Observe broadly
Alpaca assets, snapshots, daily history and completed intraday bars create a time-bounded view of tradable US stocks and ETFs.
Saentinel Trader · Active paper evaluation
A focused intraday trading research system for US stocks and ETFs. Market data ranks the candidates. Transparent deterministic rules select entries. Risk and execution remain deterministic. An optional shadow AI observes the same evidence only for comparison and research.
A clean restart
Saentinel Trader replaces an overbuilt investment-research platform with a deliberately small operating loop. It scans liquid US equities, records the evidence behind every cycle, selects entries with transparent deterministic rules, and keeps position sizing, risk, execution, monitoring and exits deterministic. An optional shadow model evaluates the same evidence independently so its recommendations can be measured without influencing orders.
One bounded cycle
Each cycle is claimed once and persisted. Selection is deterministic, stale results cannot enter, and execution rechecks the world before acting. Optional shadow-AI work is recorded independently and never delays entries.
Filter and rank eligible US stocks and ETFs from current market observations.
Add completed daily and intraday context, with explicit rejection of stale or incomplete evidence.
Apply deterministic VWAP, ATR-extension, correlation, occupancy and cycle-cap rules to the ranked shortlist.
Apply deterministic market-regime, capacity, capital, risk, spread, timing and asset-eligibility rules. Revalidate before submission; optional paper pullback entries wait at a fixed limit within a bounded window.
Manage owned positions through stop, take-profit, thesis, stagnation, winner-to-loser profit protection, daily-risk and end-of-day rules.
Persist scans, selections, orders, fills, ownership and research outcomes in SQLite. When enabled, shadow-AI recommendations are stored separately for head-to-head evaluation.
Risk before activity
The trading path is explicit, testable and restart-safe. Optional AI is observational: it can be compared with the deterministic selector, but it has no authority over capital, orders, risk or exits.
The selector assigns a bounded allocation ceiling. Position, capital, risk-mode, buying-power and market-regime controls may reduce it further; none can silently enlarge exposure.
Recovery mode reduces size, deterministic SPY regime policy can restrict or block new entries, and a hard daily-loss stop blocks new entries and flattens owned exposure.
Saentinel monitors and exits only quantities it can prove it owns. Manual positions remain outside its authority.
New entries stop before the close. Entry orders are canceled and owned positions are flattened within a bounded liquidation window.
Current paper research
Forward paper evaluation now asks two practical questions: can waiting for a pullback improve entries, and does a tightly controlled second attempt after a winner help the whole account? Each change must be judged against fills, missed opportunities, drawdown and session outcomes. Profitability remains unproven.
An optional normal-scan entry mode places a fixed limit below the decision price and waits within a bounded window. Orders are not chased or enlarged; unfilled capital and risk remain reserved until cancellation is confirmed. The separate 09:40 path stays immediate.
An optional paper-only rule permits one additional pullback attempt after a proven profitable trade, confirmed flatness and a cooldown. Fresh selection and all existing risk controls still apply. A failed or unfilled attempt cannot be recycled.
Two separate paper accounts will compare winner re-entry off versus on with otherwise matching code and settings. The planned evaluation covers at least ten matched full sessions and measures account returns, drawdown, fills and displaced opportunities—not just the second trade's result.
An aggregate planned open-risk cap can limit new entries across owned positions and pending buys. This supplements per-trade and daily-loss controls; planned risk is not a guarantee of realized loss.
Paper sessions can finish with a checked daily journal archive after the official close, confirmed flatness and resolved orders. A submitted exit or cancellation never counts as proof that the account is flat.
Deterministic rules continue to select entries. Optional AI recommendations are recorded independently for comparison and cannot place, size, block or exit trades.
Saved market observations, decisions, orders and fills support review of entry quality, feed differences and trade economics. Missing observations remain unknown; favorable selected trades alone do not establish an edge.
What failure changed
“A system can execute correctly and still fail its investment objective.”
The original Saentinel became too complex to prove useful. The replacement deliberately narrows the problem, preserves real paper evidence, and treats disappointing outcomes as inputs—not as results to explain away. When forward evidence showed that LLM picks underperformed the candidates they rejected, the model was removed from the trading decision path instead of being defended. Profitability remains unproven.
The honest boundary
Follow the experiment
Saentinel Trader is private software under active paper evaluation. Questions about the project are welcome.
contact@saentinel.com