Changelog

All notable changes to anthroposophy.ai are documented in this file.

The format follows Keep a Changelog. Versioning follows Semantic Versioning.

1.1.4 - 2026-07-06

Research workflows

This release serves quote-level verification — the byte-for-byte quotation discipline as an API operation rather than a client-side re-implementation — and fixes the bare-/mcp connector failure.

• Verify a quotation before printing it. verify_quote on a candidate inherited from the secondary literature returns near_miss with a word-level diff against the closest corpus window and the verbatim span at its locus — an inherited misquote is caught in one call instead of propagated.
• Audit a document’s quote ledger in one call. Batch mode takes up to 50 quote+locus pairs and returns per-item verdicts plus an all_quote_safe summary — a published artifact’s “verified” claim can be the API’s output rather than the author’s promise.
• Locate a quote whose wording differs across editions. Without a locus, verify_quote locates candidates corpus-wide and checks every (lang, edition) witness, so a quotation taken from any indexed translation resolves to its locus.
• Review an elision’s surroundings in one call. fetch_passage with window=N returns the passage plus up to N neighbors per side under context.before / context.after — checking that an ellipsis in a spliced quotation hides nothing meaning-inverting no longer costs one call per neighbor.

Discipline: only quote_safe: true clears a candidate for quotation marks. match_normalized means the wording matches but the typography does not — quote match.verbatim_text exactly as returned, never your candidate. near_miss means the corpus reads differently; quoting the candidate propagates drift. A locus-scoped no_match is authoritative for that locus; a corpus-search no_match is strong evidence of absence, not proof.

Added

• Tool verify_quote: quote-level verification. verify_locus proves a locus exists; this proves a quotation matches. Single mode takes candidate with optional scope (ga + lecture (+ passage_number), lang, edition); batch mode takes quotes[] (max 50) and returns per-item results plus a summary with all_quote_safe. Verdicts, strongest first: match (byte-for-byte substring of a witness; quote-safe), match_normalized (matches after typographic normalization only — quote style, dash family, whitespace runs, emphasis markers; not quote-safe; match.verbatim_text carries the exact span to quote instead), match_spliced (candidate contains ellipses; every fragment matched in order within one lecture; quote-safe only when every fragment matched byte-for-byte), near_miss (closest corpus window, with similarity and a word-level differences diff; near_miss.corpus_text carries the verbatim span), no_match. Without a locus the check locates candidates corpus-wide via full-text search (exact phrase first, then bag-of-words with 0-hit relaxation); a locus-scoped no_match is authoritative for that locus, a corpus-search no_match is strong evidence of absence, not proof. All (lang, edition) witnesses are checked by default, so a quotation from any indexed translation can be located. REST equivalent: POST /v1/steiner/verify-quote. Deployed 2026-07-06.
• fetch_passage parameter window (integer 0-5, default 0): returns up to N neighboring passages per side — same lecture, same lang + edition — under context.before / context.after, each with verbatim text, locus, and URL. One call to review the surroundings of a passage, e.g. the text elided by an ellipsis in a spliced quotation. Also on the REST passage endpoints as ?window=N. Deployed 2026-07-06.

Changed

• MCP tools/list returns 19 tools (5 GA + 7 FMC + 5 sources + 2 claim-registry).
• Server instructions extended: candidate quotations — especially those inherited from secondary literature or model memory — should be checked with verify_quote before being placed inside quotation marks; only quote_safe: true clears a candidate.
• MCP server version and SERVER_VERSION bumped to 1.1.4.
• Per-IP rate limiting is now enforced on both the REST and MCP surfaces (app-level token bucket, ~240 requests/minute/IP). Over-limit requests receive 429 Too Many Requests with a Retry-After header. This is generous flood protection, not a precise quota — a considerate client rarely encounters it; batch where the surface supports it (verify_quote ledger mode) and prefer targeted fetches over sweeps.
• Search boolean operators are now normalized operator-aware: query operands are stemmed and case-folded while uppercase OR/AND/NOT/NEAR, quoted phrases, prefix *, and column filters are preserved. Previously the query normalizer lowercased the whole query, demoting FTS5’s uppercase-only operators to literal terms, so an OR union became an implicit-AND and could return an authoritative zero where the individual terms matched — a manufactured verbatim-absence. A 0-hit now reports absence_signal all_alternatives_absent for an OR union (every alternative absent) distinctly from exact_phrase_absent for a phrase/AND/NEAR co-occurrence miss. Related surface additions: two_voice_volume on search hits and fetch_passage (with an attribution_warning on the passage) flags correspondence volumes where a passage may be a letter by a correspondent rather than by Steiner; edition_coverage ({lang: [editions]}) accompanies any single-ga search and fetch_passage, and a 0-hit in a language a volume lacks returns absence_signal=language_not_available rather than topical absence; fields=compact (REST ?fields=compact) drops per-hit available_versions[] and edition metadata for large result sets; and ga with from_date/to_date is documented as the scoping recipe for locating a passage within a long lecture. Deployed 2026-07-07.

Fixed

• MCP transport: requests to the bare /mcp path (no trailing slash) received 307 Temporary Redirect to /mcp/. MCP clients that do not follow redirects never completed initialize, so connectors configured with the bare URL attached no tools. Both /mcp and /mcp/ now serve the streamable-HTTP transport directly, with no redirect hop. The redirect had been present since the MCP endpoint first shipped; it surfaced when connector-side redirect handling changed. Deployed 2026-07-06.

1.1.3 - 2026-06-11

Research workflows

This release adds schema-set enumeration and documents the claim-registry tools deployed 2026-06-04.

• Enumerate the FMC schema set. list_fmc_schemas pages the full set in deterministic id order (total_count plus offset/limit echo). Schemas were previously retrievable only individually by FMC00.NNN id; no enumeration path existed.
• Filter schemas by topic or id range. query matches description text and linked-page slugs (descriptions are sparse upstream; linked pages carry most of the topical signal). prefix scopes an id range, e.g. prefix="FMC00.5".
• Image retrieval handoff. Entries with image_only: true carry no usable description; pass the same schema_id to fetch_fmc_schema_image for the diagram as inline ImageContent. fetch_fmc_schema remains the single-schema detail record.
• Trace registered claims. trace_claims(node) returns registered claims touching a concept or work: typed relations with up to three verbatim anchors each (leg_a, leg_b, witness), every anchor carrying its locus, edition, content hash, and fetch URL. mode="divergence" restricts results to divergence claims.
• Re-verify a claim before asserting it. verify_claim(claim_id) re-runs a registered claim against the serving corpora: each anchored leg is checked by content-hash (registry integrity) and re-fetch (corpus drift). Per-leg verdicts: anchor_present, anchor_missing, hash_mismatch, fetch_error.

Discipline: list_fmc_schemas entries are curatorial_content — cite as ‘FMC schema FMC00.NNN (FMC curator, anthroposophy.eu, CC-BY-SA)’; descriptions are paraphrase and never go inside quotation marks. For claims, quote only anchor_text, byte-for-byte, and report relation types as given. node_not_found from trace_claims means no claims are registered for that node, not that the corpora are silent — use search_steiner / search_sources for frontier research. The registry’s selection principle (steinerian-anthroposophical) is named in every envelope.

Added

• Tool list_fmc_schemas: enumeration and filtered discovery over the FMC schema set. Filters: prefix (schema-id prefix), query (case-insensitive substring over description text and linked-page slugs). Per-schema fields: schema_id, description (null when none exists), schema_image_url, linked_pages, page_url, and image_only (the signal to call fetch_fmc_schema_image). Deterministic schema_id ordering with total_count and offset/limit echo. Closes the schema-discovery gap: the wiki’s schema-listing page is outside the mirror’s indexed namespace, so fetch_fmc_page could not reach it.
• Tool trace_claims (deployed 2026-06-04; documented here): registered claims touching a concept or work — typed relations across the corpora (Steiner GA, Bible, Aquinas, Church Fathers, Bhagavad Gita, and the other /sources/ works). Each claim carries a relation type (direct_citation, commentary_on, doctrinal_appropriation, thematic_affinity, ambient_influence, doctrinal_contrast, interpretive_continuation, complementary_polarity), a mode_tag (resonance | divergence), a statement, and up to three verbatim anchors (leg_a, leg_b, witness) with loci, editions, content hashes, and fetch URLs. relation_source identifies the asserting layer (curated / registered / citation_index / summary_engagement); machine-derived edges are weaker evidence than curated ones. The registry is compiled from reviewed claim files and deployed as an artifact; no write path is exposed on the public surface.
• Tool verify_claim (deployed 2026-06-04; documented here): live re-verification of a registered claim. Each anchored leg is checked by content-hash against the registry (integrity) and re-fetch against the serving corpora (drift). Per-leg verdicts: anchor_present, anchor_missing (treat the claim as unverified), hash_mismatch (registry integrity failure), fetch_error (infrastructure, not drift). ok=true means every anchored leg verified at verified_at. Claim ids come from trace_claims envelopes (edges[].claim_id).

Changed

• MCP tools/list returns 18 tools (4 GA + 7 FMC + 5 sources + 2 claim-registry). Corrects the 1.1.2 entry, which stated 14; the count at 1.1.2 was 15 (search_anthroposophy omitted from its FMC arithmetic).
• MCP server version and SERVER_VERSION bumped to 1.1.3. The server version field now references the SERVER_VERSION constant instead of carrying a second literal; the two values had drifted apart twice before 1.1.2.

1.1.2 - 2026-05-23

Research workflows

This release adds the /sources/ corpus. It is curated by Steiner-engagement, not by canon — these are the foundational texts the GA itself takes up.

• From a Steiner lecture to its source text. find_source_chapters_for_ga_locus("GA 142, 1912-12-28") returns the Sources chapters the curator has linked to that lecture (Bhagavad Gita Discourses, Mahabharata Parvas, the Pauline epistles as the GA 142 counterpart text). fetch_source_chapter then quotes the verses verbatim.
• From a foundational text to Steiner’s treatment of it. fetch_source_summary on the Kaṭha Upanishad or De Anima Book II names the verified GA loci where Steiner engages it; fetch_passage returns his actual words.
• Cross-corpus surveys in a single call. search_anthroposophy(query=..., corpus="all") fans GA + Free Man Creator (FMC) + Sources together, distinguishing hits by [ga] / [fmc] / [sources] tags with per-corpus totals.
• Honest gap-detection. When the curator has examined a chapter and Steiner did not engage it, steiner_status: "not_engaged" marks the absence explicitly. Aristotle’s De Anima carries this flag.
• Stream-filtered comparative work. search_sources(stream="indian") scopes to Vedic / Upanishadic / Gita / Mahabharata material; stream="greco-christian" to Patrologia + Aristotle + Aquinas + Hermetic Corpus; stream="western-european" to the Beguine + Rhineland + Theosophical + Goethean strands.
• Curator’s cross-tradition congruences as named syntheses. Each summary surfaces the FMC curator’s explicit cross-tradition observations (e.g. De Anima’s tripartite soul ↔ Vedanta’s pancha-kosha ↔ Plotinus’s nous-triad ↔ Kabbalah’s nefesh / ruach / neshamah) as paraphrase the curator owns, never as quotes from those traditions.

Discipline: only fetch_source_chapter is quote-safe (quote_safe=true). fetch_source_summary, find_source_chapters_for_ga_locus, and every steiner_loci.statement field are FMC-curatorial paraphrase (quote_safe=false) — they capture the verified loci but they never go inside quotation marks.

Added

• /sources/ corpus integration — Sacred, Inspired and Foundational Texts across 5 wisdom-streams (Indian, Persian, Egyptian-Hebrew, Greco-Christian, Western-European). Vedic scripture, the Greek-philosophical tradition (Aristotle, Plato, Plotinus), Patrologia (Church Fathers), Aquinas (Summa Theologiae + Summa Contra Gentiles), Beguine + Rhineland mysticism, Hermetic Corpus, Rosicrucian Manifestos, Theosophical inheritance (Isis Unveiled, Secret Doctrine, Mahatma Letters), Goethean works.
• Tool search_sources: full-text search over /sources/ chapters. Filters: work, lang, edition, stream, tradition, form, year_from, year_to. Returns ranked chapter hits with chapter_locus, snippet, curatorial blurb, and summary_url (when present) pointing at the FMC-curatorial paraphrase block.
• Tool fetch_source_chapter: retrieve verbatim chapter text by chapter_locus (<work>[/<subwork>...]/<chapter>). Returns text_trust='verbatim', quote_safe=true, parent breadcrumbs, descriptor subtitle + blurb, and available_versions[]. Use to obtain quote-safe text from /sources/.
• Tool fetch_source_summary: retrieve the FMC-curatorial paraphrase block for a chapter or work — theme, soul-faculty, Steiner engagement (paraphrase + verified GA loci, or the honesty-gap sentinel “not engaged in the GA corpus”), cross-tradition congruences. text_trust='paraphrase', quote_safe=false. The structural verbatim-vs-paraphrase split that prevents accidental quotation of curator-summary content.
• Tool list_source_works: discovery — enumerate /sources/ works by stream, tradition, form, parent_locus, top_level_only, has_summary. Returns work_locus, name, tradition, year_approx, chapter_count, word_count, landing_url, summary_url.
• Tool find_source_chapters_for_ga_locus: reverse cross-reference (sources → GA). Given a GA locus (GA 142, 1912-12-28) or volume (142), returns which /sources/ chapters cite that lecture, with paraphrase statements. Driven by the sources_ga_xref bridge, populated from chapter and work summaries.
• search_anthroposophy umbrella extended with corpus='sources' and corpus='all'. 'all' fans GA + FMC + sources in a single call with round-robin interleave; 'both' remains the back-compat default (GA + FMC). Response carries sources_total alongside ga_total and fmc_total.
• Tool fetch_fmc_schema_image: returns inline ImageContent for FMC schemas, mirror-first with upstream fallback. Resolves the image-fetch gap where MCP clients could not follow MediaWiki redirects to the upstream JPEGs.

Changed

• Top-of-document Quotation discipline (read first) block in server instructions extended with a /sources/ paragraph. fetch_source_chapter returns quote-safe text (text_trust='verbatim'); fetch_source_summary returns FMC-curatorial paraphrase (text_trust='paraphrase', quote_safe=false) — never quote, even when paraphrasing is tight. Each steiner_loci entry inside a summary names a verified GA locus retrievable via fetch_passage; the surrounding statement is the curator’s paraphrase, not Steiner’s words. find_source_chapters_for_ga_locus returns paraphrase as well.
• search_anthroposophy.corpus enum: ['ga','fmc','both'] → ['ga','fmc','sources','both','all']. Tool description acknowledges sources as a third corpus alongside GA and FMC.
• MCP tools/list returns 14 tools (4 GA + 5 FMC + 5 sources).
• MCP server version and SERVER_VERSION bumped to 1.1.2. Both were stale prior to this release (1.1.0 and 1.0.0 respectively).

Fixed

• scripts/deploy_site.sh parity probe: added retry/backoff (8 attempts, 0.5s × attempt#) to tolerate uvicorn’s multi-worker spawn window. systemctl is-active returned active before all workers had bound to port 8044, causing a false rollback on the 2026-05-22 deploy.
• v2/server/main.py: removed a stale import sources_shared that crashed all 4 anthro-api workers with ModuleNotFoundError on startup (the module is on the MCP service’s path, not the API service’s). The /sources/ REST endpoints remain deferred per the MCP-first pattern; the import was dead code.
• find_fmc_pages_for_ga_locus.inputSchema: removed the top-level anyOf that required either ga_locus or ga_volume. Anthropic’s custom tool API rejects top-level anyOf / oneOf / allOf in input_schema; any agent call that included this tool in its toolset was failing with 400 tools.N.custom.input_schema: input_schema does not support oneOf, allOf, or anyOf at the top level. The field-presence constraint is now enforced in the dispatch handler instead (which already validated the same rule and returned error_code: "missing_argument"). A regression test (test_no_top_level_schema_combinators) sweeps every FMC tool schema for top-level combinators.

1.1.1 - 2026-05-18

Research workflows

This release adds the quotation-discipline layer for corpus citations.

• Anything inside quotation marks must be byte-for-byte from a corpus retrieval. fetch_passage, fetch_source_chapter, or a quoted_span record with status=stored — minor trimming, smoothing, or dropping a single word counts as paraphrase, even when self-labelled as quoting.
• 0-hit responses carry an authoritative “do not quote” signal. search_steiner / search_anthroposophy / search_sources return corpus_verdict: "NO_MATCH — ..." in the response payload when total == 0. The verdict is authoritative: the queried text is not in the indexed corpus.
• Bare GA-volume-and-year citations are rejected as unverified. A citation such as “GA 26, 1924” without a specific lecture-date or passage-number is not a verified citation.
• Reverse-lookup questions get the dedicated tool. “Which FMC pages cite GA 146, 1913-06-02?” routes to find_fmc_pages_for_ga_locus, not generic FTS — exact-locus questions answered by an exact-locus tool.

Citations in agent output therefore reflect retrieval rather than training memory; a “not in corpus” answer is authoritative.

Added

• corpus_verdict field on search_steiner and search_anthroposophy responses when total == 0. Carries an authoritative “do not quote” signal in the response payload, so the agent reads it directly rather than relying only on session-level instructions.

Changed

• New top-of-document Quotation discipline (read first) block in server instructions. Requires search_steiner → fetch_passage before opening quotation marks. 0-hit corpus results are authoritative. Bare GA-number-plus-year citations (e.g. “GA 26, 1924”) rejected as unverified. Byte-for-byte means byte-for-byte: dropping a single word (‘thus’, ‘then’, ‘also’) or trimming for slide-readability breaks verbatim status; self-labeling a paraphrase doesn’t exempt it.
• Additional agent-discipline blocks in server instructions: inline citation rendering at parity with web-source linking; section references must use <page_slug>#<section_anchor> form; reverse cross-reference questions must use find_fmc_pages_for_ga_locus; stay inside the corpus when verify_locus returns true; citation follow-ups must verify both page and section anchor before answering.
• Tool find_fmc_pages_for_ga_locus: description lead rewritten around the question pattern it handles (“which FMC pages cite this lecture?”); routes away from search_anthroposophy for reverse-lookup queries.
• Tool search_anthroposophy: closing “When NOT to use this tool” block redirects exact-locus reverse lookups to find_fmc_pages_for_ga_locus.
• Shared FMC source-discipline contract (in search_anthroposophy + fetch_fmc_page descriptions): closing Quotation-marks discipline sentence reserves quotation marks for quoted_span records with status='stored' or verbatim Steiner passages from fetch_passage.

Fixed

• find_fmc_pages_for_ga_locus.inputSchema now declares anyOf requiring ga_locus or ga_volume. Was permissive on {}.
• Removed dead corpus argument from search_steiner / fetch_passage / fetch_parallel dispatch. Handlers pass corpus="steiner" as a literal. fetch_parallel.inputSchema no longer advertises the single-value corpus enum.

1.1.0 - 2026-05-14

Research workflows

This release adds the Free Man Creator (FMC) curatorial layer (anthroposophy.eu — pages, schemas, structured quoted-span records, and GA cross-references) to the MCP surface alongside the GA.

• Reverse cross-reference (GA → FMC). find_fmc_pages_for_ga_locus("1913-06-02-GA146") returns the FMC pages citing a GA locus, with attribution class (adjacent / section-inferred / main-body-reference).
• Curatorial page retrieval. fetch_fmc_page("Krishna_impulse") returns ordered curatorial chunks (lead, sections, subsections, discussion notes) plus quoted_span records with provenance.
• Schema retrieval. fetch_fmc_schema("FMC00.389A") returns description, image URL, linked pages, and cited GA loci.
• Link-graph traversal. traverse_anthroposophy_links runs BFS over the FMC cross-link graph (depth 1-3, configurable max-nodes).

FMC content is CC-BY-SA paraphrase; quoted_span records carry the curator’s structured provenance (status=stored for inline verbatim, status=referenced for GA pointers that resolve via fetch_passage). The unified search_anthroposophy umbrella ranks GA + FMC hits together for queries spanning both.

Added

• FMC (anthroposophy.eu) corpus integration. The curatorial wiki at anthroposophy.eu is now retrievable through the MCP connector alongside the Gesamtausgabe. License: CC-BY-SA.

• Tool search_anthroposophy: unified full-text search across the GA and FMC corpora. Filters: corpus, chunk_class, target_corpus, status, page_slug. Inherits GA filters (lang, edition, ga, ga_range, from_date, to_date). Detects ranking mode from query content (primary_seeking, comparative, curatorial, default) and adjusts interleave order accordingly.

• Tool fetch_fmc_page: retrieve an FMC page by slug. Returns ordered curatorial chunks (lead, sections, subsections, discussion notes), quoted_span records with provenance, and the page’s GA cross-references.

• Tool fetch_fmc_schema: retrieve a schema by FMC00.NNN id. Returns description, image URL, linked pages, and cited GA loci.

• Tool traverse_anthroposophy_links: BFS over the FMC link graph from a root page slug. Depth clamped to [1, 3]; configurable max_nodes.

• Tool find_fmc_pages_for_ga_locus: reverse cross-reference. Given a GA locus (1913-06-02-GA146) or volume (146), return all FMC pages that cite it with attribution class (adjacent, section-inferred, main-body-reference).

• Source-discipline contract enforced in tool response structure. quoted_span records with status=referenced carry no verbatim text; agents follow target_locus to the primary corpus. Records with status=stored carry the full embedded_attribution chain. All FMC retrieval results carry primary_anchor (corpus, page_url, curator, license) and content_hash for citation stability.

Changed

• MCP tools/list returns 9 tools (4 GA + 5 FMC).

1.0.0 - 2026-05-07

Research workflows

Initial release. The complete Gesamtausgabe (GA 1 through GA 354, public domain) behind a single MCP surface.

• Search the full corpus by keyword. search_steiner runs FTS5 with Porter stemmer + diacritic normalization. Filters: lang, edition, ga, ga_range, from_date, to_date.
• Verify a GA reference exists before citing. verify_locus confirms (ga, lecture_date) is in the corpus — useful as a citation gate before publishing or quoting.
• Retrieve any passage verbatim. fetch_passage(ga, lecture, passage_number, lang, edition) returns text + stable citation URL in either German original or available English translation.
• View parallel translations side-by-side. fetch_parallel returns multiple editions of the same locus; per-edition translator + year + basis are surfaced, and an alignment_confidence flag warns when paragraph counts diverge between source and translation.

German originals and English translations are cross-retrievable per locus. Browseable HTML at /v1/steiner/<lang>/ga/<vol>/<date>/p<NNN>, with content-negotiated JSON via the .json suffix. The MCP surface is anonymous and per-IP rate-limited at the edge; no account required.

Added

• Initial public release.

• Gesamtausgabe corpus (GA 1 through GA 354). Rudolf Steiner’s books, lectures, and lecture cycles. Public domain.

• Multilingual retrieval. English translations alongside German originals. Per-edition tracking (translator, year, basis). alignment_confidence flag marks paragraph-count divergence between source and translation.

• MCP server at https://anthroposophy.ai/mcp. Streamable HTTP transport per MCP spec 2025-03-26. Anonymous access; per-IP rate limiting at the edge.

• Tool search_steiner: FTS5 search over the Gesamtausgabe. Porter stemmer with diacritic normalization. Filters: lang, edition, ga, ga_range, from_date, to_date, corpus.

• Tool verify_locus: confirm a GA reference exists in the indexed corpus before citing.

• Tool fetch_passage: retrieve a verbatim Steiner passage by (ga, lecture, passage_number, lang, edition). Returns text and a stable citation URL.

• Tool fetch_parallel: retrieve the same passage across multiple translation editions side-by-side.

• REST API at https://anthroposophy.ai/v1/.

• Browseable passages at /v1/steiner/<lang>/ga/<ga_volume>/<lecture_date>/p<NNN>. Content negotiation returns HTML to browsers, JSON to agents via the .json suffix.