PROOF BEFORE CAPITAL
Replay a revision, then keep the result.
The four Backtest peers - Workbench, Lab, Ledger and Edge - what each one is for, which of them cost credits, and what makes a run citable later.
THE QUESTION
A backtest answers what this would have done, and nothing else
A backtest replays a frozen strategy revision over stored history on the same kernel the chart and a live automation use, so the three cannot disagree about what an indicator said. It does not answer whether the edge will persist, and it cannot see anything the data does not contain - a fill assumption, a fee schedule and a bar boundary are all decisions the run made on your behalf.
Read a result as a record of a decision procedure, not as a forecast. The parts of Backtest that matter most are the ones that keep the record: what was run, over what, under which assumptions, and whether that answer can still be produced.
THE FOUR PEERS
Which surface answers which question
They share one evidence vocabulary and one top card. Moving between them never changes the revision you are working on.
| Peer | Question it answers | Plan |
|---|---|---|
| Workbench | What does this revision do over this window, right now, in this browser? | Every plan, with a daily allowance on Free |
| Lab | What does a bounded experiment say once every trial is counted, including the failures? | Quant |
| Ledger | What durable runs and experiments does this account already own? | Every plan |
| Edge | Which part of a parameter space survives out of sample? | Pro |
TWO KINDS OF RUN
A local run is free to compute and easy to lose
The Workbench computes in your own browser, over public candles, on your own CPU. It costs Meridian nothing, which is why the allowance is generous and why the result lives only in the tab that made it.
A durable server run is the same workload sent to the queue, where it is executed, stored and given an identity you can cite from a Trade Plan or a Review. That is what credits pay for - not the answer, which you have already seen for free, but the record of it.
- Promote, do not re-runPromotion dispatches the same spec that produced the local result, so the durable copy is the run you already read rather than a second one that happens to look similar.
- A quote comes firstEvery durable dispatch is quoted before it is charged, and the estimate you approved is the ceiling the reservation holds.
- A timeout is not retriedA deterministic computation that ran out of time will run out of time again; a silent retry would be a second bill for the same answer.
FIRST RUN
Get one honest result
Bring a saved revision
Compose the logic in Strategies first. A revision is versioned, which is what lets a result stay attached to the thing that produced it.
Open StrategiesSet the window before the parameters
Choose the history and the timeframe you actually intend to trade, then leave them alone while you vary the logic. Changing both at once produces a number nobody can attribute.
Read the evidence, not the headline
Trade count, drawdown and the fill assumptions decide whether a return figure means anything. A high return over eleven trades is a sample, not a result.
Keep the ones worth citing
Promote a run to the durable ledger when you intend to reference it later - in a Trade Plan, a Review, or a decision you will have to justify to yourself in three months.
Open the Ledger
The consoles put an inspector rail beside the run it inspects and expect a pointer for the split. On a narrow screen they say so and point at the surfaces that do read well there; the Ledger and Dashboard are the ones to reach for.
NEXT