A modern, high-performance lexical analysis and parsing system with comprehensive PCRE2 support and CognitiveGraph integration
Recorded: 2026-09-24 · schema ml-baseline/1
Pre-ML baseline of the deterministic engine over the
Minotaur-Grammars example
corpora — one record per (grammar, Examples.txt) pair, with latency
percentiles (p50/p95/p99/min) and allocation profile (allocated bytes
p50/p95/p99/min) per parse, measured over 30 runs after 2 warm-up runs,
fresh StepParserEngine per run.
This is the “today” measurement the ML-assisted prototypes (sub-issues #74 learned rule prioritization, #75 learned GLR path pruning) must be evaluated against: no semantic changes and no allocation regressions vs. these numbers (issue #58 definition of done).
baseline-2026-09-24.csv — 111 records, one per (grammar, examples) pair.
Comment header records the environment; columns:
grammar, inputLengthBytes, success, errorCount, tokenCount, pathCount,
ambiguousParseCount, parseTimeP50Micros, parseTimeP95Micros,
parseTimeP99Micros, parseTimeMinMicros, allocatedBytesP50,
allocatedBytesP95, allocatedBytesP99, allocatedBytesMin.| Field | Value |
|---|---|
| Runtime | .NET 8.0.31 (Release build) |
| OS | Linux x64 |
| Processor count | 1 |
| Corpus | Minotaur-Grammars main @ 2026-09-24 |
Timings are machine-dependent and intended for relative, same-machine comparison only; re-measure on your machine with the command below before comparing prototypes against this file.
From a checkout of this repo (with a Minotaur-Grammars checkout):
MINOTAUR_GRAMMARS_DIR=/path/to/Minotaur-Grammars \
ML_TRACE_OUT=/tmp/ml-baseline \
dotnet test -c Release --filter "FullyQualifiedName~MlBaseline_RecordsCorpus"
This writes the raw ml-baseline/1 JSONL (same numbers, plus per-record
environment fields). The always-on smoke test
MlBaseline_HarnessMeasuresKnownGrammarParse covers the
percentile/allocation math in CI without the corpus.