strawmANN benchmark report

Measured as-deployed: the scheduler was left as it ships, so a difference is what a user would see rather than the engine in isolation, and ambient load is part of the measurement. Qdrant ran equal-work: asked for one populated graph (--segments 1), as strawmANN serves, so ef means the same thing on both sides. This is the configuration §8's comparative licensing is built around, and it is a control rather than a deployment. Its optimizer treats the count as a target; what it actually held is read back from the engine and stated under Collection config.

Conformance tier reached: T2 tie-aware rank. T1 (every BM25 score within its tolerance of the f64 oracle) and T2 (tie-aware rank) passed on both engines: the same documents at the same scores, so a throughput comparison is licensed (§8). Conformance hash 5b1e22a97fb23473.

Every BM25 score strawmANN returned is within 2.249e-07 (relative) of the f64 oracle's, against T1's tolerance of 1e-05: the same documents, the same statistics and the same ranking, in f32 (§8.1).

Every figure is the median of 3 passes per row per engine, alternated A/B/A/B (§7.4), and the spread is this run's own over 5 of 6 rows. The ingest and index-build rows (W15-upload) have none, so no build time carries a verdict. Each pass rebuilds the index, so the bands include build variance and are wider than a floor measured against one standing graph.

No Qdrant developer has used this tool. Nothing here has been run, checked against something already known, or disagreed with by anyone outside the project: every number has one author and one reviewer, and they are the same person. A result you can refute is more useful to us than one you accept.

It has never been pointed at a real Qdrant regression. The harness detects a known ISA slowdown and correctly reports no difference on a control row. That is internal consistency, not evidence it would flag a regression in your tree or stay quiet through a refactor. bench/harness/qdrant_ab.py exists to run two of your commits through it blind, and that experiment has not been done.

Part of the measured throughput gap is kernel width, not architecture. Read from a dev checkout: Qdrant's distance path tops out at four 256-bit accumulators and has no AVX-512, while these kernels use 512-bit ones on a machine that has them. That is a real difference and it is not the same claim as "the design is faster".

The engines are not equivalent, by construction. §1 removes sharding, replication, consensus, snapshots, sparse vectors, multivectors and disk-resident operation — permanently. Payload storage and filtering were in that list until 2026-09-03 and are now built (M7), so W12 measures a filtered search with a keyword index on both engines. Anything still unbuilt answers UNIMPLEMENTED naming the construct and reports n/a, never a silent degradation. A strawman that was not faster would mean it was badly built; the question is by how much, and where the model was wrong.

Everything below runs from a clone. The engine has no dependencies; the analysis path is a uv project pinned by bench/uv.lock.

scripts/doctor.py                    # what can this machine measure?
bench/setup.py check                 # §7.1 host gate: governor, boost, isolation, idle
conformance/datasets/datasets.py fetch

# one engine at a time, §7.1
zig build -Doptimize=ReleaseFast
# --connections matters: W4 opens ~32 sockets (-t 16 -c 2) and a server with
# fewer closes the excess before the HTTP/2 preface, which the client reports
# only as "transport error".
./zig-out/bin/strawmann --port 6334 --connections 64
bench/harness/workloads.py run http://localhost:6334 strawmann --sink --report

# the other half of every number: recall, and the licence to compare at all
cd conformance
cargo run --release -- relevance --engine http://localhost:6334 --label strawmann \
  --base $DATA/<base>.fbin --queries $DATA/<queries>.fbin \
  --ground-truth $DATA/gt/<ground truth>.json --metric <metric> \
  --ef 32 64 128 256 512 --json ../bench/results/strawmann/recall.json
cargo run --release -- differ --strawmann http://localhost:6334 --qdrant http://localhost:6434 \
  --base $DATA/<base>.fbin --queries $DATA/<queries>.fbin \
  --ground-truth $DATA/gt/<ground truth>.json \
  --json ../bench/results/strawmann/conformance.json

uv run --project bench bench/harness/report.py strawmann qdrant

Disagreements are the point. docs/spec.md is what all of this is measured against, docs/bugs.md lists the measurement bugs found so far (each produced a plausible wrong number rather than a failure), and docs/validation.md records what has and has not been checked.