Benchmarks

This server is a thin shell over pacs008. The library is benchmarked in its own repository, so re-measuring schema validation here would only measure that again. What belongs here is the shape an agent experiences: which tool costs what, and how that changes with batch size.

Running it

python benches/bench_tool_dispatch.py           # full run
python benches/bench_tool_dispatch.py --quick   # what CI runs
python benches/bench_tool_dispatch.py --json    # machine-readable

CI runs --quick — not as a timing gate, but so a benchmark that has stopped compiling against the current API fails the build rather than rotting into a file that reads as verified and is not.

Single-call tools

tool	cost
list_message_types	~4.5 µs
get_required_fields	~171 µs
classify_address	~0.5 µs
validate_address	~0.6 µs
repair_address	~5.8 µs

list_message_types touches no records, so its cost is the dispatch floor — what every call pays before doing any work. The address tools sit near it because they are string work with no schema behind them.

The payment pipeline

records	validate ms	µs/record	generate ms	v/g	output bytes
1	2.09	2,090	0.87	2.4×	1,067
10	19.10	1,911	2.71	7.0×	6,899
100	187.13	1,871	24.94	7.5×	65,219

Two things to read here.

Output grows with the batch. Every record is rendered, not just the first. That is worth stating because the sibling acmt server behaves differently — most acmt message types describe a single account and silently ignore the rest of the batch. Here they do not.

Validation costs roughly seven times generation, and the growth exponent is 0.98 — linear. An agent that validates, fixes, and re-validates pays for the expensive half twice; batching the fixes before re-validating is worth doing.

Address screening: the November 2026 path

addresses	ms	µs/address
1	0.01	12.21
100	1.08	10.83
1,000	11.97	11.97
10,000	120.76	12.08

Flat, and about 155× cheaper per item than record validation.

That matters because the structured-address cutover is a screening problem, not a per-payment one: the question is “how many of our stored addresses will fail”, asked once across everything. At ~12 µs each, a million addresses is about twelve seconds — something you run while waiting, not a job somebody has to schedule.

Not measured

verify_bic_online makes a network call. Timing somebody else’s DNS and TLS handshake would say nothing about this code.