Scenario Manager · Scenario Discipline
A scenario is a contract, not a mood.
Saved slots
10
Prebuilt presets
5
Variables per run
1
Read this first
Three ways to define a run
01Manual, preset, or saved slot, and the difference matters.
The configuration source is itself a setting. Manual means typing values freely. Preset means following the scenario library's advisory assumption pack, which drives hit-rate, fee, tail, and volatility modifiers together rather than one at a time. Saved slot recalls a recorded configuration from one of ten stored states. The slots are what make a comparison reproducible weeks later: a saved run configuration is a claim you can re-examine, and a remembered one is not.
The one-variable law
02A scenario that changes two things explains neither.
A run that adjusts the stop policy and the branch weights simultaneously produces a result nobody can attribute. The contract format enforces isolation: one variable moves, everything else holds, and compound proposals get decomposed into a chain of scenarios run in sequence. It is slower, and it is the only version that produces answers rather than impressions. Stress packs are the deliberate exception, because a coordinated hostile environment is itself the variable under test.
- The cost model selector is a single lever with three positions: off, standard live assumptions, or the stress pack.
- Stress packs apply a coordinated set of hostile overrides in one action rather than by hand-editing each tab.
- A fixed seed across a scenario chain removes sequence luck from every comparison in the chain.
Persistence before verdict
03A profile must look better across weeks and months, not one flattering sample.
Comparing an active profile against alternates over real evidence asks one question: was the alternative structurally superior, persistently? A single window that favours a candidate is the expected behaviour of any search across enough candidates, which is precisely why the persistence requirement exists. Without it, the scenario library becomes a machine for discovering whichever configuration best fits the most recent stretch of data.
Sandbox by construction
04Scenario outputs stay isolated until deliberately promoted.
Nothing produced by a scenario touches production rules by default. The isolation is architectural rather than procedural: the engine writes to a separate run workbook, the scenario library holds its own assumption columns, and the benchmark standard forbids integrating research experiments into benchmark comparisons until the new model has been tested and accepted as a new benchmark version. Promotion is a decision, and it has its own workflow.
How MARS uses this
A scenario contract specifies expected behavior before the run — and the grid is where contracts get judged. Each column is one controlled condition: same evidence, one variable changed, outcomes rendered against the P10 tail so the contract's promise can be checked instead of remembered.
BASELINE
LIVEcurrent governing profile
TIGHTER STOPS
REJECTEDEV cost exceeds drawdown saving
TNP WEIGHT +10
SANDBOXedge up, adverse tail deepens
FEE MODEL B
CANDIDATEfriction saving survives resampling
How it benefits you
System changes stop being vibes-based. The tempting tweak that costs 0.07R of expectancy for a modest drawdown saving gets rejected by arithmetic before it silently taxes six months of trading - and promising candidates carry their evidence with them into review.
Contracts under judgment: one variable per column, same evidence throughout, the tail cost always in view.
Reference
The prebuilt scenario presets and what each one stresses
| Preset | What it changes | What it exposes |
|---|---|---|
| Baseline | Nothing; the control condition | The reference every other run is measured against |
| Conservative | Tighter assumptions across hit rate and deployment | Whether the system still completes when the edge is assumed thinner |
| Aggressive | Looser risk and deployment assumptions | Where lock probability begins to climb faster than completion |
| High-Fee Stress | Escalated commission, spread, and swap drag | How much of the modelled edge is actually friction-sensitive |
| Trend Collapse | Degraded trend-branch behaviour | Concentration risk in the branches carrying the fat tail |
Edge cases & failure modes
Where it can mislead
- !A scenario without a written expected behaviour cannot fail, which makes it a demonstration rather than a test.
- !Comparing scenarios under different seeds reintroduces the sequence luck the contract was designed to remove.
Worked example
The Dynamic 7-Tier Benchmark converts the drawdown distribution into modeled tier-usage expectations: given the gate transitions the evidence implies, T1-T3 should host most deployment while T5-T7 stay exceptional. Live tier dwell is audited against this profile every cycle.
The seven-tier ladder with modeled dwell: how often each tier should be in use given the gate history. High tiers exist to be rare.
Connected inside MARS
This module doesn't work alone.
Go deeper
Operator briefs on this territory.
Deep dive — 01
Define expected behavior before running the test.
Write the expected behavior first, or the run will mean whatever you wanted it to.
Read the full brief →
Deep dive — 02
Compound ideas get decomposed, and the decomposition has a blind spot.
Three isolated tests answer three questions and cannot answer whether the three interact.
Read the full brief →
Deep dive — 03
A library beats a blank page, especially when the question is hostile.
Ad-hoc testing tests what you thought of. A library tests what you would rather not.
Read the full brief →
Take it further
A scenario worth promoting usually survives twenty that were not. The Foundry is where the other twenty go.
Open the AlphaRail Foundry Lab ↗
Every module ships in the complete MARS package.
One price. Eleven workbooks, three TradingView indicators, and the full manual library — $497.