diff --git a/opteryx-skene/README.md b/opteryx-skene/README.md
new file mode 100644
index 0000000000..0bbda538e9
--- /dev/null
+++ b/opteryx-skene/README.md
@@ -0,0 +1,126 @@
+# Opteryx
+
+Opteryx is an in-process SQL query engine. Query **planning** (parse, bind,
+optimize) runs in Python; query **execution** is native (Cython/C++). It
+queries Parquet directly from storage with no preloading or preprocessing,
+which makes it well suited to ad hoc analytics.
+
+For more information, visit:
+
+- [Opteryx Documentation](https://docs.opteryx.app/)
+- [Opteryx GitHub Repository](https://github.com/mabel-dev/opteryx-core)
+
+This entry benchmarks Opteryx on **skene**, its native columnar storage
+format. It is the native-format counterpart to `Opteryx (Parquet, partitioned)`.
+
+---
+
+## Generating Benchmark Results
+
+### High-level Steps
+1. Set up the environment.
+2. Install Python and the required dependencies.
+3. Download the benchmark dataset.
+4. Convert it to skene (this is the load step).
+5. Run the benchmark script.
+
+### Detailed Instructions
+
+1. **Start an AWS EC2 instance**
+   - OS: Ubuntu 24
+   - Architecture: 64-bit (x86_64 or AArch64)
+   - Instance Type: `c6a.4xlarge`
+   - Root Storage: 500 GB gp2 SSD
+   - Advanced Details: ensure 'EBS-optimized instance' is **disabled**.
+
+2. **SSH into the instance** (after status checks complete):
+   ~~~bash
+   ssh ubuntu@{ip}
+   ~~~
+
+3. **Update the package list and install Git**
+   ~~~bash
+   sudo apt-get update -y
+   sudo apt-get install git -y
+   ~~~
+
+4. **Clone the ClickBench repository**
+   ~~~bash
+   git clone https://github.com/ClickHouse/ClickBench
+   cd ClickBench/opteryx-skene
+   ~~~
+
+5. **Run the benchmark script**
+   ~~~bash
+   sudo ./benchmark.sh
+   ~~~
+
+### Loading
+
+ClickBench distributes the dataset as Parquet, so this entry converts it to
+skene before querying. That conversion is the load step, and `Load time` is its
+wall-clock — comparable to any entry that ingests the source data into a native
+store, and unlike the Parquet entry, whose load performs no conversion at all.
+
+`convert.py` performs the conversion using the writer that ships inside the
+`opteryx-core` wheel. skene, draken and rugo are all packaged in that single
+wheel, so the conversion needs no additional dependency and no source checkout.
+Row groups are packed 16 per file at 262144 rows each, matching the engine's own
+mirrors: packing is per directory rather than per source file.
+
+The conversion runs across processes. A worker owns a contiguous range of input
+files end to end and writes its own output files, because morsels hold raw
+pointers and cannot cross a process boundary. Workers default to three quarters
+of the cores; `-j` overrides it.
+
+That parallelism has a cost in layout: row groups do not
+span chunk boundaries, so each worker's last row group — and last file — is
+short. The published mirror was built with 12 workers and holds **28 files**,
+against 24 for a single-worker build of the same data, and it measures about 7%
+slower across the 43 queries for that reason. `-j 1` reproduces the
+single-worker layout exactly. Row count is invariant either way: it is verified
+per chunk against the source footers and again on the total, and a mismatch
+fails the run rather than warning.
+
+The source Parquet is deleted once the conversion completes, so `data-size`
+measures the skene dataset alone.
+
+### Compression posture
+
+The mirror is written with the engine's read-first ("performance") posture:
+skene decodes substantially faster uncompressed or lz4-compressed than with
+per-section zstd, at the cost of more bytes on disk. It is a deliberate choice
+for locally attached storage, where the disk is not the bottleneck and decompression
+is pure cost. Remotely read data, where bytes dominate, is written differently.
+
+The ClickBench Parquet corpus is published pre-compressed, so
+`Opteryx (Parquet, partitioned)` reads whatever that corpus contains. In
+practice the two are close in size — the skene mirror is 15.39 GB against
+14.74 GB of Parquet, about 4% larger.
+
+### Python version
+
+`opteryx-core` publishes cp314 x86_64 and AArch64 manylinux wheels and declares
+no runtime dependencies, so `install` is a single binary-wheel download with no
+on-box compilation and no toolchain.
+
+### Query dialect
+
+`queries.sql` adapts queries to Opteryx's dialect. The adaptations are syntactic
+— they do not change what is computed, the rows returned, or the work the engine has to do:
+
+- **Q19, Q43** — `EventTime` is stored as an integer epoch, so it is cast
+  explicitly (`EventTime::TIMESTAMP[s]`) before `extract(minute FROM ...)` and
+  before truncation.
+- **Q43** — `TRUNC(<ts>, 'minute')` rather than `DATE_TRUNC('minute', <ts>)`.
+- **Q29** — the `REGEXP_REPLACE` pattern and replacement use `b''` and `r''`
+  literals so the backslash reference survives to the regex engine.
+- **Q37-Q42** — `EventDate` comparisons cast both sides to `DATE`
+  (`EventDate::DATE >= '2013-07-01'::DATE`).
+
+### Hardware coverage
+
+Results are published for instance types with **32 or fewer vCPUs**. The account
+used for these runs is limited to 32 concurrent on-demand vCPUs, so the 192-vCPU
+machines in the ClickBench fleet (`c6a.metal`, `c7a.metal-48xl`,
+`c8g.metal-48xl`) could not be launched.
diff --git a/opteryx-skene/benchmark.sh b/opteryx-skene/benchmark.sh
new file mode 100755
index 0000000000..b7ae16ae72
--- /dev/null
+++ b/opteryx-skene/benchmark.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+export BENCH_DOWNLOAD_SCRIPT="download-hits-parquet-partitioned"
+export BENCH_RESTARTABLE=no
+# Single-process engine: each query forks a fresh full-machine process with no
+# shared scheduler across connections, so the concurrent-QPS test only
+# oversubscribes RAM rather than measuring throughput. Skip it by default;
+# override BENCH_CONCURRENT_DURATION to re-enable. See issue #946.
+export BENCH_CONCURRENT_DURATION="${BENCH_CONCURRENT_DURATION:-0}"
+exec ../lib/benchmark-common.sh
diff --git a/opteryx-skene/check b/opteryx-skene/check
new file mode 100755
index 0000000000..ed305e8015
--- /dev/null
+++ b/opteryx-skene/check
@@ -0,0 +1,34 @@
+#!/bin/bash
+set -e
+
+# Smoke test that actually executes through the engine and propagates a
+# non-zero exit on failure. The old `python -m opteryx "SELECT version()"`
+# CLI path is unusable here for two reasons:
+#
+#   1. opteryx-core's single-run CLI calls the removed opteryx.query(),
+#      catches the AttributeError, and STILL exits 0 -- a false green that
+#      would let a broken install pass this check.
+#   2. `version()` is not a function in the current engine; the version is
+#      exposed as the system variable `@@version`.
+#
+# This matters more than it looks: bench_load() in lib/benchmark-common.sh
+# calls ./check unconditionally, AFTER the ~14 GB download and load. A check
+# that cannot pass burns the whole setup cost before failing.
+#
+# Assert the shape of the value rather than a literal, so the check does not
+# need editing on every release.
+"$HOME/opteryx_venv/bin/python" - <<'PY'
+import re
+
+import opteryx
+
+session = opteryx.session()
+morsels = list(session.execute_to_morsels("SELECT @@version"))
+session.close()
+
+assert morsels, "SELECT @@version returned no morsels"
+value = list(morsels[0][0])[0]
+if isinstance(value, (bytes, bytearray)):
+    value = value.decode()
+assert re.fullmatch(r"\d+\.\d+\.\d+.*", str(value)), f"unexpected version {value!r}"
+PY
diff --git a/opteryx-skene/convert.py b/opteryx-skene/convert.py
new file mode 100644
index 0000000000..ba6c13aea5
--- /dev/null
+++ b/opteryx-skene/convert.py
@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+"""
+Convert ClickBench's partitioned parquet `hits` into its skene mirror.
+
+Self-contained on purpose: skene, draken and rugo all ship inside the
+`opteryx-core` wheel, so this needs nothing but the benchmark's own install.
+(The engine repo's dev/parquet_to_skene.py is dev-only tooling and is not
+available to a ClickBench run. This file is a port of it; keep them in step.)
+
+Packing: 16 row groups per file at 262144 rows per row group, matching the
+engine's own mirrors. Packing is per DIRECTORY, not per file — a row group can
+span two source files, so output files do not correspond to input files.
+
+PARALLELISM: conversion runs across PROCESSES, not threads. Both hot phases
+release the GIL (rugo's decode, skene's add_row_group/write_to), but the morsel
+construction between them does not, which caps a thread pool at ~2.5x against
+~5x for processes. Morsels hold raw pointers and cannot cross a process
+boundary, so a worker owns a CONTIGUOUS RANGE OF INPUT FILES end to end and
+writes its own output files; only counts come back.
+
+CONSEQUENCE — the output is NOT byte-identical to a serial build. Row groups do
+not span chunk boundaries, so each worker's last row group is short, where
+serially only the very last one was. On this dataset that is ~16 short row
+groups of ~397 rather than 1. The layout is a function of the worker count;
+`-j 1` reproduces the serial layout exactly. ROW COUNT is invariant, is checked
+per chunk against the source footers, and a mismatch is a hard failure.
+
+Usage:  convert.py <src-dir> <dst-dir> [codec] [-j N]   codec: none|lz4|zstd
+"""
+
+import os
+import sys
+from concurrent.futures import ProcessPoolExecutor
+
+import skene
+from draken.morsels.morsel import Morsel
+from rugo.parquet import read_metadata
+from rugo.parquet import read_parquet
+
+ROWS_PER_ROW_GROUP = 262144
+ROW_GROUPS_PER_FILE = 16
+
+
+class Packer:
+    def __init__(self, out_dir, stem, codec, zstd_level, first_index=0):
+        self.out_dir = out_dir
+        self.stem = stem
+        self.codec = codec
+        self.zstd_level = zstd_level
+        self._writer = None
+        self._pending = []
+        self._pending_rows = 0
+        self._row_groups_in_file = 0
+        self.files = 0
+        self.rows = 0
+        self.nbytes = 0
+        # Each worker reserves a disjoint output index range, so two workers
+        # can never write the same filename.
+        self._first_index = first_index
+        os.makedirs(out_dir, exist_ok=True)
+
+    def add(self, morsel):
+        if morsel.num_rows == 0:
+            return
+        self._pending.append(morsel)
+        self._pending_rows += morsel.num_rows
+        while self._pending_rows >= ROWS_PER_ROW_GROUP:
+            merged = self._merge()
+            self._emit(merged.slice(0, ROWS_PER_ROW_GROUP))
+            remainder = merged.num_rows - ROWS_PER_ROW_GROUP
+            if remainder > 0:
+                self._pending = [merged.slice(ROWS_PER_ROW_GROUP, remainder)]
+            else:
+                self._pending = []
+            self._pending_rows = remainder
+
+    def _merge(self):
+        # Morsel has no `concat`; `combine` is the n-way merge. Getting this
+        # wrong loses rows silently, which a benchmark cannot survive.
+        return self._pending[0] if len(self._pending) == 1 else Morsel.combine(self._pending)
+
+    def close(self):
+        # The final row group is SHORT, and so is the final file - a dataset
+        # does not divide evenly and padding or dropping the tail are both wrong.
+        if self._pending_rows > 0:
+            self._emit(self._merge())
+            self._pending = []
+            self._pending_rows = 0
+        self._close_file()
+
+    def _emit(self, row_group):
+        if self._writer is None:
+            self._writer = skene.SkeneWriter(
+                read_acceleration=True, codec=self.codec, zstd_level=self.zstd_level
+            )
+            self._row_groups_in_file = 0
+        self._writer.add_row_group(row_group)
+        self._row_groups_in_file += 1
+        self.rows += row_group.num_rows
+        if self._row_groups_in_file >= ROW_GROUPS_PER_FILE:
+            self._close_file()
+
+    def _close_file(self):
+        if self._writer is None:
+            return
+        index = self._first_index + self.files
+        path = os.path.join(self.out_dir, f"{self.stem}-{index:04d}.skene")
+        # write_to() completes the file in place; finish() would double peak RSS
+        # on a wide schema for bytes nobody keeps.
+        self.nbytes += self._writer.write_to(path)
+        self.files += 1
+        self._writer = None
+        self._row_groups_in_file = 0
+
+
+def convert_dir(paths, out_dir, stem, codec, zstd_level, first_index=0):
+    packer = Packer(out_dir, stem, codec, zstd_level, first_index)
+    for p in paths:
+        with read_parquet(p) as reader:
+            for morsel in reader:
+                packer.add(morsel)
+    packer.close()
+    return packer.files, packer.rows, packer.nbytes
+
+
+def _default_workers():
+    """Three quarters of the cores, never all of them.
+
+    Saturating the cores measured ~50% SLOWER than the plateau: a worker holds
+    a whole decoded file plus up to ROW_GROUPS_PER_FILE row groups buffered in
+    its writer, so the run goes memory- and scheduler-bound.
+    """
+    return max(1, (os.cpu_count() or 1) * 3 // 4)
+
+
+def _output_files_for(rows):
+    """Exactly how many .skene files a chunk of `rows` rows produces — this is
+    what reserves each worker's output index range, so it must be exact."""
+    if rows == 0:
+        return 0
+    row_groups = -(-rows // ROWS_PER_ROW_GROUP)
+    return -(-row_groups // ROW_GROUPS_PER_FILE)
+
+
+def _plan_chunks(paths, row_counts, workers):
+    """Split into <= `workers` CONTIGUOUS chunks of whole files, balanced by ROW
+    COUNT rather than file count. Returns [(paths, rows)] in input order so
+    output names stay sequential."""
+    total = sum(row_counts)
+    # Never split into more chunks than there are full output files of rows:
+    # each chunk starts a new output file, so over-splitting a small table
+    # shatters it into undersized files with undersized row groups — precisely
+    # what packing 16 row groups per file exists to avoid.
+    rows_per_file = ROWS_PER_ROW_GROUP * ROW_GROUPS_PER_FILE
+    n = min(workers, len(paths), max(1, total // rows_per_file))
+    if n <= 1 or total == 0:
+        return [(list(paths), total)]
+
+    chunks = []
+    start = 0
+    assigned = 0
+    for _ in range(n - 1):
+        remaining_chunks = n - len(chunks) - 1
+        target = (total - assigned) / (remaining_chunks + 1)
+        rows = 0
+        end = start
+        while end < len(paths):
+            if len(paths) - (end + 1) < remaining_chunks:
+                break
+            rows += row_counts[end]
+            end += 1
+            if rows >= target:
+                break
+        if end == start:
+            break
+        chunks.append((paths[start:end], rows))
+        assigned += rows
+        start = end
+    if start < len(paths):
+        chunks.append((paths[start:], sum(row_counts[start:])))
+    return chunks
+
+
+def _convert_chunk(task):
+    """Process-pool entry point. Top-level and picklable-only arguments because
+    spawn-start platforms re-import this module in the child."""
+    paths, out_dir, stem, codec, zstd_level, first_index, expected_rows = task
+    files, rows, nbytes = convert_dir(paths, out_dir, stem, codec, zstd_level, first_index)
+    if rows != expected_rows:
+        # Catches the failure that destroyed an earlier converter quietly: a bad
+        # morsel merge dropped 76% of the rows and still produced a plausible,
+        # fast, completely wrong dataset.
+        raise RuntimeError(
+            f"{out_dir}: chunk at index {first_index} wrote {rows:,} rows "
+            f"but its sources hold {expected_rows:,}"
+        )
+    return files, rows, nbytes
+
+
+def main():
+    argv = [a for a in sys.argv[1:]]
+    workers = _default_workers()
+    out = []
+    i = 0
+    while i < len(argv):
+        if argv[i] in ("-j", "--workers"):
+            workers = int(argv[i + 1]); i += 2; continue
+        if argv[i].startswith("--workers="):
+            workers = int(argv[i].split("=", 1)[1]); i += 1; continue
+        out.append(argv[i]); i += 1
+    if len(out) not in (2, 3):
+        print(__doc__)
+        return 1
+    src, dst = out[0], out[1]
+    codec = out[2] if len(out) == 3 else "lz4"
+    if codec not in ("none", "lz4", "zstd"):
+        print(f"ERROR: unknown codec {codec!r}")
+        return 1
+    zstd_level = 9 if codec == "zstd" else 0
+
+    if not os.path.isdir(src):
+        print(f"ERROR: source not found: {src}")
+        return 1
+    stale = [f for f in os.listdir(dst) if f.endswith(".skene")] if os.path.isdir(dst) else []
+    if stale:
+        print(f"ERROR: {dst} already holds {len(stale)} .skene file(s); rm -rf it first")
+        return 1
+
+    paths = sorted(os.path.join(src, f) for f in os.listdir(src) if f.endswith(".parquet"))
+    if not paths:
+        print(f"ERROR: no parquet files in {src}")
+        return 1
+
+    # Row counts come from the source footers up front, so output index ranges
+    # can be reserved before any worker starts and no worker has to ask another
+    # where its files begin.
+    row_counts = [read_metadata(p).num_rows for p in paths]
+    expected_total = sum(row_counts)
+
+    tasks = []
+    next_index = 0
+    for chunk_paths, chunk_rows in _plan_chunks(paths, row_counts, workers):
+        tasks.append((chunk_paths, dst, "hits", codec, zstd_level, next_index, chunk_rows))
+        next_index += _output_files_for(chunk_rows)
+
+    os.makedirs(dst, exist_ok=True)
+    if len(tasks) == 1:
+        results = [_convert_chunk(tasks[0])]
+    else:
+        with ProcessPoolExecutor(max_workers=len(tasks)) as pool:
+            results = list(pool.map(_convert_chunk, tasks))
+
+    files = sum(r[0] for r in results)
+    rows = sum(r[1] for r in results)
+    nbytes = sum(r[2] for r in results)
+    if rows != expected_total:
+        raise RuntimeError(f"wrote {rows:,} rows, sources hold {expected_total:,}")
+    if files == 0:
+        raise RuntimeError("no row groups read - refusing to write an empty table")
+    print(f"codec={codec} workers={len(tasks)} files={files} rows={rows} bytes={nbytes}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/opteryx-skene/data-size b/opteryx-skene/data-size
new file mode 100755
index 0000000000..8e65ea4b35
--- /dev/null
+++ b/opteryx-skene/data-size
@@ -0,0 +1,4 @@
+#!/bin/bash
+set -e
+
+du -bcs hits | awk '/total$/ { print $1 }'
diff --git a/opteryx-skene/install b/opteryx-skene/install
new file mode 100755
index 0000000000..21ba004257
--- /dev/null
+++ b/opteryx-skene/install
@@ -0,0 +1,44 @@
+#!/bin/bash
+set -e
+
+# Opteryx (PyPI package: opteryx-core) on stock CPython 3.14.
+#
+# NOT the free-threaded build. Opteryx abandoned the free-threaded-Python
+# experiment in June 2026; the engine's parallelism target is native (C++)
+# threads under a released GIL, so the GIL build is the supported and the
+# representative configuration. opteryx-core stopped publishing cp314t wheels
+# after 0.9.16 accordingly -- installing onto a 3.14t interpreter would fall
+# through to the sdist and try to build Rust/C++ on the box.
+#
+# opteryx-core publishes cp314 x86_64 manylinux wheels and declares no runtime
+# dependencies, so this is a single binary-wheel download -- no on-box
+# compilation and no toolchain. x86_64 only, which is fine: the canonical
+# c6a.4xlarge is x86_64.
+
+sudo apt-get update -y
+sudo apt-get install -y software-properties-common
+sudo add-apt-repository -y ppa:deadsnakes/ppa
+sudo apt-get update -y
+
+# Ubuntu 24.04 (noble) ships 3.12; 3.14 comes from deadsnakes, which carries
+# python3.14 for noble. python3.14-venv provides the venv module.
+sudo apt-get install -y python3.14 python3.14-venv git wget
+
+if [ ! -d "$HOME/opteryx_venv" ]; then
+    python3.14 -m venv "$HOME/opteryx_venv"
+fi
+
+"$HOME/opteryx_venv/bin/python" -m pip install --upgrade pip
+# Pulls the cp314 wheel for the latest release.
+"$HOME/opteryx_venv/bin/python" -m pip install --upgrade opteryx-core
+
+# Fail loudly here rather than 43 queries later if pip silently fell back to a
+# source build or resolved an interpreter we did not expect.
+"$HOME/opteryx_venv/bin/python" - <<'PY'
+import sys
+
+import opteryx
+
+print(f"python  {sys.version.split()[0]}  (GIL enabled: {sys._is_gil_enabled()})")
+print(f"opteryx-core {opteryx.__version__}")
+PY
diff --git a/opteryx-skene/load b/opteryx-skene/load
new file mode 100755
index 0000000000..f714884db7
--- /dev/null
+++ b/opteryx-skene/load
@@ -0,0 +1,19 @@
+#!/bin/bash
+# ClickBench ships parquet; skene is Opteryx's native format, so the dataset has
+# to be converted before it can be queried. That conversion IS the load step,
+# and its wall-clock is what `Load time` reports -- the same shape as any entry
+# that ingests the source data into a native store.
+set -e
+
+mkdir -p parquet_src
+mv hits_*.parquet parquet_src/ 2>/dev/null || true
+
+# lz4 is the engine's read-first ("performance") posture. See README.md: it is
+# deliberately NOT the codec the published parquet corpus uses, so this entry
+# and `Opteryx (Parquet, partitioned)` differ in codec as well as in format.
+"$HOME/opteryx_venv/bin/python" convert.py parquet_src hits lz4
+
+# Drop the source once converted: `data-size` must measure the skene dataset,
+# not skene plus a parquet copy, and 500 GB does not need to hold both.
+rm -rf parquet_src
+sync
diff --git a/opteryx-skene/queries.sql b/opteryx-skene/queries.sql
new file mode 100644
index 0000000000..906e5cd6ec
--- /dev/null
+++ b/opteryx-skene/queries.sql
@@ -0,0 +1,43 @@
+SELECT COUNT(*) FROM hits;
+SELECT COUNT(*) FROM hits WHERE AdvEngineID <> 0;
+SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits;
+SELECT AVG(UserID) FROM hits;
+SELECT COUNT(DISTINCT UserID) FROM hits;
+SELECT COUNT(DISTINCT SearchPhrase) FROM hits;
+SELECT MIN(EventDate), MAX(EventDate) FROM hits;
+SELECT AdvEngineID, COUNT(*) FROM hits WHERE AdvEngineID <> 0 GROUP BY AdvEngineID ORDER BY COUNT(*) DESC;
+SELECT RegionID, COUNT(DISTINCT UserID) AS u FROM hits GROUP BY RegionID ORDER BY u DESC LIMIT 10;
+SELECT RegionID, SUM(AdvEngineID), COUNT(*) AS c, AVG(ResolutionWidth), COUNT(DISTINCT UserID) FROM hits GROUP BY RegionID ORDER BY c DESC LIMIT 10;
+SELECT MobilePhoneModel, COUNT(DISTINCT UserID) AS u FROM hits WHERE MobilePhoneModel <> '' GROUP BY MobilePhoneModel ORDER BY u DESC LIMIT 10;
+SELECT MobilePhone, MobilePhoneModel, COUNT(DISTINCT UserID) AS u FROM hits WHERE MobilePhoneModel <> '' GROUP BY MobilePhone, MobilePhoneModel ORDER BY u DESC LIMIT 10;
+SELECT SearchPhrase, COUNT(*) AS c FROM hits WHERE SearchPhrase <> '' GROUP BY SearchPhrase ORDER BY c DESC LIMIT 10;
+SELECT SearchPhrase, COUNT(DISTINCT UserID) AS u FROM hits WHERE SearchPhrase <> '' GROUP BY SearchPhrase ORDER BY u DESC LIMIT 10;
+SELECT SearchEngineID, SearchPhrase, COUNT(*) AS c FROM hits WHERE SearchPhrase <> '' GROUP BY SearchEngineID, SearchPhrase ORDER BY c DESC LIMIT 10;
+SELECT UserID, COUNT(*) FROM hits GROUP BY UserID ORDER BY COUNT(*) DESC LIMIT 10;
+SELECT UserID, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, SearchPhrase ORDER BY COUNT(*) DESC LIMIT 10;
+SELECT UserID, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, SearchPhrase LIMIT 10;
+SELECT UserID, extract(minute FROM EventTime::TIMESTAMP[s]) AS m, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, extract(minute FROM EventTime::TIMESTAMP[s]), SearchPhrase ORDER BY COUNT(*) DESC LIMIT 10;
+SELECT UserID FROM hits WHERE UserID = 435090932899640449;
+SELECT COUNT(*) FROM hits WHERE URL LIKE '%google%';
+SELECT SearchPhrase, MIN(URL), COUNT(*) AS c FROM hits WHERE URL LIKE '%google%' AND SearchPhrase <> '' GROUP BY SearchPhrase ORDER BY c DESC LIMIT 10;
+SELECT SearchPhrase, MIN(URL), MIN(Title), COUNT(*) AS c, COUNT(DISTINCT UserID) FROM hits WHERE Title LIKE '%Google%' AND URL NOT LIKE '%.google.%' AND SearchPhrase <> '' GROUP BY SearchPhrase ORDER BY c DESC LIMIT 10;
+SELECT * FROM hits WHERE URL LIKE '%google%' ORDER BY EventTime LIMIT 10;
+SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY EventTime LIMIT 10;
+SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY SearchPhrase LIMIT 10;
+SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY EventTime, SearchPhrase LIMIT 10;
+SELECT CounterID, AVG(length(URL)) AS l, COUNT(*) AS c FROM hits WHERE URL <> '' GROUP BY CounterID HAVING COUNT(*) > 100000 ORDER BY l DESC LIMIT 25;
+SELECT REGEXP_REPLACE(Referer, b'^https?://(?:www\.)?([^/]+)/.*$', r'\1') AS k, AVG(length(Referer)) AS l, COUNT(*) AS c, MIN(Referer) FROM hits WHERE Referer <> '' GROUP BY REGEXP_REPLACE(Referer, b'^https?://(?:www\.)?([^/]+)/.*$', r'\1') HAVING COUNT(*) > 100000 ORDER BY l DESC LIMIT 25;
+SELECT SUM(ResolutionWidth), SUM(ResolutionWidth + 1), SUM(ResolutionWidth + 2), SUM(ResolutionWidth + 3), SUM(ResolutionWidth + 4), SUM(ResolutionWidth + 5), SUM(ResolutionWidth + 6), SUM(ResolutionWidth + 7), SUM(ResolutionWidth + 8), SUM(ResolutionWidth + 9), SUM(ResolutionWidth + 10), SUM(ResolutionWidth + 11), SUM(ResolutionWidth + 12), SUM(ResolutionWidth + 13), SUM(ResolutionWidth + 14), SUM(ResolutionWidth + 15), SUM(ResolutionWidth + 16), SUM(ResolutionWidth + 17), SUM(ResolutionWidth + 18), SUM(ResolutionWidth + 19), SUM(ResolutionWidth + 20), SUM(ResolutionWidth + 21), SUM(ResolutionWidth + 22), SUM(ResolutionWidth + 23), SUM(ResolutionWidth + 24), SUM(ResolutionWidth + 25), SUM(ResolutionWidth + 26), SUM(ResolutionWidth + 27), SUM(ResolutionWidth + 28), SUM(ResolutionWidth + 29), SUM(ResolutionWidth + 30), SUM(ResolutionWidth + 31), SUM(ResolutionWidth + 32), SUM(ResolutionWidth + 33), SUM(ResolutionWidth + 34), SUM(ResolutionWidth + 35), SUM(ResolutionWidth + 36), SUM(ResolutionWidth + 37), SUM(ResolutionWidth + 38), SUM(ResolutionWidth + 39), SUM(ResolutionWidth + 40), SUM(ResolutionWidth + 41), SUM(ResolutionWidth + 42), SUM(ResolutionWidth + 43), SUM(ResolutionWidth + 44), SUM(ResolutionWidth + 45), SUM(ResolutionWidth + 46), SUM(ResolutionWidth + 47), SUM(ResolutionWidth + 48), SUM(ResolutionWidth + 49), SUM(ResolutionWidth + 50), SUM(ResolutionWidth + 51), SUM(ResolutionWidth + 52), SUM(ResolutionWidth + 53), SUM(ResolutionWidth + 54), SUM(ResolutionWidth + 55), SUM(ResolutionWidth + 56), SUM(ResolutionWidth + 57), SUM(ResolutionWidth + 58), SUM(ResolutionWidth + 59), SUM(ResolutionWidth + 60), SUM(ResolutionWidth + 61), SUM(ResolutionWidth + 62), SUM(ResolutionWidth + 63), SUM(ResolutionWidth + 64), SUM(ResolutionWidth + 65), SUM(ResolutionWidth + 66), SUM(ResolutionWidth + 67), SUM(ResolutionWidth + 68), SUM(ResolutionWidth + 69), SUM(ResolutionWidth + 70), SUM(ResolutionWidth + 71), SUM(ResolutionWidth + 72), SUM(ResolutionWidth + 73), SUM(ResolutionWidth + 74), SUM(ResolutionWidth + 75), SUM(ResolutionWidth + 76), SUM(ResolutionWidth + 77), SUM(ResolutionWidth + 78), SUM(ResolutionWidth + 79), SUM(ResolutionWidth + 80), SUM(ResolutionWidth + 81), SUM(ResolutionWidth + 82), SUM(ResolutionWidth + 83), SUM(ResolutionWidth + 84), SUM(ResolutionWidth + 85), SUM(ResolutionWidth + 86), SUM(ResolutionWidth + 87), SUM(ResolutionWidth + 88), SUM(ResolutionWidth + 89) FROM hits;
+SELECT SearchEngineID, ClientIP, COUNT(*) AS c, SUM(IsRefresh), AVG(ResolutionWidth) FROM hits WHERE SearchPhrase <> '' GROUP BY SearchEngineID, ClientIP ORDER BY c DESC LIMIT 10;
+SELECT WatchID, ClientIP, COUNT(*) AS c, SUM(IsRefresh), AVG(ResolutionWidth) FROM hits WHERE SearchPhrase <> '' GROUP BY WatchID, ClientIP ORDER BY c DESC LIMIT 10;
+SELECT WatchID, ClientIP, COUNT(*) AS c, SUM(IsRefresh), AVG(ResolutionWidth) FROM hits GROUP BY WatchID, ClientIP ORDER BY c DESC LIMIT 10;
+SELECT URL, COUNT(*) AS c FROM hits GROUP BY URL ORDER BY c DESC LIMIT 10;
+SELECT 1, URL, COUNT(*) AS c FROM hits GROUP BY 1, URL ORDER BY c DESC LIMIT 10;
+SELECT ClientIP, ClientIP - 1, ClientIP - 2, ClientIP - 3, COUNT(*) AS c FROM hits GROUP BY ClientIP, ClientIP - 1, ClientIP - 2, ClientIP - 3 ORDER BY c DESC LIMIT 10;
+SELECT URL, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND DontCountHits = 0 AND IsRefresh = 0 AND URL <> '' GROUP BY URL ORDER BY PageViews DESC LIMIT 10;
+SELECT Title, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND DontCountHits = 0 AND IsRefresh = 0 AND Title <> '' GROUP BY Title ORDER BY PageViews DESC LIMIT 10;
+SELECT URL, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 AND IsLink <> 0 AND IsDownload = 0 GROUP BY URL ORDER BY PageViews DESC LIMIT 10 OFFSET 1000;
+SELECT TraficSourceID, SearchEngineID, AdvEngineID, CASE WHEN (SearchEngineID = 0 AND AdvEngineID = 0) THEN Referer ELSE '' END AS Src, URL AS Dst, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 GROUP BY TraficSourceID, SearchEngineID, AdvEngineID, CASE WHEN (SearchEngineID = 0 AND AdvEngineID = 0) THEN Referer ELSE '' END, URL ORDER BY PageViews DESC LIMIT 10 OFFSET 1000;
+SELECT URLHash, EventDate, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 AND TraficSourceID IN (-1, 6) AND RefererHash = 3594120000172545465 GROUP BY URLHash, EventDate ORDER BY PageViews DESC LIMIT 10 OFFSET 100;
+SELECT WindowClientWidth, WindowClientHeight, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 AND DontCountHits = 0 AND URLHash = 2868770270353813622 GROUP BY WindowClientWidth, WindowClientHeight ORDER BY PageViews DESC LIMIT 10 OFFSET 10000;
+SELECT TRUNC(EventTime::TIMESTAMP[s], 'minute') AS M, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-14'::DATE AND EventDate::DATE <= '2013-07-15'::DATE AND IsRefresh = 0 AND DontCountHits = 0 GROUP BY TRUNC(EventTime::TIMESTAMP[s], 'minute') ORDER BY M LIMIT 10 OFFSET 1000;
diff --git a/opteryx-skene/query b/opteryx-skene/query
new file mode 100755
index 0000000000..f2ca00e871
--- /dev/null
+++ b/opteryx-skene/query
@@ -0,0 +1,42 @@
+#!/bin/bash
+# Reads a SQL query from stdin, runs it via opteryx-core (Python in-process)
+# against the partitioned parquet under ./hits/.
+# Stdout: query result as TSV (header + rows).
+# Stderr: query runtime in fractional seconds on the last line.
+set -e
+
+query=$(cat)
+
+"$HOME/opteryx_venv/bin/python" - "$query" <<'PY'
+import sys
+import timeit
+
+import opteryx
+
+query = sys.argv[1]
+
+# opteryx-core execution surface: a Session that yields native
+# morsels. `execute_to_morsels` runs entirely in the native engine; draining
+# the generator is the execution. We time the drain (matching the engine's own
+# ClickBench runner), holding the morsels so results can be emitted afterwards.
+# NB: opteryx.query() / Cursor.arrow() from the 0.x line no longer exist here.
+session = opteryx.session()
+start = timeit.default_timer()
+morsels = list(session.execute_to_morsels(query))
+end = timeit.default_timer()
+
+cols = None
+for morsel in morsels:
+    if cols is None:
+        cols = morsel.column_names
+        print("\t".join(
+            c.decode() if isinstance(c, (bytes, bytearray)) else str(c)
+            for c in cols
+        ))
+    for i in range(morsel.num_rows):
+        print("\t".join("" if v is None else str(v) for v in morsel[i]))
+
+session.close()
+
+print(f"{end - start:.3f}", file=sys.stderr)
+PY
diff --git a/opteryx-skene/results/20260818/c6a.4xlarge.json b/opteryx-skene/results/20260818/c6a.4xlarge.json
new file mode 100644
index 0000000000..d51a3a9427
--- /dev/null
+++ b/opteryx-skene/results/20260818/c6a.4xlarge.json
@@ -0,0 +1,60 @@
+{
+      "system": "Opteryx",
+      "date": "2026-08-18",
+      "machine": "c6a.4xlarge",
+      "cluster_size": 1,
+      "proprietary": "no",
+      "hardware": "cpu",
+      "tuned": "no",
+      "tags": ["C++","stateless","column-oriented","embedded"],
+      "load_time": 208,
+      "data_size": 14949726123,
+      "concurrent_qps": null,
+      "concurrent_error_ratio": null,
+      "result": [
+        [0.743, 0.097, 0.098],
+        [0.915, 0.194, 0.196],
+        [1.047, 0.231, 0.232],
+        [1.023, 0.188, 0.188],
+        [1.432, 0.74, 0.743],
+        [1.978, 0.468, 0.465],
+        [0.771, 0.098, 0.098],
+        [0.942, 0.199, 0.199],
+        [1.908, 0.933, 0.926],
+        [2.992, 1.074, 1.071],
+        [1.691, 0.328, 0.327],
+        [1.931, 0.351, 0.348],
+        [1.962, 0.866, 0.866],
+        [3.382, 1.114, 1.11],
+        [2.531, 0.974, 0.975],
+        [1.707, 0.963, 0.976],
+        [3.876, 2.067, 2.071],
+        [3.828, 2.041, 2.099],
+        [7.114, 4.55, 4.548],
+        [0.967, 0.188, 0.194],
+        [7.82, 1.317, 1.308],
+        [9.734, 1.417, 1.429],
+        [15.509, 2.167, 2.176],
+        [10.165, 1.493, 1.492],
+        [3.413, 0.429, 0.426],
+        [1.918, 0.377, 0.377],
+        [3.735, 0.454, 0.453],
+        [7.99, 1.919, 1.953],
+        [7.426, 3.519, 3.49],
+        [0.754, 0.243, 0.244],
+        [4.34, 0.844, 0.849],
+        [7.162, 1.185, 1.163],
+        [8.723, 6.55, 6.641],
+        [9.235, 4.105, 4.107],
+        [9.29, 4.151, 4.128],
+        [1.394, 0.822, 0.822],
+        [0.989, 0.276, 0.276],
+        [0.951, 0.201, 0.207],
+        [0.965, 0.224, 0.219],
+        [1.247, 0.56, 0.56],
+        [0.92, 0.177, 0.175],
+        [0.892, 0.168, 0.168],
+        [0.868, 0.159, 0.164]
+]
+  } 
+  
\ No newline at end of file
diff --git a/opteryx-skene/start b/opteryx-skene/start
new file mode 100755
index 0000000000..06bd986563
--- /dev/null
+++ b/opteryx-skene/start
@@ -0,0 +1,2 @@
+#!/bin/bash
+exit 0
diff --git a/opteryx-skene/stop b/opteryx-skene/stop
new file mode 100755
index 0000000000..06bd986563
--- /dev/null
+++ b/opteryx-skene/stop
@@ -0,0 +1,2 @@
+#!/bin/bash
+exit 0
diff --git a/opteryx-skene/template.json b/opteryx-skene/template.json
new file mode 100644
index 0000000000..36032e7542
--- /dev/null
+++ b/opteryx-skene/template.json
@@ -0,0 +1,12 @@
+{
+  "system": "Opteryx",
+  "proprietary": "no",
+  "hardware": "cpu",
+  "tuned": "no",
+  "tags": [
+    "C++",
+    "stateless",
+    "column-oriented",
+    "embedded"
+  ]
+}
diff --git a/opteryx/README.md b/opteryx/README.md
index 877bee0ed1..2338ac4c10 100644
--- a/opteryx/README.md
+++ b/opteryx/README.md
@@ -1,38 +1,40 @@
 # Opteryx
 
-Opteryx is an in-process SQL query engine written in Python/Cython that leverages Apache Arrow as its in-memory format. Designed for ad hoc queries, Opteryx directly queries data from storage without requiring any preloading or preprocessing.
+Opteryx is an in-process SQL query engine. Query **planning** (parse, bind,
+optimize) runs in Python; query **execution** is native (Cython/C++). It
+queries Parquet directly from storage with no preloading or preprocessing,
+which makes it well suited to ad hoc analytics.
 
 For more information, visit:
 
-- [Opteryx Documentation](https://opteryx.dev/)
-- [Opteryx GitHub Repository](https://github.com/mabel-dev/opteryx)
+- [Opteryx Documentation](https://docs.opteryx.app/)
+- [Opteryx GitHub Repository](https://github.com/mabel-dev/opteryx-core)
 
-This page provides instructions for benchmarking Opteryx using the split Parquet files provided by ClickBench.
+This page benchmarks Opteryx (PyPI package `opteryx-core`) using the split
+Parquet files provided by ClickBench.
 
 ---
 
 ## Generating Benchmark Results
 
-To generate benchmark results, follow these steps:
-
-### **High-level Steps**
+### High-level Steps
 1. Set up the environment.
 2. Install Python and the required dependencies.
 3. Download the benchmark dataset.
 4. Run the benchmark script.
 
-### **Detailed Instructions**
+### Detailed Instructions
 
 1. **Start an AWS EC2 instance**
    - OS: Ubuntu 24
-   - Architecture: 64-bit
+   - Architecture: 64-bit (x86_64 or AArch64)
    - Instance Type: `c6a.4xlarge`
    - Root Storage: 500 GB gp2 SSD
-   - Advanced Details: Ensure 'EBS-optimized instance' is **disabled**.
+   - Advanced Details: ensure 'EBS-optimized instance' is **disabled**.
 
-2. **SSH into the instance** (after the status checks are complete):
+2. **SSH into the instance** (after status checks complete):
    ~~~bash
-   ssh ec2-user@{ip}
+   ssh ubuntu@{ip}
    ~~~
 
 3. **Update the package list and install Git**
@@ -52,6 +54,40 @@ To generate benchmark results, follow these steps:
    sudo ./benchmark.sh
    ~~~
 
+### Python version
+
+`opteryx-core` publishes cp314 x86_64 and AArch64 manylinux wheels and declares
+no runtime dependencies, so `install` is a single binary-wheel download with no
+on-box compilation and no toolchain.
+
+### Query dialect
+
+`queries.sql` adapts queries to Opteryx's dialect. The adaptations are syntactic
+— they do not change what is computed, the rows returned, or the work the engine has to do:
+
+- **Q19, Q43** — `EventTime` is stored as an integer epoch, so it is cast
+  explicitly (`EventTime::TIMESTAMP[s]`) before `extract(minute FROM ...)` and
+  before truncation.
+- **Q43** — `TRUNC(<ts>, 'minute')` rather than `DATE_TRUNC('minute', <ts>)`.
+- **Q29** — the `REGEXP_REPLACE` pattern and replacement use `b''` and `r''`
+  literals so the backslash reference survives to the regex engine.
+- **Q37-Q42** — `EventDate` comparisons cast both sides to `DATE`
+  (`EventDate::DATE >= '2013-07-01'::DATE`).
+
+### Hardware coverage
+
+Results are published for instance types with **32 or fewer vCPUs**. The account
+used for these runs is limited to 32 concurrent on-demand vCPUs, so the 192-vCPU
+machines in the ClickBench fleet (`c6a.metal`, `c7a.metal-48xl`,
+`c8g.metal-48xl`) could not be launched. The published set spans 2 to 16 vCPUs
+on both x86_64 (`c6a.*`, `t3a.small`) and AArch64 (`c8g.*`), which covers the
+small/medium range of the standard fleet on both architectures.
+
 ### Known Issues
 
-- Queries 33 and 34 fail due to Out of Memory (OOM) errors.
+- On the memory-constrained instances the heaviest `GROUP BY` queries do not fit
+  in RAM and spill to swap rather than failing. They complete, but two orders of
+  magnitude slower — on `c6a.xlarge` (8 GB) three queries account for more than
+  half the total runtime. The benchmark environment provides the 16 GB swapfile
+  that ClickBench's `cloud-init` configures for every system; without it these
+  queries would be `null` instead of slow.
diff --git a/opteryx/check b/opteryx/check
index 4d4c12fd75..ed305e8015 100755
--- a/opteryx/check
+++ b/opteryx/check
@@ -1,4 +1,34 @@
 #!/bin/bash
 set -e
 
-"$HOME/opteryx_venv/bin/python" -m opteryx "SELECT version()" >/dev/null
+# Smoke test that actually executes through the engine and propagates a
+# non-zero exit on failure. The old `python -m opteryx "SELECT version()"`
+# CLI path is unusable here for two reasons:
+#
+#   1. opteryx-core's single-run CLI calls the removed opteryx.query(),
+#      catches the AttributeError, and STILL exits 0 -- a false green that
+#      would let a broken install pass this check.
+#   2. `version()` is not a function in the current engine; the version is
+#      exposed as the system variable `@@version`.
+#
+# This matters more than it looks: bench_load() in lib/benchmark-common.sh
+# calls ./check unconditionally, AFTER the ~14 GB download and load. A check
+# that cannot pass burns the whole setup cost before failing.
+#
+# Assert the shape of the value rather than a literal, so the check does not
+# need editing on every release.
+"$HOME/opteryx_venv/bin/python" - <<'PY'
+import re
+
+import opteryx
+
+session = opteryx.session()
+morsels = list(session.execute_to_morsels("SELECT @@version"))
+session.close()
+
+assert morsels, "SELECT @@version returned no morsels"
+value = list(morsels[0][0])[0]
+if isinstance(value, (bytes, bytearray)):
+    value = value.decode()
+assert re.fullmatch(r"\d+\.\d+\.\d+.*", str(value)), f"unexpected version {value!r}"
+PY
diff --git a/opteryx/install b/opteryx/install
index f2c4ed9349..21ba004257 100755
--- a/opteryx/install
+++ b/opteryx/install
@@ -1,20 +1,44 @@
 #!/bin/bash
 set -e
 
+# Opteryx (PyPI package: opteryx-core) on stock CPython 3.14.
+#
+# NOT the free-threaded build. Opteryx abandoned the free-threaded-Python
+# experiment in June 2026; the engine's parallelism target is native (C++)
+# threads under a released GIL, so the GIL build is the supported and the
+# representative configuration. opteryx-core stopped publishing cp314t wheels
+# after 0.9.16 accordingly -- installing onto a 3.14t interpreter would fall
+# through to the sdist and try to build Rust/C++ on the box.
+#
+# opteryx-core publishes cp314 x86_64 manylinux wheels and declares no runtime
+# dependencies, so this is a single binary-wheel download -- no on-box
+# compilation and no toolchain. x86_64 only, which is fine: the canonical
+# c6a.4xlarge is x86_64.
+
 sudo apt-get update -y
 sudo apt-get install -y software-properties-common
 sudo add-apt-repository -y ppa:deadsnakes/ppa
 sudo apt-get update -y
-sudo apt-get install -y python3.11 python3.11-venv git wget build-essential python3.11-dev
+
+# Ubuntu 24.04 (noble) ships 3.12; 3.14 comes from deadsnakes, which carries
+# python3.14 for noble. python3.14-venv provides the venv module.
+sudo apt-get install -y python3.14 python3.14-venv git wget
 
 if [ ! -d "$HOME/opteryx_venv" ]; then
-    python3.11 -m venv "$HOME/opteryx_venv"
+    python3.14 -m venv "$HOME/opteryx_venv"
 fi
 
 "$HOME/opteryx_venv/bin/python" -m pip install --upgrade pip
-# 0.26.1 only ships x86_64 wheels, so arm64 hosts (c8g.*) fell through
-# to sdist where the build failed at "opteryx/third_party/abseil/
-# containers.pyx doesn't match any files". 0.26.8 publishes
-# manylinux2014_aarch64 wheels for cp310-cp313, which fixes arm64
-# without changing anything for x86_64.
-"$HOME/opteryx_venv/bin/python" -m pip install --upgrade opteryx==0.26.8
+# Pulls the cp314 wheel for the latest release.
+"$HOME/opteryx_venv/bin/python" -m pip install --upgrade opteryx-core
+
+# Fail loudly here rather than 43 queries later if pip silently fell back to a
+# source build or resolved an interpreter we did not expect.
+"$HOME/opteryx_venv/bin/python" - <<'PY'
+import sys
+
+import opteryx
+
+print(f"python  {sys.version.split()[0]}  (GIL enabled: {sys._is_gil_enabled()})")
+print(f"opteryx-core {opteryx.__version__}")
+PY
diff --git a/opteryx/queries.sql b/opteryx/queries.sql
index fa5568c632..906e5cd6ec 100644
--- a/opteryx/queries.sql
+++ b/opteryx/queries.sql
@@ -16,7 +16,7 @@ SELECT SearchEngineID, SearchPhrase, COUNT(*) AS c FROM hits WHERE SearchPhrase
 SELECT UserID, COUNT(*) FROM hits GROUP BY UserID ORDER BY COUNT(*) DESC LIMIT 10;
 SELECT UserID, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, SearchPhrase ORDER BY COUNT(*) DESC LIMIT 10;
 SELECT UserID, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, SearchPhrase LIMIT 10;
-SELECT UserID, extract(minute FROM EventTime) AS m, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, extract(minute FROM EventTime), SearchPhrase ORDER BY COUNT(*) DESC LIMIT 10;
+SELECT UserID, extract(minute FROM EventTime::TIMESTAMP[s]) AS m, SearchPhrase, COUNT(*) FROM hits GROUP BY UserID, extract(minute FROM EventTime::TIMESTAMP[s]), SearchPhrase ORDER BY COUNT(*) DESC LIMIT 10;
 SELECT UserID FROM hits WHERE UserID = 435090932899640449;
 SELECT COUNT(*) FROM hits WHERE URL LIKE '%google%';
 SELECT SearchPhrase, MIN(URL), COUNT(*) AS c FROM hits WHERE URL LIKE '%google%' AND SearchPhrase <> '' GROUP BY SearchPhrase ORDER BY c DESC LIMIT 10;
@@ -26,7 +26,7 @@ SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY EventTime LIMIT
 SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY SearchPhrase LIMIT 10;
 SELECT SearchPhrase FROM hits WHERE SearchPhrase <> '' ORDER BY EventTime, SearchPhrase LIMIT 10;
 SELECT CounterID, AVG(length(URL)) AS l, COUNT(*) AS c FROM hits WHERE URL <> '' GROUP BY CounterID HAVING COUNT(*) > 100000 ORDER BY l DESC LIMIT 25;
-SELECT REGEXP_REPLACE(Referer, b'^https?://(?:www\.)?([^/]+)/.*$', r'\\1') AS k, AVG(length(Referer)) AS l, COUNT(*) AS c, MIN(Referer) FROM hits WHERE Referer <> '' GROUP BY REGEXP_REPLACE(Referer, b'^https?://(?:www\.)?([^/]+)/.*$', r'\\1') HAVING COUNT(*) > 100000 ORDER BY l DESC LIMIT 25;
+SELECT REGEXP_REPLACE(Referer, b'^https?://(?:www\.)?([^/]+)/.*$', r'\1') AS k, AVG(length(Referer)) AS l, COUNT(*) AS c, MIN(Referer) FROM hits WHERE Referer <> '' GROUP BY REGEXP_REPLACE(Referer, b'^https?://(?:www\.)?([^/]+)/.*$', r'\1') HAVING COUNT(*) > 100000 ORDER BY l DESC LIMIT 25;
 SELECT SUM(ResolutionWidth), SUM(ResolutionWidth + 1), SUM(ResolutionWidth + 2), SUM(ResolutionWidth + 3), SUM(ResolutionWidth + 4), SUM(ResolutionWidth + 5), SUM(ResolutionWidth + 6), SUM(ResolutionWidth + 7), SUM(ResolutionWidth + 8), SUM(ResolutionWidth + 9), SUM(ResolutionWidth + 10), SUM(ResolutionWidth + 11), SUM(ResolutionWidth + 12), SUM(ResolutionWidth + 13), SUM(ResolutionWidth + 14), SUM(ResolutionWidth + 15), SUM(ResolutionWidth + 16), SUM(ResolutionWidth + 17), SUM(ResolutionWidth + 18), SUM(ResolutionWidth + 19), SUM(ResolutionWidth + 20), SUM(ResolutionWidth + 21), SUM(ResolutionWidth + 22), SUM(ResolutionWidth + 23), SUM(ResolutionWidth + 24), SUM(ResolutionWidth + 25), SUM(ResolutionWidth + 26), SUM(ResolutionWidth + 27), SUM(ResolutionWidth + 28), SUM(ResolutionWidth + 29), SUM(ResolutionWidth + 30), SUM(ResolutionWidth + 31), SUM(ResolutionWidth + 32), SUM(ResolutionWidth + 33), SUM(ResolutionWidth + 34), SUM(ResolutionWidth + 35), SUM(ResolutionWidth + 36), SUM(ResolutionWidth + 37), SUM(ResolutionWidth + 38), SUM(ResolutionWidth + 39), SUM(ResolutionWidth + 40), SUM(ResolutionWidth + 41), SUM(ResolutionWidth + 42), SUM(ResolutionWidth + 43), SUM(ResolutionWidth + 44), SUM(ResolutionWidth + 45), SUM(ResolutionWidth + 46), SUM(ResolutionWidth + 47), SUM(ResolutionWidth + 48), SUM(ResolutionWidth + 49), SUM(ResolutionWidth + 50), SUM(ResolutionWidth + 51), SUM(ResolutionWidth + 52), SUM(ResolutionWidth + 53), SUM(ResolutionWidth + 54), SUM(ResolutionWidth + 55), SUM(ResolutionWidth + 56), SUM(ResolutionWidth + 57), SUM(ResolutionWidth + 58), SUM(ResolutionWidth + 59), SUM(ResolutionWidth + 60), SUM(ResolutionWidth + 61), SUM(ResolutionWidth + 62), SUM(ResolutionWidth + 63), SUM(ResolutionWidth + 64), SUM(ResolutionWidth + 65), SUM(ResolutionWidth + 66), SUM(ResolutionWidth + 67), SUM(ResolutionWidth + 68), SUM(ResolutionWidth + 69), SUM(ResolutionWidth + 70), SUM(ResolutionWidth + 71), SUM(ResolutionWidth + 72), SUM(ResolutionWidth + 73), SUM(ResolutionWidth + 74), SUM(ResolutionWidth + 75), SUM(ResolutionWidth + 76), SUM(ResolutionWidth + 77), SUM(ResolutionWidth + 78), SUM(ResolutionWidth + 79), SUM(ResolutionWidth + 80), SUM(ResolutionWidth + 81), SUM(ResolutionWidth + 82), SUM(ResolutionWidth + 83), SUM(ResolutionWidth + 84), SUM(ResolutionWidth + 85), SUM(ResolutionWidth + 86), SUM(ResolutionWidth + 87), SUM(ResolutionWidth + 88), SUM(ResolutionWidth + 89) FROM hits;
 SELECT SearchEngineID, ClientIP, COUNT(*) AS c, SUM(IsRefresh), AVG(ResolutionWidth) FROM hits WHERE SearchPhrase <> '' GROUP BY SearchEngineID, ClientIP ORDER BY c DESC LIMIT 10;
 SELECT WatchID, ClientIP, COUNT(*) AS c, SUM(IsRefresh), AVG(ResolutionWidth) FROM hits WHERE SearchPhrase <> '' GROUP BY WatchID, ClientIP ORDER BY c DESC LIMIT 10;
@@ -34,10 +34,10 @@ SELECT WatchID, ClientIP, COUNT(*) AS c, SUM(IsRefresh), AVG(ResolutionWidth) FR
 SELECT URL, COUNT(*) AS c FROM hits GROUP BY URL ORDER BY c DESC LIMIT 10;
 SELECT 1, URL, COUNT(*) AS c FROM hits GROUP BY 1, URL ORDER BY c DESC LIMIT 10;
 SELECT ClientIP, ClientIP - 1, ClientIP - 2, ClientIP - 3, COUNT(*) AS c FROM hits GROUP BY ClientIP, ClientIP - 1, ClientIP - 2, ClientIP - 3 ORDER BY c DESC LIMIT 10;
-SELECT URL, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-01' AND EventDate <= '2013-07-31' AND DontCountHits = 0 AND IsRefresh = 0 AND URL <> '' GROUP BY URL ORDER BY PageViews DESC LIMIT 10;
-SELECT Title, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-01' AND EventDate <= '2013-07-31' AND DontCountHits = 0 AND IsRefresh = 0 AND Title <> '' GROUP BY Title ORDER BY PageViews DESC LIMIT 10;
-SELECT URL, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-01' AND EventDate <= '2013-07-31' AND IsRefresh = 0 AND IsLink <> 0 AND IsDownload = 0 GROUP BY URL ORDER BY PageViews DESC LIMIT 10 OFFSET 1000;
-SELECT TraficSourceID, SearchEngineID, AdvEngineID, CASE WHEN (SearchEngineID = 0 AND AdvEngineID = 0) THEN Referer ELSE '' END AS Src, URL AS Dst, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-01' AND EventDate <= '2013-07-31' AND IsRefresh = 0 GROUP BY TraficSourceID, SearchEngineID, AdvEngineID, CASE WHEN (SearchEngineID = 0 AND AdvEngineID = 0) THEN Referer ELSE '' END, URL ORDER BY PageViews DESC LIMIT 10 OFFSET 1000;
-SELECT URLHash, EventDate, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-01' AND EventDate <= '2013-07-31' AND IsRefresh = 0 AND TraficSourceID IN (-1, 6) AND RefererHash = 3594120000172545465 GROUP BY URLHash, EventDate ORDER BY PageViews DESC LIMIT 10 OFFSET 100;
-SELECT WindowClientWidth, WindowClientHeight, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-01' AND EventDate <= '2013-07-31' AND IsRefresh = 0 AND DontCountHits = 0 AND URLHash = 2868770270353813622 GROUP BY WindowClientWidth, WindowClientHeight ORDER BY PageViews DESC LIMIT 10 OFFSET 10000;
-SELECT DATE_TRUNC('minute', EventTime) AS M, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate >= '2013-07-14' AND EventDate <= '2013-07-15' AND IsRefresh = 0 AND DontCountHits = 0 GROUP BY DATE_TRUNC('minute', EventTime) ORDER BY M LIMIT 10 OFFSET 1000;
+SELECT URL, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND DontCountHits = 0 AND IsRefresh = 0 AND URL <> '' GROUP BY URL ORDER BY PageViews DESC LIMIT 10;
+SELECT Title, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND DontCountHits = 0 AND IsRefresh = 0 AND Title <> '' GROUP BY Title ORDER BY PageViews DESC LIMIT 10;
+SELECT URL, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 AND IsLink <> 0 AND IsDownload = 0 GROUP BY URL ORDER BY PageViews DESC LIMIT 10 OFFSET 1000;
+SELECT TraficSourceID, SearchEngineID, AdvEngineID, CASE WHEN (SearchEngineID = 0 AND AdvEngineID = 0) THEN Referer ELSE '' END AS Src, URL AS Dst, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 GROUP BY TraficSourceID, SearchEngineID, AdvEngineID, CASE WHEN (SearchEngineID = 0 AND AdvEngineID = 0) THEN Referer ELSE '' END, URL ORDER BY PageViews DESC LIMIT 10 OFFSET 1000;
+SELECT URLHash, EventDate, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 AND TraficSourceID IN (-1, 6) AND RefererHash = 3594120000172545465 GROUP BY URLHash, EventDate ORDER BY PageViews DESC LIMIT 10 OFFSET 100;
+SELECT WindowClientWidth, WindowClientHeight, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-01'::DATE AND EventDate::DATE <= '2013-07-31'::DATE AND IsRefresh = 0 AND DontCountHits = 0 AND URLHash = 2868770270353813622 GROUP BY WindowClientWidth, WindowClientHeight ORDER BY PageViews DESC LIMIT 10 OFFSET 10000;
+SELECT TRUNC(EventTime::TIMESTAMP[s], 'minute') AS M, COUNT(*) AS PageViews FROM hits WHERE CounterID = 62 AND EventDate::DATE >= '2013-07-14'::DATE AND EventDate::DATE <= '2013-07-15'::DATE AND IsRefresh = 0 AND DontCountHits = 0 GROUP BY TRUNC(EventTime::TIMESTAMP[s], 'minute') ORDER BY M LIMIT 10 OFFSET 1000;
diff --git a/opteryx/query b/opteryx/query
index 6d37357757..f2ca00e871 100755
--- a/opteryx/query
+++ b/opteryx/query
@@ -1,7 +1,7 @@
 #!/bin/bash
-# Reads a SQL query from stdin, runs it via opteryx (Python in-process)
+# Reads a SQL query from stdin, runs it via opteryx-core (Python in-process)
 # against the partitioned parquet under ./hits/.
-# Stdout: query result.
+# Stdout: query result as TSV (header + rows).
 # Stderr: query runtime in fractional seconds on the last line.
 set -e
 
@@ -10,42 +10,33 @@ query=$(cat)
 "$HOME/opteryx_venv/bin/python" - "$query" <<'PY'
 import sys
 import timeit
+
 import opteryx
 
 query = sys.argv[1]
 
+# opteryx-core execution surface: a Session that yields native
+# morsels. `execute_to_morsels` runs entirely in the native engine; draining
+# the generator is the execution. We time the drain (matching the engine's own
+# ClickBench runner), holding the morsels so results can be emitted afterwards.
+# NB: opteryx.query() / Cursor.arrow() from the 0.x line no longer exist here.
+session = opteryx.session()
 start = timeit.default_timer()
-res = opteryx.query(query)
-# In opteryx 0.26 `opteryx.query()` returns a Cursor / Relation. Probe
-# the common materialisation methods in priority order — different
-# versions expose different combinations.
-out = None
-for method in ("arrow", "to_arrow_table", "fetchall"):
-    fn = getattr(res, method, None)
-    if callable(fn):
-        try:
-            out = fn()
-            break
-        except Exception:
-            continue
+morsels = list(session.execute_to_morsels(query))
 end = timeit.default_timer()
 
-if out is None:
-    # Final fallback: iterate the cursor directly (older API).
-    for row in res:
-        print(row)
-elif hasattr(out, "to_pylist"):
-    # pyarrow.Table — print as TSV header + rows.
-    cols = out.column_names
-    print("\t".join(cols))
-    for row in out.to_pylist():
-        print("\t".join("" if row.get(c) is None else str(row.get(c))
-                        for c in cols))
-elif isinstance(out, (list, tuple)):
-    for row in out:
-        print(row)
-else:
-    print(out)
+cols = None
+for morsel in morsels:
+    if cols is None:
+        cols = morsel.column_names
+        print("\t".join(
+            c.decode() if isinstance(c, (bytes, bytearray)) else str(c)
+            for c in cols
+        ))
+    for i in range(morsel.num_rows):
+        print("\t".join("" if v is None else str(v) for v in morsel[i]))
+
+session.close()
 
 print(f"{end - start:.3f}", file=sys.stderr)
 PY
diff --git a/opteryx/results/20260818/c6a.4xlarge.json b/opteryx/results/20260818/c6a.4xlarge.json
new file mode 100644
index 0000000000..16dfa165ff
--- /dev/null
+++ b/opteryx/results/20260818/c6a.4xlarge.json
@@ -0,0 +1,60 @@
+{
+      "system": "Opteryx (Parquet, partitioned)",
+      "date": "2026-08-18",
+      "machine": "c6a.4xlarge",
+      "cluster_size": 1,
+      "proprietary": "no",
+      "hardware": "cpu",
+      "tuned": "no",
+      "tags": ["C++","stateless","column-oriented","embedded"],
+      "load_time": 21,
+      "data_size": 14737666736,
+      "concurrent_qps": null,
+      "concurrent_error_ratio": null,
+      "result": [
+        [0.602, 0.11, 0.11],
+        [0.76, 0.29, 0.287],
+        [1.041, 0.401, 0.396],
+        [1.143, 0.346, 0.348],
+        [1.362, 0.832, 0.833],
+        [1.679, 0.77, 0.784],
+        [0.798, 0.109, 0.11],
+        [1.001, 0.291, 0.288],
+        [1.857, 1.153, 1.157],
+        [2.221, 1.495, 1.502],
+        [1.262, 0.61, 0.603],
+        [1.435, 0.712, 0.723],
+        [1.958, 1.168, 1.149],
+        [3.336, 1.461, 1.485],
+        [1.871, 1.333, 1.297],
+        [1.841, 1.146, 1.143],
+        [4.006, 2.477, 2.458],
+        [3.731, 2.408, 2.414],
+        [7.174, 4.968, 4.991],
+        [0.826, 0.217, 0.219],
+        [10.747, 3.003, 2.995],
+        [12.519, 3.406, 3.401],
+        [23.432, 5.067, 5.078],
+        [12.454, 3.782, 3.662],
+        [3.042, 0.948, 0.951],
+        [1.332, 0.711, 0.691],
+        [3.28, 0.977, 0.967],
+        [9.99, 2.743, 2.786],
+        [9.096, 4.596, 4.614],
+        [0.795, 0.379, 0.378],
+        [2.615, 1.418, 1.415],
+        [6.549, 1.886, 1.899],
+        [7.972, 6.737, 6.69],
+        [11.679, 5.618, 5.711],
+        [11.674, 5.732, 5.701],
+        [1.408, 0.962, 0.958],
+        [1.192, 0.43, 0.45],
+        [0.987, 0.257, 0.257],
+        [1.172, 0.405, 0.405],
+        [1.627, 0.801, 0.809],
+        [0.966, 0.216, 0.217],
+        [0.933, 0.212, 0.212],
+        [0.945, 0.223, 0.222]
+]
+  } 
+  
\ No newline at end of file
diff --git a/opteryx/template.json b/opteryx/template.json
index 548d083dd8..159898b59b 100644
--- a/opteryx/template.json
+++ b/opteryx/template.json
@@ -1,9 +1,10 @@
 {
-  "system": "Opteryx",
+  "system": "Opteryx (Parquet, partitioned)",
   "proprietary": "no",
   "hardware": "cpu",
   "tuned": "no",
   "tags": [
+    "C++",
     "stateless",
     "column-oriented",
     "embedded"