What does a real call return?
For y = x1 × x2 with both inputs 1 ± 1 (independent normal), first-order GUM gives a standard uncertainty of 1.414, while Monte Carlo with 200,000 draws gives 1.732, close to the exact value √3 ≈ 1.732. The tool returns linearization_inadequate: the linear approximation misses the product of the two deviations, so the Monte Carlo interval is the one to use.
Recorded on 6 November 2026 through the public MCP endpoint, service release 0.17.1. Inputs and full responses, response SHA-256 1f2e79db5bfe…. Verdict values are the literal values the tool returns. Run the same inputs yourself to compare; elapsed times vary, and a new release may change the figures.
Checked against
- JCGM 101 mass-calibration examples, exact moments of squared normals, lognormal quantiles and an independently seeded NumPy oracle.
Recorded on release 0.16.1. These checks cover the listed cases, not every input. Evidence cards.
How it works
A bounded expression parser checks dimensions and converts compatible units. Forward-mode derivatives produce a first-order GUM (Guide to the Expression of Uncertainty in Measurement) uncertainty budget. Fixed-sample Monte Carlo in ten seeded blocks propagates normal, uniform, triangular or constant inputs. The tool compares the first-order and simulated coverage intervals and reports convergence or undefined draws explicitly.
Inputs and units
Provide an expression of at most 512 characters and 1–16 named variables with values, distributions and optional units. Supply standard uncertainty for a normal input or half-width for uniform/triangular inputs. Correlations may be declared between normal inputs. Choose 10,000–2,000,000 samples, a seed, coverage from 0.5 to 0.999 and an optional compatible result unit.
What the result contains
A located dimensional inconsistency or model result with first-order sensitivities, a per-variable budget, simulated summaries and coverage intervals, linearization assessment, convergence information and a formatted statement. No uncertainty, undefined draws and insufficient Monte Carlo convergence are distinct outcomes.
Example MCP call
{
"name": "propagate_uncertainty",
"arguments": {
"expression": "x1 * x2",
"variables": [
{
"name": "x1",
"value": 1,
"distribution": "normal",
"standard_uncertainty": 1,
"unit": ""
},
{
"name": "x2",
"value": 1,
"distribution": "normal",
"standard_uncertainty": 1,
"unit": ""
}
],
"coverage": 0.95,
"samples": 200000,
"seed": 7
}
}Send this tool name and arguments through a connected MCP client. Discover the authoritative input schema with tools/list.
Call this engine in one command
Node.js 20+ and npm. This downloads a small remote MCP client; it does not install the engine locally.
npx --yes --package https://scorecompute.com/downloads/scorecompute-call-0.14.2.tgz scorecompute-call propagate_uncertainty --arguments '{"expression":"x1 * x2","variables":[{"name":"x1","value":1,"distribution":"normal","standard_uncertainty":1,"unit":""},{"name":"x2","value":1,"distribution":"normal","standard_uncertainty":1,"unit":""}],"coverage":0.95,"samples":200000,"seed":7}'Client download, PowerShell input and verification →
Direct MCP access
propagate_uncertainty is independently callable. It is deliberately not connected to the automatic planner, so nothing passes its results to another engine for you. Two tools returning similar JSON fields does not mean their results are compatible.
Cite this engine
Version 0.1.0, source SHA-256 5ee2d2b93b35…. Evidence recorded on release 0.16.1; the public service currently runs 0.17.1. Replace the access date when you cite it.
ScoreCompute. Measurement uncertainty with checked units (propagate_uncertainty), version 0.1.0, source SHA-256 5ee2d2b93b358d02768e0e7623061bea03fbc6448595f2fa99bda0c7bcedeebd. Evidence release 0.16.1, service release 0.17.1. https://scorecompute.com/engines/propagate-uncertainty/ (accessed YYYY-MM-DD).
All engine citations · Evidence card
Execution and availability
ScoreCompute exposes this tool through MCP Streamable HTTP. Rust computation runs on a connected worker; the public website and MCP gateway run separately. Among these 22 scientific tools, CUDA is implemented for simulate_pi; the other scientific engines run on CPU. The separate contributor network runs two compiled public workloads: a π estimate and fixed synthetic schedule probes. Requests are bounded and concurrent work may be refused when capacity is occupied.
Record inputs, assumptions and returned provenance when sharing a result. The public observatory displays software client names and tool activity, without publishing calculation arguments or results.