diff --git a/eksml/BENCHMARKS.md b/eksml/BENCHMARKS.md
index c70d034..af4a98f 100644
--- a/eksml/BENCHMARKS.md
+++ b/eksml/BENCHMARKS.md
@@ -111,6 +111,14 @@ Chunked streaming parse where each parser tokenizes SAX events and builds a full
| 15,003 op/s |
13,719 op/s |
+
+ | @tuananh/sax-parser |
+ 24,238 op/s |
+ 19,238 op/s |
+ 12,849 op/s |
+ 18,062 op/s |
+ 11,681 op/s |
+
| easysax |
66,566 op/s |
@@ -127,14 +135,6 @@ Chunked streaming parse where each parser tokenizes SAX events and builds a full
5,186 op/s |
9,352 op/s |
-
- | saxes |
- 27,902 op/s |
- 5,884 op/s |
- 6,687 op/s |
- 6,013 op/s |
- 5,117 op/s |
-
| htmlparser2 |
26,433 op/s |
@@ -144,12 +144,12 @@ Chunked streaming parse where each parser tokenizes SAX events and builds a full
6,791 op/s |
- | @tuananh/sax-parser |
- 24,238 op/s |
- 19,238 op/s |
- 12,849 op/s |
- 18,062 op/s |
- 11,681 op/s |
+ saxes |
+ 27,902 op/s |
+ 5,884 op/s |
+ 6,687 op/s |
+ 6,013 op/s |
+ 5,117 op/s |
| sax |
@@ -206,6 +206,14 @@ Pure scanner throughput with no downstream work -- isolates the tokenizer's raw
11,824 op/s |
11,712 op/s |
+
+ | @tuananh/sax-parser |
+ 22,011 op/s |
+ 16,577 op/s |
+ 14,692 op/s |
+ 15,889 op/s |
+ 11,361 op/s |
+
| saxes |
35,117 op/s |
@@ -222,14 +230,6 @@ Pure scanner throughput with no downstream work -- isolates the tokenizer's raw
7,028 op/s |
7,186 op/s |
-
- | @tuananh/sax-parser |
- 22,011 op/s |
- 16,577 op/s |
- 14,692 op/s |
- 15,889 op/s |
- 11,361 op/s |
-
| sax |
14,798 op/s |
@@ -348,6 +348,7 @@ All fixtures are in [`test/fixtures/`](./test/fixtures/).
## Methodology
- **Tool**: [Vitest bench](https://vitest.dev/guide/features.html#benchmarking) (wraps [tinybench](https://github.com/tinylibs/tinybench))
+- **Table ordering**: Rows are ordered by the geometric mean of op/s across the table's fixtures (an arithmetic mean would be dominated by the smallest, fastest fixture). A library that cannot parse a fixture is ranked against its per-fixture results, not rewarded for the gap: tXml sits below libraries that beat it on every fixture it completes.
- **Warmup**: Default tinybench warmup iterations
- **Environment**: Single-threaded, synchronous execution on Node.js
- **Parser reuse**: In the SAX streaming and tokenization suites, every parser is constructed once and reused across iterations (all measured parsers support this; verified by comparing event streams across runs). This measures steady-state parse throughput rather than constructor cost. XmlParseStream is the exception, web streams are single-use, so it pays its constructor per iteration.