LLM Changelog / Errata

Official leaderboard disclosure updated on 2026-09-14: four OpenAI official records, including one superseded historical record; the recorded promotion/signoff occurred on 2026-07-07. Review was delegated AI review plus human batch signoff, not per-item human reading. Claude/Gemini/Daoism-Qwen targets remain pending; DeepSeek participated as a grader, not a target.

Compatibility: Description correction only; no score, target, or passed/blocking gate logic changes.; breaking change: no.

Documented the existing llm-rag-v3 additive contract: contract_version, retrieval_mode, dense, fusion, hit lexical_rank/dense_rank/fused_score, and retrieve mode aliases hybrid/lexical/v3/v2. Corrected dense Hamming-to-float ordering before RRF, failure kinds, embedding dependency, and answer n=1..6 (default 4).

Compatibility: Legacy api_version=llm-rag-v2 and response mode=node_search_trigram remain unchanged. Retrieval values and request/response execution are unchanged; published schemas now describe existing v3 fields.; breaking change: no.

Recorded the public read-only MCP server at /mcp (Streamable HTTP): daoist_search, daoist_answer, daoist_cite, and daoeval_question. Corrected MCP envelope limit to 60 requests/minute per IP, distinct from upstream RAG 30/minute; source metadata is available only when results provide it.

Compatibility: Documents the existing server; no tools, limits, transport behavior, or retrieval behavior changed.; breaking change: no.

Added LIUS governance v1: node citation packages, LIUS-ID/version/content hashes, review status layers, public errata ledger endpoint, audit bundle table, and paper_entity_link evidence fields.

Compatibility: additive; breaking change: no.

Added public provenance, citation graph, changelog/errata, self-host guide, and offline deterministic scorer surfaces.

Compatibility: additive; breaking change: no.

Published governance, evaluation gate, DaoEval manifest, RAG contract, schemas, and expanded OpenAPI/capabilities discovery.

Compatibility: additive; breaking change: no.

Clarified that DaoEval v0.1 is a preprint-style method/data artifact and that formal scholarly claims require peer-reviewed follow-up.

Compatibility: additive; breaking change: no.

Aligned public count policy: 400 routeable public questions, 330 core-v0.1 main leaderboard denominator, 0 official verified runs.

Compatibility: additive; breaking change: no.

Before governance v1, public node citation exposed a stable URL and updated date but not a machine-readable LIUS-ID, content hash, review layer, and citation package in one object.

LIUS governance v1 adds lius_node_governance, node citation package API, RAG hit governance fields, and paper_entity_link evidence/review fields.

A RAG citation means a published LIUS node/source field was returned. It is not yet a claim-level edition/page verification certificate.

Use citation_status values source_present, node_published, source_verified, and human_verified; formal scholarship must verify primary editions separately.

Some public copy could be read as saying 500 questions are already public, while the current routeable benchmark has 400 questions and the core denominator is 330.

Public wording now distinguishes design target, routeable public table, and core-v0.1 leaderboard denominator.

Capabilities listed generate output field as text, while the route returns output.

Capabilities now lists request_id, output, citations, context_blocks, and retrieval_policy.