KERNEL ONLINE ·
MARKET INTELLIGENCE

Topological Regime Detection.
Not Statistical Guesswork.

The same H₀ persistent homology that powers Apex17 robotics now fingerprints market regimes. Birth-death barcodes replace moving averages. O(1) recall replaces retraining. <1ms replaces "overnight batch."

DECISION PIPELINE

Observe → Fingerprint → Recall → Policy → Execute → Record

Six deterministic stages from raw market data to live order execution. No human in the loop. No retraining. Every decision anchored to a regime hash.

01

Observe

OHLCV

Ingest price, volume, order book data via Polygon/Alpaca feeds

02

Topology

<1ms

H₀ persistence on price series. Union-Find elder rule. Birth-death pairs.

03

Fingerprint

SHA-256

Top-8 pairs quantized into 16-byte descriptor. O(1) exact-match hash.

04

Recall

O(1)

20-dim cosine similarity match against RegimeMemory ring buffer.

05

Risk Gate

8 limits

Position size, R:R, drawdown, exposure, confidence — all enforced.

06

Execute

LIVE

Alpaca REST API. Real broker. Real capital. Real accountability.

PROVEN NUMBERS

Every Claim Has a Source File

These are not projections. These are measured values from production code.

<1ms
Regime Detection
500-bar window on CPU
44
Labeled Regimes
regime_labels.json
O(1)
Recall Complexity
SHA-256 hash lookup
20
Fingerprint Dims
Spectral + vol + flow + momentum
ENGINE ARCHITECTURE

Market Topology Engine — Deep Dive

Pure Python chain-graph H₀ persistence. Same algorithm as Apex17's CUDA kernel, CPU-only since market data windows are tiny (100-1000 bars).

01
Core Algorithm

H₀ Persistent Homology

Sort bars by value (ascending filtration). Activate in order. Merge with temporally adjacent bars via Union-Find with elder rule. Extract birth-death pairs. Compute stability, entropy, regime hash.

02
Identity Layer

Regime Fingerprint Hash

Top-8 persistence pairs quantized into 64 bins, packed into a 16-byte descriptor, then SHA-256 hashed to a 16-char hex string. Result: O(1) exact-match recall — "I've seen this exact market before."

03
Memory System

RegimeMemory Ring Buffer

Stores fingerprints with time-decay weighting. Each regime becomes a 20-dimensional vector (spectral, volatility, flow, momentum). Cosine similarity recalls historical outcomes for identical market structures.

04
Risk Enforcement

TitanRiskAgent — 8 Gates

Every order passes through 8 configurable risk limits before execution. Position size, risk-reward ratio, confidence threshold, max drawdown, max exposure. Regime memory adjusts sizing via multiplier [0.7 — 1.3×].

REGIME INTELLIGENCE

44 Labeled Regimes.
From Real Market Data.

Each regime has a name, win rate, average return, and recommended action. The system doesn't guess — it recognizes.

Neutral Market
19
High-Win Zone
9
Quiet Bull Run
7
Bearish Slide
3
Low-Win Danger
1
# data/regime_labels.json — sample regimes
{ "0": { "name": "Quiet Bull Run", "win_rate": 0.9, "action": "Aggressive" } }
{ "16": { "name": "Quiet Bull Run", "win_rate": 1.0, "action": "Aggressive" } }
{ "38": { "name": "Bearish Slide", "win_rate": 0.4, "action": "Cautious" } }
# High-Win regimes → aggressive. Bearish → cautious. Automatic.
RISK ENFORCEMENT

8 Configurable Risk Gates

Every trade proposal passes through TitanRiskAgent validation before execution. No exceptions. No overrides.

10%
Max Position Size
Per-symbol cap
90%
Max Exposure
Total portfolio
1.5
Min Risk/Reward
Required R:R ratio
5%
Max Drawdown
Daily hard stop
200
Max Daily Trades
Churn prevention
0.65
Min Confidence
Model threshold
0.7×
Bad Regime
Tighten sizing
1.3×
Good Regime
Loosen sizing

Source: src/neural_chat/titan_risk_agent.py · RiskLimits dataclass · All values configurable per-strategy

ARCHITECTURAL COMPARISON

Two Fundamentally Different Approaches

Traditional quantitative stacks retrain models on regime shifts. SignalBrain-OS fingerprints the shift itself.

SIGNALBRAIN-OS

Topological Intelligence

market_topology.py regime_memory.py titan_risk_agent.py alpaca_broker.py

Structural identity — birth-death barcodes capture the shape of market structure, not statistical moments. No retraining on regime change.

Key advantage: When markets shift, the topology changes before the statistics do. We detect transitions, not aftershocks.
TRADITIONAL QUANT STACK

Statistical Approach

Feature Eng. ML Model Backtest Paper Trade

Descriptive statistics — moving averages, Bollinger bands, RSI. All lag the market. Models trained on past regimes fail on new ones.

Structural weakness: When the regime shifts, the model's training distribution no longer matches reality. Requires manual retraining.
"If you can fingerprint a point cloud, you can fingerprint a price series. The math doesn't care if the input is LiDAR or OHLCV — topological identity is domain-agnostic."
— SIGNALBRAIN-OS ARCHITECTURE THESIS
WHY TOPOLOGY WINS

Structural Differentiators

Three properties that statistical approaches fundamentally cannot replicate.

01

Transition Detection

Topology detects regime transitions before statistics do. When the persistence barcode changes shape, the market is shifting — even if moving averages haven't crossed yet.

02

Zero Retraining

Statistical models must be retrained when market structure changes. The topology engine adapts instantly — it computes structure from raw data on every tick. No training set, no overfitting.

03

Cross-Domain Transfer

The same compute_market_topology() function runs on price series, volatility curves, and flow data. Robotics, healthcare, defense — same algorithm, different sensor input.

HONEST ASSESSMENT

What We Can and Can't Prove — Yet

Transparency is a moat. Every other company only talks about wins.

✓ What We Can Prove

  • 44 distinct regimes labeled from real market data
  • <1ms topology computation on 500-bar windows
  • O(1) hash recall via SHA-256 fingerprinting
  • 20-dim fingerprint with time-decay cosine similarity
  • 8 risk gates enforced on every trade proposal
  • Live broker execution via Alpaca REST API
  • Regime-adjusted sizing with [0.7×, 1.3×] multiplier

✗ What We Can't Prove Yet

  • Long-term PnL — system is in live validation, not enough history for Sharpe ratio claims
  • Multi-instrument performance — currently validated on equities, not yet on crypto/forex/commodities
  • Regime transition prediction — we detect transitions, we don't predict them
  • High-frequency alpha — our edge is regime awareness, not latency arbitrage
REPRODUCIBLE

Verify It Yourself

Every number traces to a source file. No black boxes.

market topology verification
# Compute topology on live price data
from neural_chat.market_topology import compute_market_topology

result = compute_market_topology(close_prices, volumes=volumes)

# Output:
  stability:       0.847          ← regime holding strong
  entropy:         1.23 bits      ← moderate complexity
  max_persistence: 0.0312         ← dominant feature
  num_components:  7
  regime_hash:     "0x4A7F2C1D8E3B" ← O(1) recall key

# Source files:
  src/neural_chat/market_topology.py  318 lines   H₀ persistence
  src/neural_chat/regime_memory.py    671 lines   Fingerprint store
  src/neural_chat/titan_risk_agent.py 287 lines   Risk validation
  data/regime_labels.json             44 regimes  Labeled behaviors

See the full picture.

28-slide deep dive into architecture, market, team, and financials.

Request Investor Deck → View GPU Performance → Try Live Demo →