AI never owns execution
Models structure information and draft policy. Deterministic code owns the final capital decision.
NOCTURNERRESEARCH / BUILD NOTES
Trust is earned
in the calibration.
The difficult part is not producing a score. It is proving that the score behaves sensibly across sessions, assets, order sizes, and market regimes.
CORE THESIS
In 24/7 markets tied to part-time price discovery, data age is a position.
A token price can look precise while the information beneath it is decaying. Nocturner treats freshness, depth, divergence, and catalysts as first-class execution inputs.
The engine is built to say what it knows, show why it reached a verdict, and become more conservative when evidence quality deteriorates.
DESIGN PRINCIPLES
Constraints we
refuse to blur.
Every verdict explains itself
A score without factors cannot be challenged, debugged, or responsibly automated.
Utilities share core data
If a feature needs an unrelated pipeline, it is a separate product—not a free addition.
Policy is configurable
Risk tolerance belongs to the capital owner. Nocturner supplies factors, defaults, and safe boundaries.
CALIBRATION LOOP
Observe first.
Enforce when proven.
01
Collect
Historical pool, oracle, session, and catalyst state.
02
Backtest
Weights against realized slippage and post-trade outcomes.
03
Review
False positives, missed events, and regime behavior.
04
Promote
Approved thresholds into deterministic policy.
ROADMAP
Built in the order
risk demands.
Read-only scorecard
Historical collection, transparent factors, dashboard, and alert delivery.
IN PROGRESSAgent and protocol APIs
MCP queries, dry runs, corporate actions, halt mirror, and RiskFeed.
NEXTERC-4337 guard
Audited policy module, session-key integration, and production thresholds.
AFTER CALIBRATIONReputation and tuning
Outcome-driven agent reputation, threshold tuning, and protocol analytics.
USAGE-LEDDEPLOYMENT SHAPE
One Go binary. One PostgreSQL. One Redis.
TimescaleDB and pgvector extend the same database. Dedicated RPC, local inference, and Kubernetes wait until actual volume makes them rational.
EARLY ACCESS / 2026
Follow the risk layer as it becomes real.
We're opening the read-only scorecard first while the engine collects the history needed for responsible calibration.