Measured, with receipts

Source: docs/product/claims.md, “Performance-evidence claims” ↗. The main receipt on this page is class 3: AWS i7i.metal-24xl bare metal.

1. SQL point workload — 1M rows

Read checksums matched across all three engines (200,000 hits each). The prepared point read is 40% faster than either earlier receipt. SQLite's point read is still 11× faster — published here at the same weight as every number Hyphae leads on.

2. Keyspace point workload — 1M keys

The embedded GET is 3.6× faster than Redis over its UDS round trip at p50 (2.2 µs against 7.9 µs); at p99 (16.3 µs) it is the slower one — the pool-miss tail, printed as-is. Strict and Memory SETs are unchanged: root publication plus two fsyncs, as before.

3. Lexical BM25 — 100k documents

Hit totals matched (100,000 vs 100,000). Against the 2.2.0-era receipt the query is 37× faster and durable ingest 15× faster. Tantivy still leads by 1.4× on query latency (was 50×) and 41× on ingest (was 580×); the p99 gap closed from 10× to 2.2×.

4. Durability ablations and the delta-vs-materialized transaction

Hyphae-only, identical single-SET commits on the materialized path, 10,000 commits per phase:

The regression published in the 2ff8a4b receipt is closed on the same host class: the materialized single-SET commit is back at the 2.2.0 level and slightly under it. The receipt clocks themselves were unchanged across all three runs — the roughly 31 ms that lived outside the clocks was the complete-state load reading a tree of one-entry leaves.

Transaction shape (1 INSERT + 1 SET + 1 indexed document, Memory durability, 2,000 commits), by engine composition:

The delta-staged transaction across all three engines is roughly 45× cheaper than the materialized path at this shape. Standalone group-commit smoke (256 commits, 8 producers, delta path): strict p50 1.54 ms at 587/s; group p50 3.14 ms at 2,263/s — 3.86× strict throughput.

5. Delta-transaction scaling sweep

CPU 0, median of three runs. Version depth (prior committed versions of the same keys):

Unrelated items sharing the engine, per commit:

Version depth stays flat — 193 to 202 µs across three orders of magnitude of history. The population sweep is 16% and 15% cheaper than on 2ff8a4b at 256 and 4,096 items (805 → 676 µs, 1,045 → 893 µs), with fewer page appends per commit.

6. Document-cap ladder — 250k and 1M

Fresh loads on NVMe, ladders on reopened directories (p50, 16 samples, limit=10, candidate limit 1,000); the 1M rung ran with the collection bound lifted on this host only.

The 1M query ladder is now linear in document count (3.3–4.5× for 4× the documents, against 6.6–8.3× on 2ff8a4b); at 250k nothing moved, since the 1,024-frame pool already held a 250k query's segments. The first bm25 sample after a fresh load is still a cold-cache outlier (p95 17 s at 1M, 3.8 s at 250k), reported and not excluded. The collection bound stays 250,000; the 1M rows measure above the bound.

Buffer-pool bound, same host, same 1M directory:

2.3× on the scorer from residency alone. The BLAKE3/CRC32C page-verification cost that was 44% of samples on 2ff8a4b is absent from the top 40 of the perf profile at 8,192 frames.

7. Scorer equivalence at 1M

One oracle round on the 1M directory of §6, after the ladder:

stage=durable_scorer round=0 hits=1000 ms=39.5 terms=2 segments=1562 physical_entries=372418
stage=model_scorer   round=0 hits=1000 ms=12711486.6
stage=scorer_equivalence hits=1000 bit_identical=true

The retained-model scorer reproduces the durable scorer's 1,000 ranked hits — ids, order, and score bits — at 1,000,000 documents, on a tree whose physical leaf layout the B+tree fix changed. The model took 3.5 hours; the durable scorer took 39.5 ms cold and 22 ms warm. This is the fourth rung (100k, 250k, 1M ×2) on which the two scorers agree bit for bit.

Scope and non-claims

• Single host, one run per phase (three for the delta sweeps); same host class as the two earlier receipts but a different physical machine each time — deltas under ~10% are not claimed.
• Shipped defaults throughout: 8,192-frame pool, no governor or execution pool (workers=1), --scale full baselines at their documented durable defaults.
• The collection bound stays 250,000; the 1M rows measure above the bound and the R5 vector conditions remain unmeasured.
• Synthetic corpora; no relevance claim.
• No universal superiority claim. On this host Hyphae's point reads and lexical queries are within 1.4–11× of the specialized baselines and its embedded GET is faster than Redis over UDS at p50, while durable writes remain the honest cost of copy-on-write publication plus two fsyncs. The differentiator is unchanged: one CSN across three engines with a model-checked commit protocol and receipts for every number.

Full receipt: docs/gates/evidence/hyphae-3.0-metal-a443c52-2026-09-03.md ↗. Raw harness JSON: sql ↗, keyspace ↗, lexical ↗, ablation ↗, group-commit ↗.