AEGIS · LIVE / MAINNET · TECHNICAL NOTE · 25 AUG 2026
Evidence Boundary for a Spot Trading Engine
DEPLOYMENT: LIVE / VERIFIED
PERFORMANCE: NOT YET PUBLICLY VALIDATED
BACKTEST: SEPARATE EVIDENCE SET
Abstract
AEGIS is deployed on Binance mainnet with real capital. The deployment is supported by production exchange configuration and retained exchange-side order, fill, commission, execution and closure records. Its deterministic path combines five-minute market candles, regime and ATR calculations, and live order-book imbalance. A local LLM may approve or reject a fully formed candidate but cannot alter entry, stop-loss or take-profit values. Live execution is long-only. Historical replay cannot reproduce order flow and therefore forces an OFI direction when evaluating numeric entry/exit logic. Deployment, live performance and backtest are treated as separate evidence sets.
1 · Architecture OBSERVED
The current configuration names BTCUSDT, ETHUSDT, SOLUSDT and BNBUSDT; consumes 5-minute candles with a 150-candle lookback; and schedules the decision cycle hourly. The 5-minute timeframe is the data resolution; the hourly scheduler is the current decision-cycle cadence, with each cycle evaluating the latest 150 candles. Signal construction is numeric and synchronous. Regime, ATR and order-flow determine whether a card exists. The live executor rejects non-LONG cards on the spot path.
Risk sizing budgets 2% of available balance at the stop distance. A 5% current-balance daily loss cap trips a persistent circuit breaker. Reconciliation, stale-position checks and closure recording run around the execution cycle.
2 · LLM boundary OBSERVED
The optional local-model layer receives an already locked candidate. It may approve or reject and provide narrative rationale; it is instructed not to change numeric trade parameters. Signal generation remains available when the narrative enrichment layer fails. Contribution has not yet been isolated through controlled ablation. Its presence therefore does not establish causal performance contribution.
3 · Historical simulation BACKTEST
The scheduled replay fetches 5,000 recent 5-minute candles per configured symbol. Historical best-bid/ask depth changes are unavailable, so the real OFI gate cannot be replayed. The engine instead evaluates forced bullish and forced bearish OFI scenarios. These are conditional numeric simulations, not end-to-end replicas of live operation.
- Signals use a 150-candle rolling window.
- Entry is assumed filled immediately at the card entry price.
- If one OHLC candle touches SL and TP1, stop-loss is assumed first.
- Only TP1 is used.
- No overlapping positions are simulated.
- Totals are additive per-trade percentages, not compounded equity.
- Bearish results are numeric-only and cannot be executed by the long-only live spot path.
4 · Live deployment LIVE / MAINNET
AEGIS has executed real-capital Binance mainnet orders. Retained records include exchange order identifiers, realized entry fills, commissions, partial-fill-aware quantities, protected exits and closure journals. This verifies deployment and operational execution. It does not establish profitability or equivalence with backtest results.
Current journal P&L is not published as a strategy-performance result: entry commissions are recorded from fills, while some exit commissions remain estimated rather than independently retrieved from exchange trade history, and not every closure has a fully comparable fill path. Account-balance growth is not used as a substitute because deposits and withdrawals would contaminate it.
5 · Limitations
- Backtest/live discrepancy: historical OFI is absent and replaced with a forced direction.
- Fees, spread and slippage are not included in the replay engine.
- Instant entry fill is an execution assumption.
- OHLC data cannot reveal intrabar touch order; the conservative stop-first rule is used.
- Results depend on market regime, the recent-window dataset and the fixed symbol universe.
- The live executor is long-only; bearish backtests are not live-executable.
- Additive P&L and drawdown do not model account compounding.
- Sample size, parameter sensitivity, overfitting and live/backtest discrepancy remain unvalidated without a versioned result artifact.
- LLM and order-flow contribution have not been isolated by controlled ablation.
6 · Reproducibility / Evidence
| Item | Audited value |
|---|---|
| Market | Binance Spot |
| Symbols | BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT |
| Timeframe / history | 5m / 5,000 recent candles |
| Window | 150 candles |
| Exit model | SL or TP1; stop first on same-candle ambiguity |
| Costs | Fees, spread and slippage not modeled |
| Result status | NOT YET VALIDATED for public performance reporting |
Source boundary: `ordis_prime/core/aegis_engine/` as audited on 2026-08-25. Private code, credentials and operational account data are intentionally not published.