Every performance and cost claim on this site should come with a receipt. This page is that receipt drawer: each measurement below is a real run, dated, with the method written out and — where you can run it yourself — the exact command. When we measure something new, it lands here.
The rule we hold ourselves to: we never round a measured figure into a nicer-sounding one, and we say plainly what a number does not cover.
Security · latency
Shield adds ~0.7 µs per turn
~0.7 µsmedian added per turn (observe mode)
Shield is the gateway's outbound security layer — it scans the prompt leaving for the model, and the model's streamed output coming back, for vendor secrets (API keys, tokens, private keys). The question a security layer always has to answer: how much does it slow every request down? We measured it directly.
Median added per turn (observe)
~0.7 µs
0.0007 ms — the mode every tenant runs today
Worst case (enforce, a secret redacted)
~2.0 µs
0.0020 ms, when a match is actually rewritten
For comparison — an LLM-as-guard call
5,600–11,000 ms
why we use purpose-built regex, not a model, on this path
How we measured it
Timing of the exact production scan function (scan_messages) compiled in release mode (LTO on, as production runs), over 100,000 iterations on a real 4-message turn. To smooth the noise of a single run on a busy host, we take the median across 5 independent bench runs — a median of medians. A live turn's ONLY extra work when Shield is on is this one scan before the request is forwarded; everything else (routing lookup, the forward, the streamed relay) is byte-identical whether Shield is on or off. So the scan-on-minus-scan-off delta IS the added latency. No proxy and no network are involved on purpose — a proxy round-trip would measure the provider's network, not Shield.
Reproduce it
cargo test --release shield::bench -- --ignored --nocapture
In the switchyard-gw crate. Prints median/mean/p99 for the benign-turn scan and the enforce-with-redaction case. This page is auto-refreshed from that same run by scripts/publish-shield-receipt.mjs.
What this number does not claim: Measured on the Mac Mini (Apple Silicon) that runs the gateway. Absolute nanoseconds vary run-to-run and with hardware; the point is the ORDER OF MAGNITUDE — microseconds, ~4–5 orders below a purpose-built ML guard and utterly below a provider round-trip (hundreds of ms). Shield today detects vendor SECRETS only (not PII), is observe-by-default (audits, doesn’t block, unless a tenant opts into enforce), and covers the gateway’s outbound prompt and its streamed output.
Routing · cost
7–10× cheaper than a popular router's autopilot, at equal-or-better quality
7–10×cheaper at equal-or-better quality
A head-to-head bake-off of our measured routing against a popular router's automatic mode, on the same real questions with the same quality bar. Ours came back 7–10× cheaper without losing resolution quality.
Cost vs. the router's autopilot mode
7–10× cheaper
at equal or better measured quality
Chat platforms' recommended tiers
5–27× our cost
for the same conversations
How we measured it
Same task set, same inputs, same quality judge — only the routing decision varies. Quality is graded by an LLM judge that is itself self-tested before the run is trusted. Full methodology and the per-model breakdown live on the Router Comparison page.
Where to see it: Detailed breakdown and methodology: /router-comparison.
What this number does not claim: Measured on our workload. Your mix of question types determines your own number — which is why we offer a one-time benchmark of your actual traffic before anyone wires anything in.
Serving · cost
$0.56 blended cost per 1,000 answered chatbot turns
$0.56per 1,000 answered turns (blended)
The routing engine behind Switchyard runs on a real production support chatbot. This is its blended model-serving cost across live traffic — not a rate card, the actual measured spend divided by the turns it answered.
Blended serving cost
$0.56
per 1,000 answered turns, live production traffic
How we measured it
Actual model spend on real answered turns over the measurement window, divided by the number of turns. Read from the same live serving logs the client's savings receipt is built from.
Where to see it: Hosted-assistant clients see the live version of this figure in their portal.
What this number does not claim: Blended across whatever models the traffic actually routed to; a different traffic mix yields a different blended figure.
Part of EyesInAI OS
These receipts back the products they belong to: Switchyard (routing + the Shieldyard security layer), the router comparison, and the savings receipt hosted-assistant clients see in their portal. If a claim on this site doesn't have a receipt here yet, ask us for it — the honest answer is either a link or "not measured yet."
Where a claim came from a head-to-head run — same task, only the model changes — the bake-offs are collected here, each with what it found and what it does not prove.