Diligence.

01Control philosophy

STRATA operates on the premise that AI is fallible and markets are adversarial. Every control is designed to fail safe — meaning when a component degrades or disagrees, the system stops trading rather than continues trading.

Three modes govern every deterministic gate:

• OFF — the gate does not evaluate. Used only for development or controlled testing.
• ADVISORY — the gate evaluates and logs its decision, but does not block execution. Used to measure new rules against live behavior before enforcement.
• ENFORCING — the gate actively blocks, reduces, or modifies proposals that fail its evaluation.

02Risk matrix

Every material risk vector the platform faces, with the active mitigation. Reviewed monthly.

03Defense in depth

Every order passes through at least four independent checkpoints before reaching an exchange. A failure at any checkpoint blocks the order.

1. Primordial anchoring     AI has internalized invariant
2. Cycle payload reminder    cycleLoopAwareness field
3. Safety net                code checks AI output structure
4. Review unit gate          ERU / WVU / IB Equity Review
5. Portfolio Manager         posture + exposure + redundancy
6. Entry State Machine       chase / extension / breakout
7. Risk Brake                session / shock / giveback
8. Material Change Controller justification requirement
9. Exposure Graph            book-level what-if
10. Price Sentinel           60s independent safety loop

Additionally, the Post-Trade Forensics layer runs after every close and labels the trade along six failure modes: thesis failure, expression failure, entry timing, stop placement, profit harvesting, premature exit. Labels feed back to Analyst and trading units.

04Failure modes

We reject the framing that AI trading systems are black boxes. Every material failure mode is documented, instrumented, and has a defined response. The table below covers the categories we actively monitor.

05Operational posture

Four traffic states govern the overall aggressiveness of the system at any given moment. The Analyst sets the state; the Portfolio Manager enforces it.

Performance state layers on top: HEALTHY, CAUTIOUS, DEFENSIVE based on recent PnL trajectory. Performance degradation shortens cycle intervals (more monitoring) and biases Portfolio Manager toward smaller sizes.

06Observability & audit trail

Every material decision in the system is logged with enough context to reconstruct it later. This is not optional — the forensic label feedback loop requires it.

• Cycle logs capture the complete payload, REI response, parsing outcome, review verdict, and governance pipeline result.
• Prior verdicts (last 5 per symbol) persisted in Redis, injected into every cycle for continuity.
• Previous cycle context captures structured data about the last decision — what, why, timestamp.
• Sentinel actions logged and injected into the next cycle so the AI knows what happened between cycles.
• Post-trade forensics write a labeled record for every closed trade, timestamped and searchable.
• Teaching interactions (primordials, corrections, research teaching) tracked in Redis with TTL-based dedup.
• Governance decisions at every layer (PM, Entry SM, Risk Brake, Sentinel) are logged with the exact reason for approve/reduce/reject.

07Kill switches and manual controls

Operators can intervene at multiple levels without restarting the system.

• Global kill switch. A single flag (tradingPaused) stops all new execution across all units. Existing positions continue to be monitored by Price Sentinel.
• Per-unit disable. Individual trading units can be disabled via config. Their cycles suspend; review units and Analyst continue.
• Governance mode overrides. Each of the 10 governance layers can be independently flipped between OFF, ADVISORY, ENFORCING.
• Manual corrections. Operator can send an explicit correction to any unit at any time via API.
• Operator overrides. Per-symbol override instructions inject directly into the cycle payload (e.g., "do not short X", "cap size at Y", "close Z on any weakness").
• Manual cycle trigger. Every unit has a manual trigger endpoint for ad-hoc review outside the scheduled cadence.

08Data integrity and sources

STRATA runs on commercial-grade data sources. No web scraping for anything that touches trade decisions.

Raw payloads (especially Laevitas) are not sent to AI units directly. The data services layer processes, normalizes, and summarizes data first. This preserves REI's native numerical processing and keeps token usage predictable.

09Deployment and infrastructure

10What we do not claim

Honest posture on the limits of the system:

• We do not claim AI understands markets. REI Core processes patterns across research, news, regime, and price data. It does not have a theory of markets. The governance layer assumes AI reasoning can be wrong.
• We do not claim backtests predict live performance. Every new governance rule ships in ADVISORY before ENFORCING. Production behavior informs calibration.
• We do not claim HFT capability. STRATA operates on minutes-to-hours cadence. We do not compete with colocated tick-level systems.
• We do not claim zero drawdown. Drawdowns are bounded by posture gates, profit ratchet, and session risk brake. They are not eliminated.
• We do not claim the system is complete. Forensic labels and explicit corrections continuously reshape behavior. This is by design, but it means the system of today is not identical to the system of a month from now.

Complete technical architecture detail is in architecture. Full unit inventory and teaching-endpoint reference in units. For a walkthrough of how a single cycle executes, see how it works.

Questions beyond the scope of this documentation should be directed to founder@ecliptica.ai.