FREEDOM FUND / TRACK B
Build a trading system
that can survive.
Strategy Map research for Track B. Experiments stay paper until Mason promotes them. The real Agentic sleeve is the Track A live lab on /ff. Agents can research and propose. Deterministic code and a human decide what is allowed.
Research snapshot · September 11, 2026 · evidence first
TRACK B · PAPERStrategy Map experiments until promoted. Track A LIVE lab: /ff
01 / GOAL MATH
Run the math
Change the inputs to make the constraint visible. This calculator assumes no withdrawals and smooth daily compounding; real returns are lumpy.
Days remaining345
Daily pace2.50%
Weekly pace18.86%
30-day pace109.72%
Current capital0.02% of target
A smooth compounding path is a mathematical baseline only. It does not imply that this return is achievable or that taking more risk improves the odds.
02 / OPERATING THESIS
Use agents where judgment is expensive.
- Research agentFinds and cites point-in-time evidence across filings, prices, macro data, and news.
- Quant layerTurns evidence into reproducible features, signals, and portfolio candidates.
- Risk criticChecks concentration, liquidity, drawdown, regime change, and whether the thesis is falsifiable.
- Execution layerUses deterministic sizing, limit checks, cost estimates, and a kill switch.
- Human gateRequired before any promote. Track B stays paper until Mason says yes.
03 / NON-NEGOTIABLES
The system earns autonomy.
- Strategy Map stays paperTrack B has no live order path until Mason promotes it. Human gate required.
- Risk is code, not proseSize, exposure, loss, leverage, and liquidity limits must be enforced outside the LLM.
- Point-in-time dataNo look-ahead, survivorship, revised data, or untracked prompt changes.
- Cost-adjusted evidenceInclude spread, fees, slippage, funding, borrow, and taxes where relevant.
- Stop conditionsPause on drawdown, drift, stale data, tool failure, or unexplained behavior.
04 / RESEARCH FINDINGS
What the evidence supports
Our working conclusion: use AI to improve the research loop and operational discipline before asking it to control risk capital.
Open
04 / RESEARCH FINDINGS
What the evidence supports
Our working conclusion: use AI to improve the research loop and operational discipline before asking it to control risk capital.
AI is strongest around information.
Retrieval, source comparison, structured extraction, research synthesis, and exception handling are more defensible starting points than unconstrained price prediction. FINRA describes these use cases ↗
Smart models still lose to simple baselines.
Recent sequential trading benchmarks find that many agents struggle to beat buy-and-hold and that general intelligence does not automatically produce robust returns. AI-Trader ↗ StockBench ↗
Backtest discipline is a competitive edge.
Every new prompt, feature, universe, parameter, and model is another trial. The ledger must count them and apply selection-bias checks such as the Deflated Sharpe Ratio ↗.
Time matters as much as price.
Filings and macro observations can arrive late or be revised. Use EDGAR submission timestamps and FRED/ALFRED real-time periods so historical experiments only expose what was knowable at the decision time. SEC ↗ FRED ↗
Costs can erase the signal.
Fees, spread, slippage, funding, borrow, latency, and partial fills belong in the model from day one. An agent should know when not to trade because expected edge is smaller than friction.
Autonomy has to be earned.
Govern, map, measure, and manage the system continuously; define human oversight and a way to deactivate it when outcomes drift. NIST AI RMF ↗
05 / MODEL BLUEPRINT
Separate the agent from the portfolio engine.
The LLM proposes, explains, and critiques. Typed data contracts and deterministic code decide what is executable.
Open
05 / MODEL BLUEPRINT
Separate the agent from the portfolio engine.
The LLM proposes, explains, and critiques. Typed data contracts and deterministic code decide what is executable.
01Canonical dataPrices, filings, macro, events, vintages
02Feature codeReproducible transforms and signals
03Agent reviewResearch, thesis, counter-evidence
04Risk engineExposure, liquidity, cost, loss limits
05Paper executorFills, reconciliation, audit trail
EVALUATION GATES
A candidate can advance only when it has:
- A locked hypothesis and declared data cutoff
- Chronological out-of-sample results
- Realistic fees, spread, slippage, and liquidity
- Walk-forward and regime-slice performance
- Baseline comparison and complete trial count
LEDGER FIELDS
Every decision should retain:
- Dataset and data-vintage identifiers
- Model, prompt, tool, and retrieval versions
- Signal, score, position, and risk decision
- Expected vs. realized execution and cost
- Outcome, drift flag, and post-mortem
06 / STRATEGY MAP
Promising research lanes
These are the most defensible ways to use agents around trading today—not promises of alpha. Start with benchmarks and falsifiable experiments.
01BASELINE
Diversified allocation
Keep a boring benchmark and cash policy so every active idea has a fair comparison.
- Agent role
- Research summary and rebalance checklist
- First test
- Buy-and-hold / equal-weight / simple trend baselines
02SIGNAL
Trend & momentum
Use the agent to classify regime and explain signals while code handles entries, exits, and sizing.
- Agent role
- Regime context and evidence extraction
- First test
- Cross-sectional momentum with turnover and cost controls
03SPREAD
Mean reversion / stat arb
Trade relative relationships rather than a single direction, with explicit tests for stationarity and regime breaks.
- Agent role
- Pair discovery, thesis tracking, anomaly review
- First test
- Cointegration / spread z-score with a hard exposure cap
04EVENTS
Event-driven research
Extract structured facts from filings, earnings, and verified news; never let a prose opinion become an order by itself.
- Agent role
- Source verification and catalyst timeline
- First test
- Point-in-time earnings/news features vs. a neutral benchmark
05EXECUTION
Cost-aware execution
Probably the most reliable near-term use: reduce friction, missed fills, bad sizing, and operational mistakes.
- Agent role
- Order plan and exception handling
- First test
- Slippage model, limit-vs-market simulation, and fill quality
06LATER
RL for allocation
Useful only after the data, reward, costs, and constraints are trustworthy; highly vulnerable to unstable training.
- Agent role
- Experiment orchestration and critique
- First test
- Multi-seed walk-forward evaluation against simple policies
07 / RESEARCH PROTOCOL
A ladder, not a leap.
Each stage should produce an artifact that makes the next stage safer. A backtest is evidence about a simulation, not permission to trade.
Open
07 / RESEARCH PROTOCOL
A ladder, not a leap.
Each stage should produce an artifact that makes the next stage safer. A backtest is evidence about a simulation, not permission to trade.
01Specify
Write the hypothesis, universe, horizon, data cutoff, costs, and invalidation rule.
02Replay
Run a point-in-time backtest with locked assumptions and no tuning on the test set.
03Stress
Use walk-forward windows, multiple seeds, cost sensitivity, liquidity shocks, and regime slices.
04Paper
Run the exact production path without real orders; compare expected vs. realized fills.
05Review
Human review decides whether a tiny, capped live allocation is justified.
08 / EVIDENCE LOG
What the current research says
The field is moving quickly, but live evidence is thin. These links anchor the system’s initial assumptions.
FINRA / AI in securitiesAI is being explored for portfolio research and execution, but autonomous systems can fail in unusual regimes.↗ Read the report
AI-Trader / 2025 benchmarkGeneral intelligence did not reliably translate into trading returns; risk control was tied to cross-market robustness.↗ Read the paper
CFTC / investor advisoryAI cannot predict sudden market changes, and guaranteed high-return bot claims are a fraud warning.↗ Read the advisory
NIST / AI RMFGovern, map, measure, and manage the system continuously, including a way to deactivate it when outcomes drift.↗ Read the framework
SEC / EDGAR APIsOfficial submissions and XBRL facts provide a canonical foundation for point-in-time filing research.↗ Read the API guide
DSR / selection biasPerformance can be inflated by multiple testing and non-normal returns; the trial history belongs in the ledger.↗ Read the paper
09 / SOURCES
Source inventory
Primary regulators and data providers lead; original research and open-source implementations provide context, not guarantees.
Open
09 / SOURCES
Source inventory
Primary regulators and data providers lead; original research and open-source implementations provide context, not guarantees.
- 01AI Applications in the Securities IndustryFINRA · portfolio research, execution, and autonomous-system risks
- 02Algorithmic TradingFINRA · testing, validation, supervision, and post-deployment review
- 03AI Won’t Turn Trading Bots into Money MachinesCFTC · investor protection and AI-return-scam warnings
- 04AI Risk Management Framework CoreNIST · govern, map, measure, manage, oversight, and deactivation
- 05EDGAR Application Programming InterfacesSEC · submissions and extracted XBRL facts
- 06FRED API: Real-Time PeriodsFederal Reserve Bank of St. Louis · revisions and historical vintages
- 07AI-Trader: Benchmarking Autonomous Agents in Real-Time Financial MarketsFan et al. · 2025 preprint · live-style multi-market benchmark
- 08StockBench: Can LLM Agents Trade Stocks Profitably In Real-world Markets?Chen et al. · 2025 preprint · contamination-free sequential benchmark
- 09FinMem: A Performance-Enhanced LLM Trading AgentYu et al. · 2023 preprint · layered memory and decision architecture
- 10FinRobot: AI Agent for Equity Research and ValuationZhou et al. · 2024 preprint · multi-agent research architecture
- 11The Deflated Sharpe RatioBailey and López de Prado · Journal of Portfolio Management · 2014
- 12A Reality Check for Data SnoopingWhite · Econometrica · 2000
- 13FinRL-XAI4Finance Foundation · open-source modular infrastructure reference
Full research notes and the source rationale live in the versioned research brief ↗.
NEXT BUILD
Make the ledger real.
Next, connect this surface to a versioned research ledger: datasets, hypotheses, agent prompts, trades, fills, costs, and post-mortems. The first success metric is reproducibility.