Name, license, source, content characteristics, preprocessing, and a stable digest.
RESOLVING ROUTEFollowing the evidence boundary…
This lab publishes the evidence contract before the leaderboard. Results stay withheld until corpus, revision, machine, protocol, uncertainty, raw observations, and claim boundary can travel with every point.
A beautiful chart without stable inputs and downloadable evidence creates confidence faster than it creates knowledge. The public snapshot identifies the benchmark substrate, but not a normalized cross-system dataset.
Each gate is part of the result, not supplementary fine print. Missing information blocks publication or appears as an explicit, visible limitation.
Name, license, source, content characteristics, preprocessing, and a stable digest.
Exact query, train, validation, and test boundaries with leakage controls where relevant.
Repository commit, crate versions, features, compiler, lockfile, and benchmark harness revision.
CPU, memory, accelerator, operating system, power mode, and relevant runtime configuration.
Warmup, repetitions, concurrency, cache state, timeout, and failure-handling rules.
Aggregation, variability or uncertainty, exclusions, and the complete denominator.
Machine-readable observations, logs, errors, and receipt identifiers—not only a chart image.
What the measurement supports, what it does not support, and the exact fallback path.
These suites follow the public receipt-bench scope—semantic search, compression, and memory operations. They describe what the lab intends to measure, not completed performance claims.
Exact, lexical, vector, hybrid, and temporal-state paths evaluated against a disclosed corpus and query set.
Experimental codecs measured with wire format, profile, quality budget, resource ceiling, and exact/raw fallback in view.
Ingest, retrieve, supersede, forget, reconcile, and replay behavior evaluated as operations—not collapsed into a single speed score.
When normalized data is ready, visual polish will serve inspection: filters, uncertainty, accessible tables, and downloadable data—not ornamental certainty.
No dual axes, truncated context, or units that shift between series without warning.
Defaults are declared and defensible, with weaker or failed paths still discoverable.
Every chart has a readable data table and does not require color perception to decode.
CSV or JSON, configuration, raw observations, and receipt references ship beside the graphic.
The Proofroom demonstrates receipt anatomy with a deterministic fixture. The changelog reports public registry updates from a named snapshot. Neither is a substitute for normalized benchmark results.