> ## Documentation Index
> Fetch the complete documentation index at: https://docs.iqrush.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Preserve decision-grade measurement

> Use stability and uncertainty to decide how much weight a measurement can carry. A precise-looking percentage is not enough on its own.

Use stability and uncertainty to decide how much weight a measurement can carry. A precise-looking percentage is not enough on its own.

## Check stability after selecting the run

Call `get_results_stability` with the same project, run, and supported filters as your analysis. The default aggregation is `domain`.

| Response field | Plain meaning |
| - | - |
| `data.rank_corr_ok` | The rank ordering has settled under the service's stability check. |
| `data.sufficiency_ok` | The service's signal sufficiency check has passed. |
| `data.is_converged` | Both required checks have passed. |
| `data.structural_snr` | Signal-to-noise ratio for the reported aggregation. |
| `data.structural_snr_grain` | Whether the SNR describes domains or URLs. |
| `data.target_sufficiency_ratio` | The returned sufficiency target. The target in this edition is 1.00. |
| `data.results_stability.value` | The stability result's SNR value; it is not a rank correlation. |

The signal-to-noise ratio is key to interpreting whether the measurement has enough separation to support a decision. Below the target, differences can be too small relative to uncertainty. Reaching the target alone is not sufficient: the ranking check must also pass.

Use the returned flags. Do not replace them with your own fixed rank-percentage threshold or treat “95% confidence” as a 95% stability requirement. Domain and URL SNR describe different scopes; never compare one against the other as if they were the same number.

## What to do with each result

| Result | Recommended report treatment |
| - | - |
| Converged, usable intervals, correct scope | Describe the measurement as decision-grade for that scope, while retaining uncertainty. |
| Rank check passes, sufficiency does not | Explain that the ordering has settled but separation is still too uncertain. |
| Rank check does not pass | Explain that ranking is still unsettled. |
| Stability or interval unavailable | State the limitation; do not add a “stable” badge. |

Collecting too little data can leave you acting on noise. Once the relevant checks pass and the intervals are adequate for the decision, collecting more of the same data may add limited value and consume resources. These read tools do not automatically stop or launch a measurement, and no single sample count guarantees convergence.

## Keep row quality and run quality separate

An individual estimate can have `stability: "stable"`, `"danger_zone"`, or `"insufficient"`. These are row-level indicators. They do not replace the run's convergence result. Mark limited rows clearly and avoid presenting insufficient rows as findings.

A decision-grade run still contains uncertain estimates and does not make an assistant's interpretation infallible. For change claims, use the confidence interval on the difference and the returned comparison result. See [compare measurements](/developers/guides/11-compare-measurements).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.