8000 blog: XERJ Query-1, a neurosymbolic model that compiles prompts to verifiable search queries by xerj-org · Pull Request #2976 · xerj-org/xerj · GitHub
[go: up one dir, main page]

Skip to content

blog: XERJ Query-1, a neurosymbolic model that compiles prompts to verifiable search queries - #2976

Merged
xerj-org merged 1 commit into
mainfrom
blog/xerj-query-1
Nov 5, 2026
Merged

xerj-org merged 1 commit into
mainfrom
blog/xerj-query-1

Conversation

@xerj-org
@xerj-org xerj-org commented Nov 5, 2026
Copy link
Copy Markdown
Owner

The rc.101 launch post for XERJ Query-1 (opt-in, --facts-model-dir <DIR>).

  • landing/blog/xerj-query-ai-model.html: the post. Every number traces to a measured run or is labelled arithmetic; known limits are stated in the post.
  • landing/blog/index.html: index card on top.
  • scripts/seo/pagedata.py: page registration (published 2026-11-05).
  • scripts/seo/blog_covers/: a cover motif drawn from the post's head-to-head counts, plus og / hero / thumb renders.

Checks run locally: generators --write, then every --check (fix_heads, set_nav, fix_links, gen_sitemap, mk_og_card, build_articles, build_skills); seo_lint 0 gate violations; test_article_schema, test_urlmap, test_verify_agent_discovery, test_llms_field_report; landing-constants-guard.sh passes.

🤖 Generated with Claude Code

…rifiable search queries

The rc.101 launch post for XERJ Query-1, the opt-in model that turns a
question into a small program the engine runs over the user's own facts
(enable with --facts-model-dir; off by default).

Eight chapters: why search needs a new model, how AI search works today,
what generic LLM search costs, the neurosymbolic approach (three diagrams),
how it was made and trained, how it works (records, program, trace,
verifier, every op the model writes), tests and benchmarks, and the
disclosure with what it does not do yet.

Every number is from a measured run or labelled arithmetic: the 160-question
head-to-head against RAG, GraphRAG and Opus 5.5 on one store; the 1,600-item
standard evaluation (90.29); 10 GB of Wikipedia indexed in 17 min vs 62 to
97 h to embed; per-million and per-month costs at list prices fetched
2026-11-05. Known limits are stated in the post: counting needs subject_id
and is set counts only, a count does not check the question's noun, large
namespaces are slow, the model over-declines.

Also: pagedata entry (published 2026-11-05), blog index card, a cover motif
drawn from the post's own head-to-head counts (blog_covers wrong160), og /
hero / thumb renders. Generators run with --write; every --check, seo_lint
(0 gate violations), test_article_schema, test_urlmap,
test_verify_agent_discovery, test_llms_field_report and the landing
constants guard pass.
@cla-bot cla-bot Bot added the cla-signed label Nov 5, 2026
@xerj-org
xerj-org merged commit 71d60a4 into main Nov 5, 2026
29 checks passed
YYW0228 pushed a commit to YYW0228/xerj that referenced this pull request Nov 7, 2026
The public half of the v1.0.0-rc.102 cut. The release tag goes on this
PR's merge commit, so the notes and the tree they describe are one commit.

- CHANGELOG.md: [Unreleased] cut over to [1.0.0-rc.102] - 2026-11-06.
  Every PR merged since v1.0.0-rc.101 is cited (xerj-org#2969 xerj-org#2982 xerj-org#2983 xerj-org#2984
  xerj-org#2985 xerj-org#2988 xerj-org#3023 xerj-org#3028 xerj-org#3029 xerj-org#3030) or exempted with its reason (site,
  blog and donation pages: xerj-org#2971 xerj-org#2972 xerj-org#2976 xerj-org#2978 xerj-org#2980 xerj-org#2990 xerj-org#3032;
  CI test-only: xerj-org#3031). The release-notes gate passes locally against
  origin/main 34a2f24.
- The XERJ Query-1 numbers in the summary trace to the R8 evaluation
  (private harness, runs/r8_2026-11-06/R8_RESULT.md): Score v2 on the four
  development families, 1,600 items, seed 47 90.29 -> 94.69 (+4.40,
  95% interval +3.32 to +5.64), seeds 11/29 +3.95/+4.17; stated as
  in-sample because R8 was chosen on those items; wrong answers 9 -> 10
  of 1,600 (s47), 12 -> 15 (s11); held-out fresh_v1 Net 0.3333 -> 0.3481.
  The 1,600/1,600 engine-vs-Python parity was re-run on the merged rc.102
  build tree.
- Upgrade notes say plainly that the rc.101 promise to land the storage
  and ingest lanes on public main before rc.102 was not kept, and that the
  rc.102 bundle needs an rc.102 binary.
- Fixed lists the rc.101 aarch64 Linux assets built with a 4 KiB jemalloc
  page (checked in the rc.101 build tree: "checking LG_PAGE... 12" for
  both aarch64 targets), which xerj-org#1663 had fixed in release.yml; rc.102's
  native aarch64 builds set JEMALLOC_SYS_WITH_LG_PAGE=16.
- ROADMAP.md: review line names v1.0.0-rc.102 (the
  docs_capability_lists check); the open-defects shortlist re-verified
  live: xerj-org#1122 and xerj-org#1170 open.
- engine/Cargo.toml workspace version 1.0.0-rc.102, Cargo.lock restamped
  (19 workspace packages, no dependency moved).
- Landing footer stamps V1.0.0-RC.101 -> V1.0.0-RC.102 (59 pages) and the
  nav's fallback latest-tag label; landing-constants-guard passes.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

0