3.5 movalidation cycle, not a one-chart demo
15+ yrhistorical replay across multiple market regimes
100 × 15 yrillustrated historical design; not current Engine results
155historical check count—not current coverage
50historical module count—not a robustness guarantee
These are archived research categories, not a certification of the locked Engine. Historical returns, illustrative scenarios and current deterministic safeguards are separate kinds of evidence.
Cadence and rebalance testing
We tested daily, weekly, 60-session, quarterly, and smart-cadence behavior so the system does not churn strong books just because a calendar flipped.
- Purpose
- Routine rotation can wait; de-risking cannot.
- Protects
- Overtrading, stale exits, and fake activity.
Long-history backtests
Earlier strategy rules were replayed through a 15+ year market path with benchmark and execution-lag assumptions. Those figures are not current Engine results.
- Purpose
- Performance has to survive the full window, not a cherry-picked slice.
- Protects
- One-regime winners and hindsight-only rules.
Illustrative alternate histories
The diagram illustrates 100 alternative 15-year paths and a 7.5-year held-out half. It is not measured Engine returns, a forecast or proof against overfitting.
- Purpose
- Risk mechanics cannot depend on one lucky order of returns.
- Protects
- Hindsight, invented alpha, and pretty one-path curves.
Drawdown and breaker studies
Deep-bear periods, volatility shocks, credit stress, and re-entry timing were tested around the risk meter, circuit breakers, and drawdown kill behavior.
- Purpose
- Risk can step down before the book is already broken.
- Protects
- Panic exits, whipsaw re-entry, and uncontrolled downside.
Weights, leverage and profile fit
Testing reaches portfolio construction: position caps, hedge floors, concentration, profile budgets and leveraged-instrument substitutions—not merely individual stock picks.
- Purpose
- The right stock still needs the right portfolio weight.
- Protects
- Hidden leverage, crowded names, and risk-profile mismatch.
Data integrity and source freshness
Adjusted prices, split handling, fundamentals, missing rows, stale source checks, and no-data states are treated as gating inputs instead of glossed-over blanks.
- Purpose
- Garbage-in never earns a confident customer answer.
- Protects
- Bad prices, stale metrics, and silent fallback data.
AI-capex and macro risk sweeps
Research panels show capex, semiconductor volatility, rates, liquidity and breadth. These are not all automatic Engine sizing or entry inputs.
- Purpose
- AI winners need sector-specific stress context.
- Protects
- Buying the hottest names as financing conditions turn.
Engine → LLM → Engine
Equity Quant records its output; Claude Opus and GPT Sol review selected facts and suspected sizing errors where available. Independent deterministic checks repair proven defects; AI cannot pick replacements or set weights.
- Purpose
- Model judgment can be useful without owning the portfolio.
- Protects
- Scope escape, hallucinated weights, and provider failure.
What was testedWhat we rejectedWhy it matters
Ranking formulasSingle-metric CAGR screens, risk-ratio ties, and factor mixes that only won in narrow regimes.Better marketing numbers are useless if they break under stress.
Risk timingRules that waited too long to de-risk, re-entered too fast, or treated every volatility spike the same.Equity Quant needs patience in noise and urgency in real damage.
Portfolio constructionOversized winners, hidden leverage, one-theme crowding, and hedge settings that looked safe only on average.Customer outcomes depend on the whole book, not the best ticker.
Evidence qualityUnpaired forward tests, stale-source shortcuts, blank analyst outputs, and fallback numbers presented as facts.If the evidence is not clean, the answer stays restrained.
Research archive: earlier methods and validation questions, not verified performance of the current locked rules.