AR-SIM-047·PL-01 — The calibration bench. Stylised brand illustration of the laboratory, not a photograph of a physical facility. The charts and figures inside the plate are simulated.

Record AR-SIM-047 · Parameter Laboratory

Two hundred and eighty-eight controls.
Nothing moves alone.

This is the Foundry's central instrument. The general parameter catalog specifies 288 controls; 159 of them are running on the bench below right now, open to everyone, for a limited time. Every one feeds a single deterministic model of how the parts of a trading framework interact. Move a slider and the expectancy, the drawdown, the trade count, the friction burden, the tail behaviour and the survival estimate all respond — because in a real system they do. Tap or hover the beside any control to read what it is, what moving it does, where the sensible band sits, and how it is usually misused.

Status Operational Functionality Level 1 Catalog 288 Live on this bench 159 Engines 10 Instant preview measures 73 Groups 15 Engine Deterministic, seeded Updated 2026-08

Open-catalog window — all 159 rails, everyone, until 1 September 2026

Every control on this bench is reachable right now with no account and no payment. That is a one-month window rather than the permanent free tier: after 1 September the free bench becomes fifteen named parameters and the rest of the catalog moves behind the paid ladder. The published general catalog for the planned application is 288 controls — the 159 here plus 129 specified and not yet modelled — tiered at 5, 40, 70 and 100 per cent. Twelve correctness-critical controls sit outside that ladder and stay visible at every tier.

Interactive demonstration — not a historical backtest

The engine reads no market data, no broker history and no trade journal. It models the direction and rough magnitude of the relationships between parameters. It cannot tell you whether a strategy is profitable, and it is not trying to.

Parameter bench

Control set

Strategy
Control set
Disclosure mode

—

Organised by
—

The address bar tracks this bench. Every configuration has its own link, and the link records only what you changed from the defaults.

Sample configurations

Six starting points. Each is a deliberate design position, not a recommendation — two of them are configurations the model rates badly, included because the failure is more instructive than the success.

Interactive demonstration — not a historical backtest

AlphaRail Preview Score

Illustrative

—

—

—

Figure 1 · Illustrative equity and decline Simulated

—

Figure 2 · Distribution of trade outcomes Simulated

—

Reading

—

Strongest factor

—

Weakest factor

—

AR-SIM-047 · Output catalogue

The complete output board —
and which of it is real.

Everything the laboratory is specified to report, in one place: seventy-three measures across five groups. Each one carries its own state. A measure the engine computes today shows a number from the same evaluation as the readouts above. A measure shown illustratively says so. A measure that needs evidence this phase does not hold shows no number at all, and states what it would need.

Simulated demonstration data

These figures illustrate product behaviour and do not represent historical or live trading performance. The split between computed, illustrative and planned is printed below rather than buried in documentation, because a board that renders every row identically is a board that is lying about most of them.

—

—

AR-SIM-047 · Figures 6 to 10

Five pictures the new numbers were missing.

Ratios, drawdown shape and the return-for-risk trade-off are hard to read as a column of figures. Every sweep below re-runs the same deterministic engine at a reduced path count, so no figure can disagree with the number beside it, and each states its own resolution.

Figure 6 · Risk / return frontier Simulated

—

Figure 7 · Rail influence, ranked Simulated

—

Figure 8 · Expectancy surface Simulated

—

Figure 9 · Risk-adjusted ratios Simulated

—

Figure 10 · Drawdown profile — depth against duration Simulated

—

Interactive demonstration — not a historical backtest

AR-SIM-047 · Sensitivity

One control at a time,
across its whole range.

Hold the current configuration fixed, sweep a single control from its minimum to its maximum, and plot what happens. This is the fastest way to see that a parameter has a useful zone rather than a direction.

Figure 3 · Parameter sweep Simulated

—

Figure 4 · Sensitivity — which rails move the score Simulated

—

Figure 5 · Component profile against the balanced default Simulated

—

What the sweep shows

—

A peak is not a setting

If the best value on a sweep sits on a narrow spike with poor values either side of it, that is a warning rather than a discovery. Robust settings sit on plateaus. This is the parameter-neighbourhood condition in the validation protocol, and the reason the score penalises configurations that sit far from the middle of their own safe bands.

Parameter dependencies

Which controls interact, and what the interaction produces. These are the relationships that make single-parameter optimisation misleading.

How this preview works

Stated honestly and in full. If this section and the engine ever disagree, the engine is right and this text is a bug — the source is in assets/js/engine.js and is commented throughout.

Confirmation count, signal and trend thresholds, regime fit, stop fit and the separation between trigger and authority timeframe are combined into a single quality figure between 0 and 1. Each contributes with diminishing return — the second confirmation adds far more than the fourth.

Quality raises the probability that a trade reaches its first target. It does not raise the sample size; every filter that improves quality removes opportunity, and the opportunity cost is charged separately.

The probability of reaching the first target falls exponentially as the target moves away. This single relationship is what stops the model rewarding ever-larger targets, and it is the arithmetic reason a 3R target with the same signal does not produce three times the expectancy.

p₁ = edge × (0.70 + 0.62Q) × e−0.26(T−1) − chase + expansion

Q is the quality composite, T the first target in R. The chase term charges for entries accepted past the trigger; the expansion term is a small credit for trend-family archetypes operating in expanding volatility.

A trade either fails to reach its first target, or it reaches it and is managed. Winners are split by the partial percentage: the banked portion takes the first target, and the remainder either extends to a tail multiple or exits on the trail above break-even.

Break-even activation converts part of the losing population into scratches and part of the winning population into scratches too. Both effects are modelled; the second one is the one traders forget.

Each archetype also carries a probability that a loss gaps through the stop and costs materially more than 1R. That term is what stops a high hit-rate archetype from appearing free.

Spread, two sides of slippage and commission are summed in pips and divided by the ATR-derived stop distance. The result is a cost in R, which is subtracted directly from gross expectancy.

frictionR = (spread + 2·slippage + commission) ÷ (coefficient × ATRtimeframe)

This is why cost is a timeframe decision before it is a broker decision: identical assumptions consume roughly nine times more of the edge on a five-minute trigger than on a four-hour one.

Raw opportunity is the timeframe's base rate reduced by every filter: confirmations, thresholds, the width of the volatility window and the expansion requirement. Chase distance buys a small amount of it back.

Whatever survives is then capped twice — by the number of concurrent slots and by the open-risk ceiling divided by per-trade risk. The lower of the two binds, and the difference between raw and realised opportunity is reported as rejected signal.

The drawdown figure is the median of nine independent seeded histories of the evaluated configuration. Trades are grouped into concurrency blocks; inside a block, correlation pulls outcomes toward a shared sign, which is how four positions quietly become one position held four times. Capital authority is reviewed at block boundaries rather than after every trade.

The equity curve you are looking at and the drawdown number quoted beside it are produced by the same code path, so they cannot disagree.

With the governance layer on, risk is stepped down as the decline from the equity peak deepens, and deployment stops entirely past the lock boundary. Nothing else changes: the edge, the frequency and the friction are identical.

The result is a shallower decline and a slower ascent. That trade is the whole proposition, and the panel reports both halves of it.

Eight weighted components produce a raw score, and then five safeguards cap it. Negative net expectancy caps the score below 50. Risk per trade above 6% caps it below 60. Modelled drawdown beyond 34% caps it below 46. Friction consuming more than 62% of gross expectancy caps it below 55. High correlation across several concurrent positions caps it below 52.

An invalid configuration — for example a volatility window whose minimum sits above its maximum — withholds the score entirely rather than issuing a misleading one.

Score components and weights

Expectancy quality · 19%Net expectancy after friction, in R.
Drawdown control · 16%Modelled maximum decline from the equity peak.
Survivability · 14%Estimated probability the configuration is still operating a year out.
Risk efficiency · 13%Mean outcome divided by outcome dispersion.
Robustness · 11%Sample adequacy, regime fit, and distance from the extremes of the safe bands.
Execution friction · 10%Cost as a share of gross expectancy.
Exposure discipline · 9%Effective open risk, correlation and concurrency, credited for governance.
Tail quality · 8%Balance between tail participation and capture efficiency.

Weights are a design decision, not a measurement. They encode the view that expectancy and drawdown matter most and that tail behaviour is a preference rather than a virtue.

Safe, caution and danger bands

Every rail draws its own operating envelope into the track. The bands are design judgements about where a control stops being a tuning decision and starts being a risk decision — they are not measured optima.

Operating bands for every range control in the parameter laboratory
ControlGroupRange SafeCautionDanger Default

Controls without bands are selectors or toggles, where no ordering of the options implies risk.

Limitations

The honest list. Everything here is a reason not to treat an output of this page as a finding.

It reads no market data

No prices, no fills, no historical sample. The engine is a model of relationships between design decisions, calibrated to produce plausible magnitudes. It cannot confirm that an edge exists, only show you what would follow if one did.

Archetype constants are design choices

Each archetype's baseline edge, tail affinity, preferred volatility band, design coefficient and shock probability are numbers chosen to express how those families are generally understood to behave. They are not measurements of any market.

Outcomes are independent within a block

Correlation is modelled as a shared driver inside a concurrency block and nothing more. Real correlation regimes shift, and they shift hardest in exactly the conditions that matter most.

The shock term is a single number

Gap risk is modelled as one probability of an oversized loss per archetype. Real gap risk clusters, arrives with correlated positioning, and is worse than any stationary term can express.

Nine histories is a small median

Nine seeded runs give a stable illustration, not a distribution. The Monte Carlo Core exists precisely because nine is not enough when the question is about the tails.

Nothing here models you

The largest source of variance between a modelled system and a traded one is whether the rules are followed. That is not in this engine and cannot be.

Take a configuration to the Monte Carlo Core Risk and drawdown Framework library