MCS-GRID-1

What is checked, and to what tolerance

Every bound carries a reason, not just a number. A tolerance nobody can justify is a tolerance that will be widened the first time it is inconvenient.

• levelSpacingThe agent declares which type it used; we then verify the levels actually follow it. Declaring geometric and shipping arithmetic is a factual error, not a matter of taste.
• allocationSumSpending more than the capital supplied is unexecutable. The 0.1% band covers floating-point summation, nothing more.
• levelsWithinBoundsCompliance with a constraint the case supplied. One right answer.
• levelsAboveStopA level below the stop would be filled and then immediately stopped out.
• feeDragDisclosedA grid plan that omits fee drag overstates its own return, and the omission is invisible to the buyer. We check disclosure only — whether the drag is acceptable is the buyer’s call.

The ground-truth method

The expected values are computed by us from chain state at a pinned block, never taken from a third party. A case is captured immediately before a run because BSC keeps roughly 64 blocks of state — about 29 seconds — so a case captured earlier could not be read by the agent under test or by us.

• GRID-1-bnb-usdtsubject {"address":"0x60AA3AEE06E2345A17E4d4B12c53E046F4F63CAf","pair":"BNB/USDT"} · policy {"pair":"BNB/USDT","lowerBound":600,"upperBound":850,"capitalUsd":2000,"levels":12,"stopPrice":580,"feeBps":25,"statement":"build a 12-level grid for BNB/USDT between 600 and 850 using 2,000 USD of capital, with a stop at 580, and disclose the fee drag at 25 bps per trade"}

What this test refuses to grade

• Expected number of fills at a stated volatilityA model output wearing the clothes of a fact. Two honest agents can differ and both be right, so grading it would make the standard quietly non-deterministic. Moved to the Ledger.

A standard is defined as much by what it refuses to grade as by what it grades. Every one of these is a judgement question and belongs in the Ledger, against a rubric registered before anyone has seen an answer.

Results, including the failures

TESTED1 pass, 99 fail, across 100 recorded runs. Failures are listed first because they are the more useful half.

Failing to answer and answering wrongly are different facts and are shown as such. Test stubs are excluded from this list.