modules of intake. Multi-method processing. ML ensemble scoring. Eight-layer capital preservation. Capability counts public; specific source / processor / model / infra inventories are reserved for Tier-2 subscribers and above.

LayerRoleScale
CollectionOSINT + financial + geospatial intake across 6 modules (RTM · GR · FFT · CT · SP · CC)41 sources
ProcessingConvergence engine + signal classifier + multi-method deduplication + enrichment80+ processors
ModelingProprietary ML ensemble — structural risk, event acceleration, regime classification5+ model classes
AllocationMulti-factor portfolio sizing with adaptive regime weightingConservative Kelly
ExecutionRisk-gated trade routing with three operating modes (DRY_RUN · PAPER · LIVE)8-layer guard
StorageHot working memory · warm rolling archive · cold analytics3-tier · queryable
InfrastructureContainer fleet on single VPS · monitoring · self-healing22+ containers · 84 cron
RhythmsHourly pipeline · 10–45s wallet polling · 30-min self-heal · 100-cycle walk-forward~82s pipeline
03 / THE LOOP

Six stages. One organism.

The pipeline is not a DAG. It is a heartbeat. Every cycle starts at COLLECT and ends at LEARN — the learn step writes back into convergence weights, exit genes, and Kelly priors before the next collect fires.

// 01 Collect Parallel ingest of 41 sources. Async, checkpointed, deduped at the base class. ~46s · 6 modules
// 02 Converge Entity-pair clusters across sources. Trigram dedup. Dempster-Shafer fusion. Geo-enrich. ~65 clusters / cycle
// 03 Score Tier T1/T2/T3, source quality, anomaly z-score, structural risk, event acceleration. 5 ML models
// 04 Size Multi-factor allocator. Conservative Kelly fraction modulated by regime, HEAT zone, and wallet-tier consensus. quarter-Kelly cap
// 05 Execute Watcher copies oracle entries. Five strategy engines. TP/SL/trailing exits. 10-45s polling
// 06 Learn Trade audit, drift detection, walk-forward validation, evolutionary strategy improvement in the arena. shadow population
04 / PROTECTION

Eight layers between signal and ruin.

Every entry passes through eight independent guards before it touches capital, and a ninth — the trade auditor — inspects every exit. No single point of failure has authority to size up a position.

// 01 Emergency Cutoff ARMED Hard daily floor on realised PnL. Trading halts until manual reset.
// 02 Drawdown Ceiling ARMED Bankroll high-water-mark drawdown guard. Trips at threshold, forces re-evaluation.
// 03 Consecutive-Loss Pause ARMED Streak of consecutive losses pauses new entries for one full pipeline cycle.
// 04 Daily Exposure Cap ARMED Total capital deployed per UTC day, summed across all strategies.
// 05 Per-Position Cap ARMED No single position can exceed the configured notional, regardless of Kelly suggestion.
// 06 Edge Decay Monitor ARMED Real-time edge-decay detection. Auto-pauses a strategy when its edge degrades.
// 07 Drawdown HEAT 3-tier WARM · HOT · MELTDOWN escalating drawdown throttles.
// 08 Post-Trade Audit ARMED Slippage, execution quality, and outcome attribution captured for every exit.
05 / STORAGE TIER

Three layers of memory.

Hot working memory, warm rolling archive, cold analytics. No vendor lock — every byte lives on the same VPS, indexed and queryable from one process.

// 01 · Working JSON · 72h evict

Flat JSON files. The pipeline's working set — signals, divergences, edges, briefing drafts. 3000-item cap with sliding 72-hour eviction.

signals.jsonlive
smart_signals.json2h re-gen
whale_consensus.jsonlive
convergence_signals.jsoncycle
// 02 · Archive JSONL · 180 day

Daily JSONL archives under data/archive/YYYY-MM-DD/. Append-only, line-delimited, replayable. 180-day rolling retention.

formatJSONL
retention180 d
replaynative
backupdaily ×7 · weekly ×4
// 03 · Analytics SQLite WAL · 9 tables

Embedded SQLite in WAL mode, indexed for backtest and review. Multi-table schema covering signals, edges, divergences, and trade history. Schema and row-level access available to Tier-2 subscribers.

engineSQLite WAL
tablesmulti-schema
indexedyes
queryablein-process
06 / TRANSPARENCY

What we do not ship.

Every page on this domain — including this one — is auditable cold. No third-party fonts. No analytics pixel. No external CDN. ML inference runs in-process, not over a vendor API. The threat model assumes adversaries inspect the network tab.

[ TLS ]
HTTPS + HSTS
TLS 1.3 only. Strict-Transport-Security preloaded. No mixed content. No HTTP fallback.
[ ZERO CDN ]
Self-Hosted Fonts
woff2 served from origin. No googleapis. No gstatic. No unpkg. No jsdelivr. No fingerprint leak.
[ IN-PROC ]
Local ML Inference
Every ML model runs inside the pipeline container. No external inference API, no model weights phoned home.
[ NO TRACK ]
Zero Analytics
No GA. No Plausible. No Mixpanel. No tracking pixel. No third-party JS. Inspect the network tab.
// END · TRANSMISSION

The pipeline is already running.

Read the track record, query the public API, or watch the live signal feed. All three are produced by the same loop documented above. The numbers are what they are.

Track Record Public API Live Signals