and a great deal of measurement. The Foundry began as the analytical layer of a working governance system. The corridor is how it is organised: separate questions, identical evidence.
Record AR-FAC-001 · Facility statement
A laboratory built because
the analysis kept mattering more.
The AlphaRail Foundry Laboratory is a strategy-engineering and market-research environment. It exists to make the parts of a trading framework that decide the outcome — stop geometry, concurrency, correlation, friction, capital authority — as inspectable as the entry rule that usually gets all the attention. This page states where it came from, how it works, what it will and will not claim, and who it is for.
01Why the facility existsAR-FAC-001·01
Traders do not lose
to bad entries.
They lose to the parts of the system nobody specified. A strategy is usually described by its trigger, because the trigger is the part that was consciously chosen. Everything downstream of the trigger — how wide the invalidation sits, whether a partial is taken, how many correlated positions are permitted to be open at once, what the spread costs as a share of the working stop, and what happens to position size after four losing weeks — tends to be inherited from a default, a habit, or a mood.
Those inherited decisions are where the equity curve is actually determined. A strategy with a mediocre entry and disciplined exposure control survives; a strategy with an excellent entry and uncontrolled concurrency does not. Yet almost every tool a private trader can reach answers the question “did this rule make money on this history?” and no other.
The Foundry is built around a different question: under what conditions does this stop working, and would you notice before it cost you the account? Answering it requires separating a framework into its components, moving them independently, and watching the disagreements. It requires treating an equity curve as one draw from a distribution rather than as evidence. And it requires a stated set of conditions a result has to survive before anyone is allowed to call it a conclusion.
None of that is exotic. It is ordinary research practice, applied to an activity where research practice is unusually rare and unusually expensive to omit.
Four gaps the facility addresses
The facility in one sentence
A place to decompose a complete trading framework, move its assumptions deliberately, and record what survived — with every number labelled as the model output it is.
02OriginAR-FAC-001·02
It started as a spreadsheet
that would not stop growing.
The Foundry did not begin as a product idea. It began as the analytical layer of a working governance system, and it became a product idea when that layer turned out to be more generally useful than the system it was serving.
MARS — the Montex AlphaRail System — is a spreadsheet-based trading governance system. Its purpose is narrow and specific: to decide how much capital a trader is permitted to deploy, given the current state of the account, the measured evidence of edge, and the exposure already open. It is built on two instruments. Monte Carlo provides the benchmark for what should happen. Expectancy provides the grade for what is happening. MARS is the layer that decides what you are permitted to do about the gap between them.
Building that system required a great deal of machinery that was not, strictly, governance. Drawdown had to be decomposed into depth, duration, time under water and recovery efficiency before a capital state could be derived from it. Execution quality had to be measured in adverse and favourable excursion before a stop policy could be defended. Concurrency had to be adjusted for correlation before an exposure cap meant anything. Branch populations had to be separated before attribution was possible at all. Each of those became a workbook module, and then a family of modules, and then a body of measurement that outgrew its original purpose.
The observation that produced this site was simple: the analytical layer deserved to be a product in its own right. The governance rules in MARS are calibrated to one operator's risk doctrine and are not transferable. The analysis underneath them — how friction scales against the working stop, how a partial trades capture efficiency for tail participation, how four correlated positions become one bet held four times — is transferable, and almost nobody has access to it in a usable form.
So the Foundry rebuilds that analytical layer as an environment rather than as a workbook, and the MARS governance chamber on this site describes the structure of the system it came from without publishing its calibration.
MARS · name decode
What is deliberately withheld
Drawdown gate boundaries, pool percentages, tier authorisation tables and the workbook's Monte Carlo completion figures are part of the MARS product and are not published on this site. The structure is described because the structure is the educational content; the calibration is not, because a number lifted out of the doctrine that produced it is worse than useless.
01
A governance problem
How much am I allowed to risk today, given a drawdown, an open book and a record of mixed evidence? The question that produced MARS.
02
A measurement problem
Answering it required decomposing drawdown, execution, exposure, correlation and branch attribution — each of which became its own module.
03
A transferability problem
The governance calibration is personal. The measurement layer is not. One of those can be published usefully; the other cannot.
04
A build
The Foundry is that measurement layer rebuilt as a research environment, with the governance chamber describing structure rather than settings.
03Operating principlesAR-DOC-001
Decomposition, experimentation,
validation — in practice.
The three principles are stated briefly on the homepage. Here is what each one costs to actually follow, because a principle that is free to obey is not a principle — it is a slogan.
Decomposition is easy to agree with and hard to sustain, because aggregation is where comfort lives. A composite equity curve is smoother than any of its components. A blended win rate is steadier than the branch rates inside it. A single drawdown figure hides the fact that the account spent sixty per cent of the year below its previous peak.
The operational rule is that no measurement is reported only in aggregate. Expectancy is reported per branch before it is reported per system. Drawdown is reported as depth, duration, time under water and recovery efficiency rather than as a maximum. Concurrency is reported both nominally and after correlation adjustment, because those two numbers routinely differ by a factor of three. Friction is reported as a share of the working stop, not as a currency amount, because that is the form in which it is comparable across timeframes.
The cost is that the outputs are less flattering and harder to read. A dashboard that resolves to one number is more saleable than one that resolves to eight disagreeing ones. The eight are kept.
Experimentation here does not mean trying settings until one works. That procedure has a name — it is a search over a noisy surface — and it reliably produces a configuration that describes the sample rather than the market.
The rule is that every control propagates. In the parameter laboratory, moving the stop coefficient does not simply widen the stop: it changes friction as a share of that stop, changes the outcome ladder, changes trade frequency because fewer setups clear a wider invalidation, changes tail participation, and changes the drawdown envelope. If a tool lets you move one input and holds the rest fixed, it is not modelling a system — it is modelling a spreadsheet cell.
The second rule is that the experiment has to be declared before it is run. A configuration arrived at by wandering, then justified afterwards, is a story about a search path. The research archive records the proposal, the calibration and the verdict as separate steps for that reason.
Validation is the principle most often confused with its opposite. A result is not validated by being good. It is validated by surviving a declared set of attacks: resequencing, friction stress, hit-rate degradation, out-of-sample separation, regime partitioning, drawdown stress, parameter neighbourhood inspection and sample adequacy.
The eight conditions are listed in full on the laboratory overview, and each one has a defined failure state that is recorded as readily as a pass. Failed experiments stay in the archive with their reasoning intact. A negative result that is deleted gets rediscovered later at greater expense, usually with real capital attached.
The uncomfortable implication is that most configurations do not validate, including configurations that look excellent on their primary metric. That is the expected outcome of the process working, not a fault in it.
There is a fourth commitment that does not appear in the doctrine because it is a matter of hygiene rather than of method: no figure is published without the configuration that produced it, and no model output is displayed without a marker at the point of display saying that it is a model output.
This is why every chart on this site carries a simulation marker in its own header rather than in a footnote at the bottom of the page. A disclosure that requires scrolling is a disclosure designed not to be read.
Principles are constraints
Each of the three makes the outputs less impressive and the build slower. That is the test of whether they are real.
Where they conflict
Decomposition wants more detail; usability wants less. When the two collide, the detail is kept and the interface is asked to work harder.
What they cannot do
No amount of method converts a model into a market. The principles govern how this site reasons, not whether the conclusions transfer to your account.
04Research standards · methodologyAR-STD-001
What has to be true before
a number appears on this site.
Five conditions. They are procedural rather than aspirational: each one is checkable, and a figure that fails any of them does not get published, however interesting it is.
| Standard | Requirement | How it is enforced | Failure mode it prevents |
|---|---|---|---|
| Deterministic | The same inputs produce the same outputs on every machine, every reload, forever. | No unseeded randomness anywhere in the page layer. Every stochastic figure draws from a seeded generator held in the engine. | A chart that changes when you refresh it, so no reader can check a claim twice. |
| Seeded | Where randomness is part of the model — resequencing, Monte Carlo path generation — the seed is an explicit input. | Seeds are exposed as controls where they matter and fixed where they do not. Changing a seed changes the picture; that is a feature and it is stated. | Cherry-picking a favourable draw and presenting it as the model's behaviour. |
| Documented | The configuration that produced a figure is stated near the figure, in readable form. | Figure captions carry the interpretation; parameter pages carry the full control state; methodology blocks state what the engine computes. | A number with no provenance, which cannot be argued with and therefore cannot be trusted. |
| Disclosed | Every model output is labelled as a model output at the point of display, not in a footer. | Simulation markers sit inside figure headers. Future capability is labelled with an explicit level pill and future-tense language. | A reader taking an illustration of a relationship for a measurement of history. |
| Reproducible | A reader with the page open can reconstruct the result by moving the same controls to the same values. | The whole engine runs client-side. There is no server computation to take on trust, and the source is readable in the browser. | An unverifiable claim protected by an opaque backend. |
The standard that does most of the work
Determinism. It sounds like an engineering detail and it is actually the load-bearing one, because it is what makes disagreement possible. If two readers open the same page with the same controls and see different numbers, there is nothing to discuss. If they see identical numbers, the discussion can move to whether the model deserves belief — which is the only discussion worth having.
It also imposes a discipline on the build. There is no Math.random() in any page script on this site. Stochastic behaviour is generated from a seeded generator inside the engine, so a Monte Carlo panel showing a thousand paths shows the same thousand paths to everybody, and a changed seed is a deliberate act with a visible cause.
The standard that is hardest to keep
Documentation. Writing the configuration next to the figure is tedious, and the temptation to publish a striking chart with a thin caption is constant. The rule holds because the alternative is a site full of assertions, and an assertion about trading performance with no method attached is the exact genre this facility exists in opposition to.
Where a claim rests on the MARS workbook rather than on the preview engine, the source is named and the calibration is withheld — which is itself a documented position rather than a silence.
05Honesty policyAR-STD-002
Seven commitments,
and what each one costs.
Stated as commitments rather than as values, because a commitment can be broken visibly and a value cannot. Each entry gives the rule and then what following it actually means in practice.
How to hold this to account
If you find a figure without a disclosure marker, a claim without a stated assumption, a button that does nothing, or future capability described in the present tense, that is a defect and it should be reported. The contact form has a report a problem enquiry type specifically for it. Corrections are made in the build log on the status page rather than silently.
06Scope · limitationsAR-FAC-001·06
Five things the Foundry
is not, stated plainly.
Negative definitions are usually more informative than positive ones, and they are certainly harder to weasel out of later.
Also not, more briefly
Read this before anything else
The full statement of what the preview engine does and does not model, and the risks that no model addresses, is set out on the risk disclosure. It is the most important page on this site.
Open the risk disclosure →07AudienceAR-FAC-001·07
A narrow audience,
deliberately.
Most trading products are built to be for everybody, which is why they end up teaching nobody anything. This one assumes a specific reader and would rather be wrong for you than vague for everyone.
Built for
Not built for
08Construction · how this worksAR-ENG-001
No frameworks, no trackers,
no external requests.
This is unusual enough in 2026 to be worth stating explicitly, and it is a research decision rather than an aesthetic one. A site that makes claims about determinism and reproducibility should be inspectable by the person reading it.
Why build it this way
Because the alternative undermines the argument. A site that insists on reproducibility while computing its charts on a private server is asking to be taken on faith. A site that publishes an honesty policy while loading a tracking pixel is publishing a different document from the one it appears to be publishing.
There is also a practical benefit. Determinism is much easier to guarantee when there is no network in the loop: the same code, the same data and the same seeds produce the same output on every visit, on every device, indefinitely. No API version drifts underneath a published figure.
What it costs
The first load is heavier than a conventional site, because the fonts and the full data layer arrive up front rather than being streamed in from elsewhere. There is no personalisation, no saved state between visits beyond the motion toggle, and no server-side search. Those are acceptable prices for a preview whose job is to be inspectable.
When the application ships, parts of this necessarily change — accounts and uploaded data require a server. The privacy statement sets out the intended handling principles for that phase in advance, so the change can be measured against something.
09Following the buildAR-FAC-001·09
Three ways to watch this develop,
none of which involve a newsletter cadence.
01
The facility status page
Module-by-module operational state, a dated build log of what was constructed when, the current known issues, and what is being worked on next. Updated whenever something material changes rather than on a schedule.
02
The research archive
Every study carries a status — proposed, in calibration, simulated, under review, conditional, validated, failed, archived — and moves through them visibly. New records appear as they are written, including the ones that fail.
03
The development register
A single low-volume list for people who want to be told when a module becomes usable. It is a register, not a marketing funnel: no drip sequence, no urgency, no offer that expires.
The roadmap, in short
The planned application is organised into five stages — build, test, diagnose, govern, refine — and twenty-five modules across them. The status page states which are operational today; the roadmap states what each is intended to become. Neither page describes an unbuilt module in the present tense.
Read the application roadmap →Contact, correction
and collaboration.
The most useful message this facility can receive is a specific disagreement: a relationship the engine models the wrong way round, a claim that overreaches its evidence, a figure whose caption does not match what it shows. That kind of correction changes the build. General encouragement, while pleasant, does not.
Research collaboration
If you hold a validated dataset, a methodological objection, or a study you want replicated under these conditions, say so. Joint work is credited and its limitations are published alongside it.
Corrections
Errors of fact, missing disclosures and broken controls are treated as defects. They are fixed and the fix is recorded in the build log rather than applied quietly.
Accessibility
If any part of this site is unusable with your assistive technology, that is a defect of the same severity as a wrong number. The accessibility statement explains what was checked and what was not.
No account required · nothing is uploaded · every output is a labelled model result