Choose an agent pattern
from measured runs.

Compare identical workloads across source-named patterns. Reliability is a floor; latency or cost is the objective. No opaque score decides for you.

Benchmark run CSV

Observed success floor, %

Minimize after the floor

Load the sample or paste run evidence. Analysis stays on this device.

One decision, inspectable evidence

Match the work

Every candidate must cover the same workload IDs. Unequal evidence is blocked before comparison.

Keep uncertainty

Observed success stays beside its 95% Wilson interval, latency distribution, cost, and retries.

Declare the rule

Clear the reliability floor first. Then minimize one named objective without a weighted score.

The source supplies a pattern catalog, not benchmark outcomes. This lens compares only the rows you provide and does not verify author experience, hidden reasoning, or live agent behavior.

Selected by evidence

No result yet

Re-run same rows:

Observed success / 95% Wilson--

Median / p90 latency--

Mean cost--

Mean retries--

Latency x cost tradeoff--

clears floorbelow floorcircle size is observed success

Pattern evidence

PatternSuccessMedianMean cost

Decision rule

Boundary

Export has not been created.

Pattern benchmark lens: matched evidence, declared floor and objective

Your browser does not support video.

Read the explanation

This page is designed to compare supplied CSV measurements across agent patterns. It asks for matching workload IDs before comparing candidates, then applies an observed success floor. Consider an explicitly illustrative example: pattern A succeeds in five of six runs, or eighty-three point three percent, and pattern B succeeds in all six. Both clear an eighty percent floor; only B clears one hundred percent. The bars use six pixels per percentage point. This is an observed fraction, not a guarantee that the next run will succeed. The page displays a ninety-five percent Wilson interval alongside success and says those bounds do not silently replace the declared floor. The saved engine file is absent here, so this narration does not claim to execute its coverage checks or compute its statistical implementation. After reliability admission, the user chooses median latency or mean cost as one named objective. Imagine two eligible patterns with medians two and three seconds. A wins the latency comparison, shown at one hundred twenty pixels per second. Now imagine their mean costs are four cents and two cents: B wins the cost comparison. Neither dominates both objectives, so a blended score would introduce a preference the user did not declare. The visible result text says candidates are ordered by the selected objective, then higher observed success. These numbers are teaching examples, not the missing worked sample or benchmark outcomes. The source catalog supplies pattern names; the measurement rows, workload comparability and experiment design must supply the evidence. The saved page references Papa Parse, simple statistics, D3, GSAP and a local engine script. Those files are missing from this saved site directory. Its runtime assertion therefore reports that a pinned local runtime failed to load. Load sample also depends on the absent engine, so the advertised twenty-four rows cannot be inspected or treated as measured evidence here. Native objective controls and reset can still be checked against the original page, but there is no computed winner to export. The zero lower bar denotes no verified computed result, not zero latency, zero cost or perfect reliability. This local video explains the declared workflow while preserving the source and disclosing the missing dependencies. It sends no model, HTTP or MCP request and makes no public completion claim.

Choose an agent patternfrom measured runs.