SLAUNT / BENCHMARKS

The only benchmark that
matters is your workspace.

We don't print numbers we can't stand behind. Slaunt ships with an evaluation harness that measures the context layer on your data — and the results belong to you.

1.0

What we measure

Four metric families, defined before they're flattering. The value column fills in when the harness runs on your workspace — not before.

Context recall

When an agent answers from workspace memory, does it retrieve the decision that was actually made — including the ones recorded months ago?

Method Held-out decisions are queried back through the same MCP surface agents use, and scored against the record.

measured on your workspace
Routing accuracy

Does work reach the right human or agent without a dispatcher touching it?

Method Routed outcomes are compared against post-hoc owner corrections in your own workspace.

measured on your workspace
Retrieval latency

How long does an agent wait on workspace memory before it can act?

Method Timed at the MCP boundary across your real event volume, not a demo dataset.

measured on your workspace
Grounding

How often do agents state things the workspace record doesn't support?

Method Sampled answers are traced back to source events; unsupported claims are counted against the run.

measured on your workspace
2.0

How the harness works

The evaluation runs against your workspace through the same interface agents use — so the number you get is the number you'll live with.

01Replay

Your workspace events are replayed into an isolated evaluation copy — same shape, same volume, same mess as production.

02Hold out

A slice of recorded decisions and outcomes is hidden from the context layer under test.

03Query

Agents query through the exact MCP surface they use in production. No benchmark-only fast path.

04Score

Answers are scored against the held-out record. The report lands in your workspace — and stays yours.

slaunt eval — harness run

$ slaunt eval --workspace your-org --window 90d

replaying workspace events into isolated eval copy …

holding out recorded decisions and outcomes …

querying through production MCP surface …

scoring answers against the held-out record …

done report written to your workspace — private to your org

3.0

Why there's no leaderboard here

Three reasons this page has dashes where other pages have percentages.

Workspaces aren't interchangeable

Recall on a 90-day-old design decision depends on how your team records decisions. A single global number would tell you about our demo workspace, not your rollout.

Synthetic leaderboards reward overfitting

It's easy to tune a retrieval stack to a public benchmark. It's harder to make it hold up on a workspace it has never seen. We optimize for the second thing.

Numbers without provenance erode trust

A percentage with no dataset, no method, and no way to reproduce it is marketing, not measurement. We'd rather show you the dash than show you that.

When we publish public numbers, they'll ship with the harness, the dataset, and the method — so you can reproduce them. Until then, the dash stays.

Benchmark it on your workspace

Run the evaluation harness on your own data during a pilot — and keep the report.