Record AR-LAB-001 · Laboratory Overview
This page is the map of the facility. It states what the laboratory is trying to establish, what a complete trading framework is made of, how you would work through it, what the parameter engine actually models, how an experiment moves from proposal to verdict, and — at length — what none of this can tell you.
01Laboratory MissionAR-LAB-001·01
It is “under what conditions does this stop working, and would I notice before it cost me the account?” That is a harder question, and it cannot be answered by looking at a single equity curve produced by a single set of assumptions.
Almost every tool available to a retail trader answers the first question. Enter a rule, get a history, read a profit factor. The rule is the part the trader consciously chose, so the tool is built around it — and the parts that actually determine the outcome are left as defaults nobody examined.
The Foundry inverts that. The rule is one of seventeen inputs. Everything around it — stop geometry, partial policy, concurrency, correlation, friction, and the authority that reduces exposure when the account is in trouble — is treated as a first-class object with its own behaviour, its own failure modes and its own diagnostics.
Four operating commitments
What the laboratory is not
It is not a backtester, a signal service, a broker, an optimiser that hands you settings, or a route to a guaranteed result. It reads no market data in this phase, and the roadmap is explicit about which parts of that are still absent.
02Framework AnatomyAR-FRM-SPEC
The seventeen are not a checklist to be worked through in any order. They form a dependency chain: the definition phase decides what may be traded at all, trade geometry decides the shape of a single outcome, population architecture decides how many of those outcomes overlap, cost and authority decide what the account is charged and who may stop it, and evidence decides whether any of it can be believed. A change in an early phase invalidates measurements taken in a later one.
| Phase | # | Component | What it determines | Studied in |
|---|
The phase that is skipped
Population architecture. Most traders run three or four management styles under one name and evaluate the blend, which is why a deteriorating population can sit hidden inside a stable composite for months.
The phase that is assumed
Cost and authority. Friction is usually modelled as a fixed number rather than as a share of the working stop, and governance is usually a private intention rather than a rule with a defined trigger.
The phase that is retrofitted
Evidence. Diagnostics chosen after a result is known will tend to explain that result. The measurements have to be declared before the sample is examined, which is why they are part of the framework specification rather than a reporting afterthought.
03Preview Engine · MethodologyAR-SIM-047
Every interactive number on this site comes out of one deterministic model. It is worth being precise about what that model is, because the difference between it and a backtest is the difference between a claim about relationships and a claim about history.
The engine takes 159 named controls and a strategy archetype. From the archetype it reads a baseline win probability, a tail multiple, a preferred volatility band and the ATR coefficient the logic was designed around. From the controls it derives a working stop in pips, a friction charge expressed against that stop, an outcome ladder, a trade frequency, an effective concurrency after correlation, and a drawdown envelope.
Those feed a seeded path simulator. Trades are grouped into concurrency blocks; inside a block, correlation pulls outcomes toward a shared sign, which is how “four positions” quietly becomes one position held four times. Capital authority — if the governance layer is enabled — is reviewed at block boundaries rather than after every trade, and risk is compressed as the decline from the equity peak deepens.
There is no market data anywhere in that chain. The engine encodes directional relationships that are argued for in the research records: that friction scales against the stop rather than against the trade, that partials trade tail participation for capture efficiency, that correlation converts concurrency into concentration. It shows you what moves when you change something. It does not show you what happened.
The same function is used everywhere. The figure below runs it twelve times — once for each archetype, each at its own design coefficient and otherwise at the default configuration — so the comparison is between strategy shapes rather than between tunings.
Twelve archetypes evaluated at their own design ATR coefficient.
Interactive demonstration — not a historical backtest
Strategy specimen
Decorative scene: a single framework rendered as a rotating specimen with its four management populations orbiting a shared trunk. Everything it depicts is stated in the text on this page and on the branch architecture record.
04Visitor WorkflowAR-LAB-001·04
Five stages, in order, each answering a different question about the same framework. This is a method of enquiry rather than a product tour: the stages are worth following even where the corresponding module is still described rather than built. Each card states what you would be asking, what you would do here today, and what goes wrong when the stage is skipped.
Why the order matters
Diagnosing before testing produces explanations for noise. Governing before diagnosing compresses risk without knowing which population caused the decline. Refining before governing optimises a system that has no defined survival condition. The sequence is the method.
05Parameter Engine MapAR-SIM-047·MAP
The 159 controls are organised into fifteen groups. The grouping is not cosmetic — it reflects which outputs each group is capable of moving. Reading the right-hand column is the fastest way to understand why a change in one place shows up somewhere apparently unrelated.
| Group | Controls | What it governs | Outputs that respond |
|---|
The Friction group moving one output across all twelve archetypes.
Reading the map
Three groups can invalidate a configuration outright rather than merely degrade it. The Regime group can define a volatility window with no interior, in which case no score is issued at all. The Risk group can set a per-trade exposure at which the modelled drawdown makes recovery unrealistic. The Friction group can consume the entire gross edge, at which point nothing downstream is worth reading.
The remaining eight groups trade one desirable property for another. There is no setting that improves everything, and a panel that appeared to offer one would be modelling something other than a market.
06Experiment LifecycleAR-VAL-020
Every record in the archive sits in exactly one of these states, and every state has an explicit exit criterion. A record does not advance because it looks promising; it advances because a stated condition was met. Two of the states — Failed and Archived — are terminal, and both are used.
Failed is a destination, not an accident
A record marked Failed keeps its abstract, its method and its reasoning. The archive currently holds a regime-classification study that failed precisely where it would have been useful — at the transitions — and that record is more instructive than several of the successful ones.
Validated is a description of process, not of profit
It means a result survived eight stated conditions. It does not mean the relationship will hold in a market, that it will hold next year, or that a trader implementing it will experience anything resembling the model.
07Validation ProtocolAR-EXP-102
The protocol is deliberately ordered by cost. Resequencing and friction stress are cheap and kill most candidates, so they run first. Sample adequacy sits last not because it matters least but because it is the condition most likely to send a result back to the beginning.
| # | Condition | What it does to the result | What it catches |
|---|
A result that has not been through all eight is recorded as Simulated. It is never recorded as Validated, regardless of how favourable it looks.
08Data Source OptionsAR-SIM-063
Only one of these is operational today: the configuration you build by hand in the parameter laboratory. The other five are described honestly with the functionality level they currently occupy. Nothing on this site uploads, processes or transmits a file in this phase.
Source 01
Level 1One hundred and fourteen controls set by hand. No history required, and the only route that works end to end today. Appropriate for asking structural questions — what happens to the tail if the partial rises — rather than questions about your own trading.
Source 02
Level 2A closed-position report from a trading platform. The most common source and the most inconsistent one: fee columns differ by broker, partial closes appear as separate rows, and timestamps rarely arrive in the timezone the sessions were traded in.
Source 03
Level 2A manually kept record. Usually the only source carrying the fields that matter most — intended stop, branch tag, setup identity, the reason for the exit — and usually the source with the largest number of missing rows.
Source 04
Level 3Bar data for the instruments a framework is permitted to touch. Required for anything that deserves to be called a backtest, and the single largest gap between the laboratory as it exists and the laboratory as it is described. Not present in this phase in any form.
Source 05
Level 2Aggregated statistics per management population rather than per trade: count, hit rate, mean outcome, dispersion. Enough to compare architectures and to run resequencing, and not enough for execution diagnostics, which need trade-level excursion data.
Source 06
Level 3A structured extract from the Montex AlphaRail System workbook, carrying branch identity, gate state and tier authority alongside the trade record. The richest possible input, and the one that depends on a separate product being present.
09Planned Application ModulesAR-FAC-002
The planned application is described in full on the roadmap. This is the short version: five stages, twenty-five modules, and a level attached to each so that nothing on this site claims to exist when it does not. Level 1 is built and working now, Level 2 is an interactive preview of intended behaviour, Level 3 is future production capability.
Level distribution
The distribution is the honest summary of the project's state: a working analytical core, a substantial band of behaviour that is designed and demonstrated but not yet built, and a group of modules that are currently sentences rather than software.
Full roadmap with dependencies →10LimitationsAR-LAB-001·10
These are not disclaimers attached for legal comfort. They are the specific boundaries of the model, and each of them changes how the outputs on this site should be read.
No market data is read
Not one price, in any instrument, on any timeframe. The archetype parameters are chosen to be plausible and are argued for in the research records, but they are authored values. A configuration that scores well here has demonstrated internal coherence, not market performance.
No fills are simulated
There is no order book, no queue position, no partial-fill logic and no modelling of what happens when a stop is triggered in a thin market. Slippage is an average charged symmetrically, which is precisely wrong in the conditions where slippage matters most.
This is a model of relationships, not of markets
The engine encodes how parameters trade against one another — tail against capture, frequency against quality, cost against timeframe. Those relationships are defensible in general and approximate in every particular case. The directions are the content; the magnitudes are illustrative.
The sample problem does not go away
Several archetypes model fewer than a hundred trades a year. At that frequency, the confidence interval around an expectancy estimate is wide enough to contain both a good system and a bad one. The engine warns when a configuration falls below its adequacy threshold; it cannot manufacture evidence that does not exist.
Tail risk is the weakest term in the model
Shock losses — trades that gap through the stop and cost materially more than 1R — are represented by a per-archetype probability. That is the term with the smallest real-world sample behind it and the largest influence on survival estimates for reversion-family strategies.
The model cannot tell you whether a strategy is profitable
It can tell you that a configuration is internally inconsistent, that friction consumes the modelled edge, that concurrency is concentrated, that the sample is inadequate, or that the drawdown envelope makes recovery unrealistic. It cannot tell you that something works. Nothing that reads no market data can.
One more, about MARS
The governance layer is described structurally throughout this site: capital authority is a function of decline from the equity peak, it is reviewed on cycle boundaries rather than continuously, and it steps down before a situation becomes critical. The specific gate boundaries, pool percentages and tier tables are part of the MARS workbook product and are deliberately not published here. Where a figure on this site applies governance, it uses an illustrative ladder, not the workbook's calibration.
11Related Research RecordsAR-EXP
Each of these produced a design decision rather than an observation. They are the reason the engine charges friction against the stop, treats concurrency as correlated, resequences before reporting a drawdown, and refuses to mark anything Validated on a single history.
The fastest way to understand what this laboratory claims is to move one control and watch eight outputs disagree with each other. Raise the partial percentage and watch the tail collapse. Drop the trigger timeframe and watch friction eat the edge. Turn the governance layer off and watch the drawdown envelope open.
No account required · nothing is uploaded · every output is a labelled model result
AR-CFG · Configuration inheritance
Access · Two ways in
What the application calls things
What the application asks on your first run