Evaluation

The public synthetic workloads were rerun on v0.30.0, commit 7d81314, on 2026-10-30. The Apple M4 / Node 24.21.0 machine met a quiet-start gate; it remained shared and load was sampled during the longer stages. See the benchmark report for methodology, load samples, raw results, charts and reproduction commands.

Retrieval quality

Suite	Probes	Target-record hit@1	Target-record hit@5
Synthetic card cues and body facts	80	100%	100%
Demo direct search	16	100%	100%
Demo auto-recall query	16	100%	100% (only 3 results requested)

These are synthetic search tests, not end-to-end answer tests. The synthetic suite generates identifiers with Math.random(); its results are reported in the raw output. The demo suite’s cold recall field can credit a hot hit for a probe whose acceptable set also contains a cold record, so it is not evidence of cold-record retrieval.

A previous run against a private, curated 40-question archive reported hit@1 73%, hit@3/hit@5 95%, cold recall 100%, and search p50 1.3 ms; a prior baseline was reported as 67% hit@5. That archive and its raw probes are not public.

Reported v0.5.1 re-check (same private 40-question archive): hit@1 68%, hit@3 90%, hit@5 93%, cold recall 94%, p50 1.5 ms. These private-archive numbers cannot be independently reproduced from the public fixture. The separate reported DeepSeek multi-seed results have a retained-tail methodology caveat.

Capacity and storage

Synthetic pool	Write	Cold spill	Cold GC
1,000 records / 10M assigned tokens	1.0 s	700 records / 7M in 1.07 s	350 records / 3.5M in 0.24 s
10,000 records / 100M assigned tokens	9.4 s	5,000 records / 50M in 14.45 s	2,500 records / 25M in 5.36 s

The 100M run used 384.8 MiB of original body files and a 10.7 MiB SQLite index after writing, growing to 28.3 MiB after cold spill. The earlier v0.14.0 run had a SQLite lock failure; see the fix and investigation. Token counts are assigned synthetic values. The x-only bodies do not measure identifier-body index overhead; numeric anchors in the card fields are indexed in v0.5.0. Cold GC deletes the removed records. The read-path timings and raw output are included in the benchmark report.

An older private 1.9M-token archive was reported to shrink from 18.19 MB to 0.25 MB after the body-less FTS migration and VACUUM. That migration was not rerun here; it should be treated as a historical observation rather than a current public-fixture measurement.

Measurement limits

• The long-horizon and scale runs use LocalStoreCore with default rerank weights and synthetic fixtures; they do not measure model-specific profiles, MCP/DSH integration, or an agent answering a task.
• The capacity helper’s printed p95 is based on five samples and selects the second-largest value. It is not a reliable 95th percentile. This rerun does not establish that the former p95 < 50 ms target is met.
• The offline harness passed 131/131 tests at the tested commit after building the bundle; this validates its test cases, not production or real-agent performance.