Repository navigation
Releases
rust-vulns corpus pack (pack-rust-vulns-2026-11-09)
The rust-vulns corpus pack, built and signed by the scheduled
pack publish workflow from tools/packs/rust-vulns/recipe.toml.
One release per build; under immutable releases a same-day
rebuild publishes under -r2, -r3, … instead of replacing
the day's earlier release. The newest is always at
https://github.com/xerj-org/xerj/releases?q=pack-rust-vulns
Build history lives in the releases themselves and in the
workflow runs.
Install (public key: tools/packs/keys/rust-vulns.pub in the repo):
xerj corpus add rust-vulns --from rust-vulns-pack.zip \
--verify-sig rust-vulns.pub
xerj corpus index rust-vulns
rust-vulns-freshness.json carries the build date and record
count; the same 30-day staleness refusal that governs every
reference corpus governs this one.
agent-session-trajectories corpus pack (pack-agent-session-trajectories-2026-11-09)
The agent-session-trajectories corpus pack — the first
external-contributor pack on the hub (chunxiaoxx / Nautilus,
#1255; demand anchored to #1138, wishlist slot #1210). 70
sanitized multi-session agent trajectories with failure turns
kept verbatim, built and signed by the pack publish workflow
from tools/packs
8000
/agent-session-trajectories/recipe.toml.
One release per build; under immutable releases a same-day
rebuild publishes under -r2, -r3, … instead of replacing
the day's earlier release. The newest is always at
https://github.com/xerj-org/xerj/releases?q=pack-agent-session-trajectories
Install (public key: tools/packs/keys/agent-session-trajectories.pub in the repo):
xerj corpus add agent-session-trajectories \
--from agent-session-trajectories-pack.zip \
--verify-sig agent-session-trajectories.pub
xerj corpus index agent-session-trajectories
agent-session-trajectories-freshness.json carries the build
date and record count; the same 30-day staleness refusal that
governs every reference corpus governs this one.
hub pack: agent-session-trajectories @ 13d0fbe1
Pre-indexed, signed Corpus Hub v2 pack for the agent-session-trajectories record corpus.
- content pin:
13d0fbe1ac89198caf3d24ef622636c6b47edd32(records of 2026-11-09) - licence: CC-BY-4.0 (adapt-with-attribution)
- definitions: 70, files: 1
- built by: xerj 1.0.0-rc.102
Use it: xerj hub pull agent-session-trajectories. The archive is checked against the hub key (sha256, then ed25519 signature) before anything is extracted.
Built by https://github.com/xerj-org/xerj/actions/runs/42562030700 from tools/packs/agent-session-trajectories/recipe.toml
xerj v1.0.0-rc.102
Linux-only release. This release ships Linux binaries (x86_64 and aarch64, glibc and musl, plus .deb for both glibc targets) and the XERJ Query-1 model bundle. macOS and Windows builds follow in an update of this release. Until then the installer and releases/latest keep serving v1.0.0-rc.101 on every platform; install rc.102 on Linux by downloading the assets below.
[1.0.0-rc.102] - 2026-11-06
rc.102: the XERJ Query-1 serve rule R8, a guard against counts the counting
path cannot express, and the rc.102 public window. The model bundle that
ships with this release, candidate_a_s47_r8, holds the same weights as
rc.101's candidate_a_s47 (every weight file has the same sha256) and changes
only how answers are served: rule R8 (details under Changed). On XERJ Score v2
(credit measured against Opus 5.5 = 100), on the four development families
teacher_dev, bench_select, s1b_dev and d1_select (400 items each,
1,600 in all), seed 47 moves from 90.29 to 94.69 (+4.40, 95% interval
+3.32 to +5.64, paired world-cluster bootstrap); seeds 11 and 29 move +3.95
and +4.17. These are development items, and R8 was chosen from a diagnosis of
those same items, so 94.69 is an in-sample number, not a held-out one. Wrong
answers rise on every seed: 9 to 10 of 1,600 on seed 47, 12 to 15 on seed 11.
The one held-out read is the fresh_v1 set, read once on 2026-11-04 and
re-scored under R8 without a second read: Net 0.3333 to 0.3481, right 112 to
118, wrong 11 to 12. Robustness to new wordings and fresh-set counting are not
separately measured. The engine serves the evaluated rule exactly: on the
merged rc.102 tree the engine's rows for all 1,600 development items give the
same served outcome as the evaluation's Python rows (1,600 of 1,600; 284 rows
differ only in the raw status, UNKNOWN versus UNSUPPORTED, and both are
declines). Without --facts-model-dir a node is the node rc.101 was.
Upgrade notes
- The rc.102 bundle needs an rc.102 binary. rc.101 refuses its
bundle.json(unknown ruledecline_v2_r8, model stateunavailable).
The rc.102 binary still loads the rc.101 bundle and serves it under
decline_v2with partner agreement, byte for byte as before. - Counting no longer waits for a partner under R8. With the rc.102
bundle a count is served when the question routes to counting, the hard
checks pass and the engine answers; the partner model does not run, and
partner_agreementisnullinstead oftrue. Readfinal_statusand
final_answeras before. - Data directories. rc.102 keeps rc.101's storage defaults (ZPS4
postings, row-dictionary stored text, on-disk format 4), so an rc.101 data
directory opens unchanged. The one-way upgrade from rc.100 described under
rc.101 still applies to a directory rc.100 last wrote. Not done in this
window, and carried forward: rc.101 said its storage and ingest lanes would
reach publicmainbefore rc.102. They have not; they are in the private
build tree only, and a build from publicmainstill writes format 3 and
refuses a format-4 directory. - Provenance. The rc.102 binaries are built from a private tree: public
mainat34a2f244(this window's merges) plus the XERJ-Query engine lanes
plus the model cratesxerj-infer,xerj-serveandxerj-exec, which are
licensed Proprietary and ship as binary + model bundle only. Those
crates are not Apache-2.0 and are not in the public repository; the
Apache-2.0 LICENSE covers the engine source onmain. - XERJ-Query stays opt-in. The flag is
--facts-model-dir <DIR>(or
XERJ_FACTS_MODEL_DIR); without it every/_factspath answers 404
xerj_query_not_enabled.
Added
xerj autoindex <path> --facts <namespace>writes table rows as XERJ Query-1
facts: SQLite files, SQL dumps, CSV/TSV and XLSX (#2988). Each row is one subject
whosesubject_idis the key column (id,<table>_id, … or--key), each
other column one typed fact quoting the row, so counting questions engage
without hand-written facts. Re-runs write only changed rows (withdraw the old
version, write the new one) and withdraw removed rows; unchanged rows are a
no-op. Measured on rc.101: 1,000 rows → 6,000 facts in 0.36 s; "How many
people have office Berlin?" → 366 viapath: countingwith partner
agreement (gold 366). The mapping is a per-row-change function
(facts_rows::RowFactMapper::map_change) so streamed sources can reuse it.- Live Postgres sources inside the node (#2984):
xerj connect postgres://…snapshots the chosen tables in a consistent snapshot, then
applies every insert, update, delete and truncate through logical
replication, survives restarts and resumes from its acknowledged LSN.
HTTPGET /_xerj/sources,GET|PUT|DELETE /_xerj/sources/{name}; MCP tool
xerj_sources(the MCP lists 21 tools without the hub key, 24 with it).
Measured in that PR on an x86_64 laptop (32 threads) against Postgres 16 on
the same host: a 1,000,000-row table snapshotted at 54,776 rows/s;
commit-to-readable p50 16.9 ms, p95 26.6 ms over 6,000 samples at 991
commits/s; after akill -9mid-stream, 27,696 rows = 27,696 docs with 0
missing, extra, stale or duplicate. Seedocs/LIVE_SOURCES.md. xerj autoindexindexes zip, tar, tar.gz/tgz and tar.zst archives, one
level deep, refusing absolute and..member paths, links and special
files, with byte, member and expansion-ratio caps (#2982).xerj sources [--json]anddocs/SOURCES.mdlist every way to get data in.- Cohere Rerank as a second
_searchreranking provider,"rerank": {"provider": "cohere"}, off until an operator sets a key; the endpoint is
configurable so document text can stay on a private Cohere deployment
(#2985). Seedocs/RERANK.md. - Cohere examples (#2983): an embedder config for Cohere's OpenAI-compatible
route, a stdlib Command RAG example whose citations resolve to file and
page, and the XERJ tools as a Cohere Chat v2 tool schema. Replay-tested
against a mock Cohere API; no live Cohere run yet. - The rc.101 documentation set and
/get --with-model(#2969): the
/_factsAPI, model bundle and quickstart pages, and an installer flag that
fetches the model bundle with the binary and checks its sha256.
Changed
- Serve rule R8 (opt-in per bundle; the rc.102 bundle opts in).
bundle.jsonmay namefacts.rule = "decline_v2_r8"and
counting.signal = "H". Underdecline_v2_r8the facts path serves the
raw program's answer wheredecline_v2would have declined it with policy
codeD2C_COUNT_COMPARISONorD2D_TYPE_DATE; every other code declines as
before. Under signalHa count is served on its own hard checks, without
the partner model. A bundle that names neither keepsdecline_v2and
signalA. The rule and its numbers are summarised at the top of this
section. - XERJ Query-1 launch claims corrected across the site, README and docs after
a review of 1,370 statements on 13 surfaces (#3023): latency is answer time
against Messages API time, never CLI start-up; the accuracy headline names
its scope (each question's own records given, development-set questions);
corpora are named next to numbers. - The
/_factswait loop inlanding/llms.txtanddocs/MODEL_QUICKSTART.md
stops onunavailableandxerj_query_not_enabledinstead of polling
forever (#3028). This is the docs half of issue #3021; the start-up half
stays open.
Fixed
- A count the counting op-set cannot express is no longer served (partial
fix of issue #3015). rc.101
served "How many people are in Berlin or Munich?" (Berlin x2, Munich x1) as
a verified count of 1, because both models dropped theorand agreed on
the program for Munich alone. When the serve rule would serve a count and
the question joins alternatives withor(either … or,nor) or asks
for a sum or a total of a quantity ("total height"), the answer is now a
decline on the counting path:final_status: "UNKNOWN",reason_code: "COUNT_CONNECTIVE_UNSUPPORTED", a reason naming the connective, and the
withheld program's sha aswithheld_program_sha256. Anorthat only makes
a threshold inclusive ("12 lines or more", "on or after 1990-01-01") is a
count the op-set expresses and is still served. The guard matters more
under R8, which drops the partner check: on the rc.102 release binary with
the rc.102 bundle, both "How many people are in Berlin or Munich?" and "How
many people have office Berlin or Munich?" reached the serve point with a
one-city program (the withheld shas are those of "office Munich" and
"office Berlin") and are declined; "How many people have office Berlin?"
is still served as 2. Regression tests: the three Berlin-or-Munich probes
(xerj-serveunit tests, and an HTTP test over the real models). What
stays open on #3015: returning the counted member ids. _searchwithprofile: truereports measured phase timings, in
nanoseconds, instead of 0/1 ns placeholders (#3029, issue #3005). Only
phases are measured: a non-zero breakdown timer is its phase total (an
upper bound),*_countfields are not measured, and the new field
profile.xerj_timingsays so.- Tied kNN neighbours are ranked by score, then
seq_no, then_idon every
executor (#3030, issue #3017), so a kNN or hybrid ranking with ties is the
same in the indexing process and after a restart. On a synthetic 5,000-doc
set with 2,500 distinct 384-d vectors and 160 queries, indexing process
versus restart went from 234 of 320 identical rankings to 320 of 320. - The proxy embedder accepts embedding responses without a per-item
index
(Cohere's OpenAI-compatible route omits it), sosemantic_textwrites
through such a route no longer fail withresponse parse(#2983). - The aarch64 Linux release assets are built with a 64 KiB jemalloc page
agai...
hub pack: rust-vulns @ caf0406d
Pre-indexed, signed Corpus Hub v2 pack for the rust-vulns record corpus.
- content pin:
caf0406df2ffaa4b81c9878fa7a13215981dd714(records of 2026-11-08) - licence: CC0-1.0 (adapt-with-attribution)
- definitions: 2916, files: 2
- built by: xerj 1.0.0-rc.102
Use it: xerj hub pull rust-vulns. The archive is checked against the hub key (sha256, then ed25519 signature) before anything is extracted.
Built by https://github.com/xerj-org/xerj/actions/runs/42446821657 from tools/packs/rust-vulns/recipe.toml
hub pack: rust-vulns @ caf0406d
Pre-indexed, signed Corpus Hub v2 pack for the rust-vulns record corpus.
- content pin:
caf0406df2ffaa4b81c9878fa7a13215981dd714(records of 2026-11-07) - licence: CC0-1.0 (adapt-with-attribution)
- definitions: 2916, files: 2
- built by: xerj 1.0.0-rc.101
Use it: xerj hub pull rust-vulns. The archive is checked against the hub key (sha256, then ed25519 signature) before anything is extracted.
Built by https://github.com/xerj-org/xerj/actions/runs/42338433850 from tools/packs/rust-vulns/recipe.toml
rust-vulns corpus pack (pack-rust-vulns-2026-11-08)
The rust-vulns corpus pack, built and signed by the scheduled
pack publish workflow from tools/packs/rust-vulns/recipe.toml.
One release per build; under immutable releases a same-day
rebuild publishes under -r2, -r3, … instead of replacing
the day's earlier release. The newest is always at
https://github.com/xerj-org/xerj/releases?q=pack-rust-vulns
Build history lives in the releases themselves and in the
workflow runs.
Install (public key: tools/packs/keys/rust-vulns.pub in the repo):
xerj corpus add rust-vulns --from rust-vulns-pack.zip \
--verify-sig rust-vulns.pub
xerj corpus index rust-vulns
rust-vulns-freshness.json carries the build date and record
count; the same 30-day staleness refusal that governs every
reference corpus governs this one.
rust-vulns corpus pack (pack-rust-vulns-2026-11-07)
The rust-vulns corpus pack, built and signed by the scheduled
pack publish workflow from tools/packs/rust-vulns/recipe.toml.
One release per build; under immutable releases a same-day
rebuild publishes under -r2, -r3, … instead of replacing
the day's earlier release. The newest is always at
https://github.com/xerj-org/xerj/releases?q=pack-rust-vulns
Build history lives in the releases themselves and in the
workflow runs.
Install (public key: tools/packs/keys/rust-vulns.pub in the repo):
xerj corpus add rust-vulns --from rust-vulns-pack.zip \
--verify-sig rust-vulns.pub
xerj corpus index rust-vulns
rust-vulns-freshness.json carries the build date and record
count; the same 30-day staleness refusal that governs every
reference corpus governs this one.
agent-session-trajectories corpus pack (pack-agent-session-trajectories-2026-11-08)
The agent-session-trajectories corpus pack — the first
external-contributor pack on the hub (chunxiaoxx / Nautilus,
#1255; demand anchored to #1138, wishlist slot #1210). 70
sanitized multi-session agent trajectories with failure turns
kept verbatim, built and signed by the pack publish workflow
from tools/packs/agent-session-trajectories/recipe.toml.
One release per build; under immutable releases a same-day
rebuild publishes under -r2, -r3, … instead of replacing
the day's earlier release. The newest is always at
https://github.com/xerj-org/xerj/releases?q=pack-agent-session-trajectories
Install (public key: tools/packs/keys/agent-session-trajectories.pub in the repo):
xerj corpus add agent-session-trajectories \
--from agent-session-trajectories-pack.zip \
--verify-sig agent-session-trajectories.pub
xerj corpus index agent-session-trajectories
agent-session-trajectories-freshness.json carries the build
date and record count; the same 30-day staleness refusal that
governs every reference corpus governs this one.
agent-session-trajectories corpus pack (pack-agent-session-trajectories-2026-11-07)
The agent-session-trajectories corpus pack — the first
external-contributor pack on the hub (chunxiaoxx / Nautilus,
#1255; demand anchored to #1138, wishlist slot #1210). 70
sanitized multi-session agent trajectories with failure turns
kept verbatim, built and signed by the pack publish workflow
from tools/packs/agent-session-trajectories/recipe.toml.
One release per build; under immutable releases a same-day
rebuild publishes under -r2, -r3, … instead of replacing
the day's earlier release. The newest is always at
https://github.com/xerj-org/xerj/releases?q=pack-agent-session-trajectories
Install (public key: tools/packs/keys/agent-session-trajectories.pub in the repo):
xerj corpus add agent-session-trajectories \
--from agent-session-trajectories-pack.zip \
--verify-sig agent-session-trajectories.pub
xerj corpus index agent-session-trajectories
agent-session-trajectories-freshness.json carries the build
date and record count; the same 30-day staleness refusal that
governs every reference corpus governs this one.