On-chain Forensics · operations

Status, latency and versioning

Infrastructure is judged on numbers, so here are ours with the sample size and the timestamp attached. Nothing on this page is a marketing figure.

Last measurement

Friday November 6, 2026, 19:47 CT · single client, sequential requests, no connection reuse · 115 samples across 5 routes

Route	Samples	p50	p95	Slowest	Notes
`verdictFast` cached verdict	12	385 ms	466 ms	470 ms	Primary cache hit path
`verdictFast` second token	5	412 ms	455 ms	460 ms	Same shape, different cache row
`forensicsGateway` informational	10	280 ms	340 ms	360 ms	No chain reads
`forensicsMints` challenge	10	435 ms	737 ms	901 ms	Paywall only, no data returned
`alertsSubscribe` challenge	10	257 ms	320 ms	340 ms	Paywall only, no data returned

Read this honestly: the cached routes answer in about 0.4 seconds, which is fine for a pre-trade check and not fine for a per-swap hook. Live scan routes that walk fresh chain state take seconds, not milliseconds. If you need us inside a hot path with a hard deadline, use verdictFast and treat anything else as a background or batch call.

Verify it yourself

for i in $(seq 1 20); do
  curl -s -o /dev/null -w "%{time_total}\n" \
    "https://api.onchainforensics.dev/functions/verdictFast?address=0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913&chain=base"
done

Latency depends on where you call from, so measure from your own egress rather than trusting ours.

Targets we hold ourselves to

Property	Target	Current
Cached verdict p95	under 800 ms	466 ms measured
Paywall challenge p95	under 900 ms	737 ms worst route measured
Cache refresh	every 5 minutes	every 5 minutes, 429 tokens and climbing
Read source	dedicated node, public fallback	private Alchemy endpoint, public rotation behind it
Truthfulness of a read	never turn an unreadable chain into a zero	unreadable sections return `PARTIAL` or `UNAVAILABLE`

No formal SLA contract is offered today. If you need one for a production integration, say so and we will put terms in writing before you build on us.

Versioning and deprecation

Every response and the OpenAPI document carry a semantic version. What that means in practice:

Patch and minor versions only add fields. Existing field names, types and enum members never change meaning inside a major version.
Breaking changes ship under a new route name (for example forensicsVerdict2) with the old route kept live for at least 90 days after the announcement.
Deprecation notice is published here and in the OpenAPI document at least 90 days before a route stops answering, and the old route returns a Deprecation header once the notice is live.
Removal requires the notice period to have elapsed and is never silent. A route that 404s without notice is our bug, not your problem to absorb.
Cache rebuilds are not API changes. Grades can change because the chain changed. A grade change is data, not a version bump.

Install inside an agent

The same routes ship as an MCP server so a framework can load them without code:

npx -y onchain-forensics-mcp

Six tools, stdio transport, zero dependencies. On a spent free allowance the tool returns the decoded x402 challenge instead of an error, so the calling agent knows the exact price.