Log

Append-only history. Each entry: ## [YYYY-MM-DD] <op> | <title> where <op> is ingest, query, or lint. Query with grep "^## \[" log.md | tail -5.

[2026-08-30] ingest | Opus 5 vs the founding substrate (research pass, synthesis-only)

Model: claude-opus-5 · effort: high. Curator's question: the wiki now runs on Opus 5 — how does that compare with the model it started on? Hunted one source at the bar, claude-opus-5-system-card (Anthropic, T1), ingested cross-wiki in ../llm-providers-wiki on the Sonnet-5-card precedent. No new page here — claude-opus-5 already exists in ../llm-providers-wiki and duplicating the name would make every cross-wiki link to it ambiguous. What changed here is the reading:

• synthesis.md → Model substrate rewritten. The founding substrate was claude-opus-4-8 (the corpus's oldest pages are dated the day after its 28 May release), then claude-fable-5 from 9 June, then Opus 5 from 24 July. Against 4.8, Opus 5 saturates the procedural-honesty evaluations (lazy investigation — first Claude model to score perfectly; overconfidence before a state-changing command) and regresses on unaided recall (+11% accuracy, +6% hallucination on AA-Omniscience). That split argues for the provenance rule rather than against it: a claim carrying [[source]] links is checked against a document, which is the mode that improved.
• Two cautions recorded: illegible thinking rose over 4.8 while unfaithful thinking fell, and verbalized evaluation awareness fell while white-box analysis still found unverbalized grader awareness and fabricated data — so behavioural audits are a weaker proxy than they were, inverting the trend the claude-sonnet-5-system-card flagged.
• claude-opus-4-8 corrected: it said "it is also the model maintaining this wiki," which stopped being true in June. It keeps two live roles — safety fallback, and the baseline every later honesty number is stated against.
• The "runs on the guardrailed Fable tier" line in the safety paragraph updated: Opus 5 ships under the same ASL-3 / CB-1 treatment as Opus 4.8 and is explicitly below Fable 5 in overall capability, so the tier question resolved toward the cheaper, more permissive rung. Left open on purpose: whether Opus 5 holds a long maintenance session as orchestrator rather than worker. The card's less-character-drift finding points the right way; the best-known commentary on the question is an unevidenced practitioner argument (T4), which the research pass's own bar excludes. Recorded as a visible gap rather than filled below the bar.

[2026-06-25] ingest | sift-kg — document→knowledge-graph CLI (hub-routed)

Hub-routed Telegram drop (github.com/juanceresa/sift-kg, msg 596). T3 (vendor README, no independent eval). Clear cluster-A fit — a packaged, domain-agnostic CLI that LLM-extracts a typed knowledge-graph from arbitrary documents (the two-pass schema-then-populate idea of ontologies-knowledge-graphs-ai, shipped as a tool). Created sift-kg (SoftwareApplication) and juan-ceresa (Person, thin). Two synthesis contributions: (1) its human-approved deduplication is the augment-pole counterpart to gbrain's unattended dream-cycle dedup — the augment→automate axis now has a both-ends pair on the same operation (entity resolution); (2) its property-graph (NetworkX, non-RDF) typed edges mark a middle between open-knowledge-format (no typed edges) and the RDF/OWL tradition (../knowledge-representation-wiki). Cross-linked knowledge-graph + ontologies-knowledge-graphs-ai. Runner-ups: ../agentic-tooling-wiki (agent-memory framing) + ../knowledge-representation-wiki (the KG/property-graph contrast). +2 pages.

[2026-06-05] ingest | Building Agent Memory with Knowledge Graphs (hub-routed)

Hub-routed Telegram drop (The Neural Maze, Substack). Clear cluster-A fit (extends knowledge-graph + the RAG-critique thread); runner-up ../agentic-tooling-wiki (agent memory is an agent-building concern) declined — the substance is knowledge-graph / retrieval architecture, and gbrain already bridges agent-memory to that spoke. Ingested:

• New source agent-memory-knowledge-graphs (BlogPosting, url-only): temporal KG agent memory via graphiti (Zep) over Neo4j; bi-temporal modeling; "graphs quietly replacing RAG" for agents.
• New concept temporal-knowledge-graph (DefinedTerm) — bi-temporal facts; the temporal-validity gap; extends knowledge-graph with a time axis.
• New tool graphiti (SoftwareApplication) — Zep's incremental temporal-KG framework; the agent-memory analog of gbrain's self-wiring graph.
• Updated knowledge-graph (added the temporal turn + sources frontmatter), retrieval-augmented-generation (added the fourth retrieval gap — temporal validity — to the BM25/graph/no-accumulation taxonomy), and synthesis (retrieval critique now four clean gaps; agent memory pulled toward a first-class A-core subject + the live agentic-tooling seam).
• Updated index (source + temporal-knowledge-graph + graphiti). ~6 pages touched. Cross-spoke: ../agentic-tooling-wiki is the runner-up; carried via the existing gbrain bridge, no dup. Site rebuilt + verified. Created directory tree, CLAUDE.md schema, and spine files (index, log, synthesis). No sources ingested yet.

[2026-05-29] ingest | LLM Wiki (Karpathy gist)

Source: raw/llm-wiki.md (fetched from gist.github.com/karpathy/442a6bf...). Created 8 pages: source summary llm-wiki-gist; concepts llm-wiki, retrieval-augmented-generation, memex; people andrej-karpathy, vannevar-bush; software obsidian, qmd. Updated synthesis (first thesis

• 5 open questions) and index (grouped by @type). No contradictions (single source).

[2026-05-29] ingest | As We May Think (Bush, 1945)

Source: raw/as-we-may-think.md (full text, W3C copy of the July 1945 Atlantic essay). New pages: source summary as-we-may-think; concepts associative-trails, microfilm. Enriched memex and vannevar-bush from the primary source (both previously flagged thin). Synthesis: resolved the "memex-ancestor faithfulness" open question and recorded a tension — the gist says Bush "couldn't solve who maintains the trails," but the essay assigns it to human "trail blazers"; the LLM Wiki's real claim is that human trail-building doesn't scale. Added Engelbart/NLS as a noted gap. Index updated.

[2026-05-29] ingest | Claude for Financial Services (repo, via Telegram)

Source: raw/claude-financial-services.md (README of github.com/anthropics/financial-services), delivered over Telegram. Per the new Telegram auto-ingest rule, ingested without a topic-fit gate. New pages: source summary claude-financial-services (SoftwareSourceCode); model-context-protocol, claude-cowork, claude-managed-agents. Opens a new topic cluster B (agentic LLM products / FSI), off the main knowledge-management thesis. Bridged to cluster A honestly via model-context-protocol (qmd is an MCP server; the repo uses MCP connectors) — also cross-linked from qmd. Synthesis: added a "Topic clusters" section and an open question on whether B should split into its own wiki. No contradictions.

[2026-05-29] lint | full health check (3 sources, 15 pages)

Clean: no orphans, no broken links (16/16 resolve), no unresolved contradictions. Applied 1 fix: added missing associative-trails back-link from llm-wiki (the pattern's wikilinks are associative trails, but the link was one-directional). Recommendations logged for user decision: (1) ingest an Engelbart/NLS source — the biggest lineage gap in cluster A; (2) retype claude-managed-agents from SoftwareApplication → WebAPI (more specific); (3) thin/self-flagged pages needing a dedicated source — retrieval-augmented-generation; (4) tools-for-thought comparison gaps (Zettelkasten, Roam) mentioned but unpaged; (5) decide whether cluster B (FSI) splits into its own wiki.

[2026-05-29] ingest | GBrain (Garry Tan) — via Telegram

Source: raw/gbrain.md (README of github.com/garrytan/gbrain). Major on-thesis source. New pages: source summary gbrain (SoftwareSourceCode); garry-tan (Person); knowledge-graph (DefinedTerm). Enriched retrieval-augmented-generation (first concrete benchmark: +31.4 P@5 from graph over vector-only RAG), llm-wiki (added a "compared to gbrain" spectrum note), qmd, associative-trails, model-context-protocol. Synthesis substantially advanced: thesis moved from single-source advocacy to TWO independent instantiations (one in production); framed a maintenance/automation spectrum (llm-wiki ↔ gbrain); noted gbrain's schema packs as independent corroboration of typed pages; updated A/B clusters (gbrain bridges them via MCP + Cowork). Resolved/clarified the index-scale and RAG-comparison open questions. No contradictions; one reinforced framing.

[2026-05-29] ingest | Godot Game Template — via Telegram

Source: raw/godot-game-template.md. Off-thesis (game dev). New pages: source summary godot-game-template + godot-engine, kept as ISOLATED cluster C with no forced links to A/B. Flagged in synthesis as the prime split-out candidate. Per Telegram rule, ingested without a topic gate.

[2026-05-29] ingest | Agentic SEO Skill — via Telegram

Source: raw/agentic-seo-skill.md. LLM agent skill pack (16 sub-skills, 10 agents, 89 scripts) for IDEs. New pages: source summary agentic-seo-skill + new concept agent-skills (capability as portable markdown). The concept retroactively connects claude-financial-services and gbrain (both ship markdown skillpacks) — added inbound links from both. Synthesis: agent-skills logged as a genuine sub-theme of cluster B; raised the "is this still one wiki?" question (3 clusters now).

[2026-05-29] ingest | Godot XR Community Game Jam V — via Telegram

Source: raw/godot-xr-game-jam-v.md (godotengine.org blog post, Bastiaan Olij). Cluster C (game dev). New page: source summary godot-xr-game-jam-v (BlogPosting), linked to godot-engine (added a backlink there so it's not an orphan). Cluster C now 3 pages, internally coherent, still isolated. Per Telegram rule, ingested without a topic gate.

[2026-05-29] ingest | 14 KB landing page (dev.to) — via Telegram

Source: raw/landing-page-14kb.md (web-performance minimalism article). New page: source summary landing-page-14kb (BlogPosting). No honest connection to any existing cluster — deliberately left as an INTENTIONAL ORPHAN (isolated cluster D) rather than force-linked, per the "don't fabricate connections" principle. Per Telegram rule, ingested without a topic gate. NOTE: orphan-by-design; expected to be flagged by lint.

[2026-05-29] ingest | Introduction to Lean for Programmers — via Telegram

Source: raw/lean-for-programmers.md (TDS article, Ronen Lahat; captured via WebFetch — direct fetch Cloudflare-blocked 403). New pages: source summary lean-for-programmers (BlogPosting) + lean-theorem-prover (SoftwareApplication, with Curry-Howard/dependent types/tactics/Mathlib folded in). New cluster E (formal methods) — but NOT isolated: established an honest bridge to as-we-may-think via the Leibniz→Gödel "mechanizing reasoning" lineage (Bush foresaw formal-logic machines). Synthesis: added cluster E + a meta-observation that A, B, and E all descend from "mechanizing human thought" (Bush + Leibniz as shared roots) — first real thread linking an off-thesis cluster to the core.

[2026-05-29] ingest | Three Spec-Driven AI Tools — via Telegram

Source: raw/spec-driven-ai-tools.md (ranthebuilder.cloud, Itzhak Eretz Kdosha). Cluster B. New pages: source summary spec-driven-ai-tools (compares BMAD / Spec-Kit / OpenSpec; OpenSpec highest overall, BMAD best for course-correction) + concept spec-driven-development (DefinedTerm). Tools folded into the summary rather than 3 stub pages (promote later if they recur). Linked to agent-skills; extended the B sub-theme to "process-as-markdown." Noted reflexive parallel to this wiki's own brainstorm→spec→plan build. No contradictions.

[2026-05-29] ingest | Hetzner Cloud (product page) — via Telegram

Source: raw/hetzner-cloud.md (vendor marketing page). New page: source summary hetzner-cloud (WebPage). No honest connection to any cluster — intentional orphan, isolated cluster F (cloud infrastructure). Per Telegram rule, ingested without a topic gate; left unlinked rather than fabricate a cloud-hosting tie to gbrain.

[2026-05-29] ingest | Oracle Cloud Free Tier Signup — via Telegram

Source: raw/oracle-cloud-signup.md. Client-rendered signup SPA — fetch returned NO substantive content ("enable JavaScript to run this app"). Created a thin, honest source page oracle-cloud-signup (WebPage) that records the no-content limitation rather than back-filling Oracle facts from general knowledge. Paired with hetzner-cloud to make cluster F a coherent "cloud providers" pair (added mutual links). Per Telegram rule, ingested without a topic gate.

[2026-05-29] ingest | Best Godot Plugins 2025 — via Telegram

Source: raw/godot-best-plugins-2025.md (godotawesome.com listicle). Cluster C. New page: source summary godot-best-plugins-2025 (BlogPosting), linked to godot-engine and godot-game-template. Cluster C now 4 pages — the largest off-thesis island; flagged as prime split candidate. No individual plugin pages (listicle detail). Per Telegram rule, ingested without a topic gate.

[2026-05-29] ingest | LLM Wiki Agent (SamurAIGPT) — via Telegram

Source: raw/llm-wiki-agent.md. Strongly on-thesis (cluster A): an MIT coding-agent skill implementing the llm-wiki pattern — near-identical to THIS wiki (raw/→wiki/, index/ log/overview synthesis, ingest/query/lint, auto entity/concept pages, wikilink graph). New page: source summary llm-wiki-agent (SoftwareSourceCode). Cross-linked into llm-wiki (new "Implementations" section), gbrain, knowledge-graph (its two-pass EXTRACTED/INFERRED graph), agent-skills. Synthesis: added a 4th source and new point #3 — the design is a "natural attractor" (Karpathy spec + gbrain + llm-wiki-agent + this wiki = 4 independent convergent realizations). Reflexive: the wiki ingested a description of itself. No contradictions.

[2026-05-29] split | spun off off-thesis islands into their own wikis

Moved clusters C/D/F out of this wiki, each into a standalone wiki with full structure (CLAUDE.md + index/log/synthesis + raw/ + wiki/):

• game-dev (C) → ~/projects/godot-wiki (3 sources, 4 pages)
• web-performance (D) → ~/projects/../webperf-wiki (1 source, 1 page)
• cloud providers (F) → ~/projects/../cloud-wiki (2 sources, 2 pages) Removed their entries from index.md and the cluster C/D/F sections from synthesis.md; cleaned cluster-framing notes in the migrated pages and removed a now-dangling gbrain link from the migrated hetzner-cloud page. This wiki retained clusters A (knowledge-mgmt core), B (agent products/skills), and E (formal methods) — both B and E keep honest bridges to A. Research-wiki now 8 sources / 25 pages.

[2026-05-29] ingest | Claude Opus 4.8 review (Simon Willison) — via Telegram

Source: raw/claude-opus-4-8-review.md. Routed to research-wiki (not an island): the model is the substrate under the whole ecosystem tracked here. New pages: source summary claude-opus-4-8-review (BlogPosting) + claude-opus-4-8 (SoftwareApplication, with specs/pricing + the honesty improvement). Linked to claude-cowork, claude-managed-agents, gbrain, llm-wiki-agent. Synthesis: added a "Model substrate" note — 4.8's flagship honesty gain (≈4× fewer unsupported claims) reinforces the "maintenance is near-free" bet by reducing the worst failure mode (confident fabrication). Reflexive: the model reviewed is the one maintaining this wiki. No contradictions.

[2026-05-29] lint | health check (9 sources, 27→28 pages)

Clean: no orphans, no broken links, no unresolved contradictions. Applied: (1) retyped claude-managed-agents SoftwareApplication → WebAPI (more specific; new index section); (2) created alphaproof — referenced 4× across the Lean pages but unpaged (missing-page fix), linked from both Lean pages; bridges cluster E to the AI-agent/ honesty thread. Recommendations (need a fetch/decision): ingest an Engelbart/NLS source (still the top lineage gap, flagged 2 lints running); a neutral retrieval-augmented-generation source; expand thin single-source pages (claude-cowork, microfilm, claude-opus-4-8). Optional: a Simon Willison person page if he recurs.

[2026-05-29] ingest | Augmenting Human Intellect (Engelbart, 1962) — chased per lint

Source: raw/augmenting-human-intellect.md (WebFetch of dougengelbart.org full text; HTML too long for verbatim). Fills the lint's #1 gap — the Bush→modern lineage middle. New pages: source summary augmenting-human-intellect (Report) + douglas-engelbart (Person). Wired the lineage: forward-links from memex and associative-trails (trails → Engelbart's hypertext → wikilink). Synthesis: inserted Engelbart into the through-line, added the augment→automate axis and the bootstrapping↔self-improving- wiki tie, and resolved the Engelbart open question. HONESTY FLAG: could not confirm an explicit in-text Bush citation in the 1962 report — recorded the influence as documented history but the in-document quotation as unverified (noted on both pages and in synthesis). Zettelkasten/Roam remain the open tools-for-thought gap.

[2026-05-29] ingest | Zettelkasten Method (zettelkasten.de) — chased per lint

Source: raw/zettelkasten-introduction.md (WebFetch of zettelkasten.de/introduction). Fills the Zettelkasten tools-for-thought gap. New pages: source summary zettelkasten-introduction (Article) + zettelkasten (DefinedTerm) + niklas-luhmann (Person). Framed in synthesis as a parallel ancestor to Bush/Engelbart — a third independent convergence on association-over-categorization — and, via Luhmann's "communication partner" framing, the closest historical anticipation of the LLM-as-knowledge-partner. Linked to associative-trails, memex, obsidian, llm-wiki, knowledge-graph. Resolved the Zettelkasten open question; Roam/Obsidian-as-tool/Notion remain the modern-PKM gap. Honesty: omitted the widely-cited ~90k-slip figure since this source didn't state it. No contradictions.

[2026-05-29] ingest | Roam Research (The Sweet Setup guide) — chased per lint

Source: raw/roam-research-guide.md (WebFetch; SitePoint was 403, used The Sweet Setup). New pages: source summary roam-research-guide (Article) + roam-research (SoftwareApplication). Positioned as the modern tools-for-thought operationalization: bidirectional links + backlinks-with-context + personal knowledge-graph; its wikilink→auto-page→backlink mechanic IS the llm-wiki's linking UX, and its "Unlinked References" prefigure automated association. Cross-linked from knowledge-graph, obsidian, zettelkasten. Synthesis: added the "modern tools-for-thought wave" paragraph and the full arc (Bush→Engelbart→Luhmann→Roam→LLM). HONESTY: this source covers Roam's features + a 1960s-hypertext lineage but does NOT discuss Obsidian/Logseq — attributed the launch year (2019) and "inspired other tools" to general knowledge, flagged on the page. Resolved the Roam open question; Logseq/Tana/Notion remain.

[2026-05-29] ingest | TiddlyWiki5 — via Telegram

Source: raw/tiddlywiki5.md (repo README, HTML-stripped). On-thesis (cluster A), routed to research-wiki. New page: source summary tiddlywiki5 (SoftwareSourceCode). Positioned as the earliest tool-for-thought node (~2004, predates Roam ~15 yrs): self-contained single-file personal wiki, atomic "tiddlers," transclusion, own-your-data ethos that prefigures the llm-wiki plain-files-you-own model. Linked to zettelkasten, roam-research (added backlink), obsidian, llm-wiki, associative-trails. Synthesis: prepended TiddlyWiki to the tools-for-thought-wave paragraph. No contradictions.

[2026-05-29] ingest | Build an orchestration mode (Claude API doc) — via Telegram

Source: raw/orchestration-mode.md. Cluster B, routed to research-wiki. New pages: source summary orchestration-mode (TechArticle) + concept agent-orchestration (DefinedTerm). Concept ties together the orchestrator→subagent fan-out across claude-managed-agents (leaf-workers + handoff loop), gbrain (Minions queue), and this doc; built on claude-opus-4-8's mid-conversation system messages (added backlinks from both that page and managed-agents). Synthesis: added agent-orchestration as cluster B's execution-time sub-theme (counterpart to spec-driven-development's authoring-time; the concrete "automate" end). Reflexive: same fan-out pattern the maintaining agent uses. No contradictions.

[2026-05-29] ingest | Logseq (repo README) — chased per lint

Source: raw/logseq.md. New page: source summary logseq (SoftwareSourceCode). Framed as the clearest convergence of the two lineage strands: Roam's networked-thought UX + the own-your-plain-files ethos (privacy/longevity/user control + local Markdown/Org) = the llm-wiki's substrate choice. Linked to roam-research, obsidian (backlink added), tiddlywiki5, zettelkasten, knowledge-graph. Synthesis: added Logseq + the "networked notes as plain markdown you own" convergence point. HONESTY: README covers KM/privacy/plain-files but not the block-outliner/bidirectional-links/journals details — attributed those to general knowledge, flagged on the page. Tana/Notion remain the gap.

[2026-05-29] ingest | Notion (Wikipedia) — chased per lint

Source: raw/notion-wikipedia.md (neutral, third-party — also addresses the lint's wish for non-vendor sources). New pages: source summary notion-wikipedia (Article) + entity notion (SoftwareApplication). Framed as the counterpoint that bounds cluster A's convergence claim: the mass-market winner (20M+ users, $500M+ ARR) is cloud/proprietary, structured-DB, not local-plain-files — so "own your plain files" is a deliberate value choice, not an inevitability (a reason the LLM-wiki picks plain markdown on purpose). Synthesis: added the Notion counterpoint paragraph. Mild cluster-B bridge (Notion's 1M+ Custom Agents). Tana now the only remaining PKM-tools gap. No contradictions.

[2026-05-29] ingest | Tana (supertags guide) — chased per lint (final PKM gap)

Source: raw/tana-supertags-guide.md (third-party; Tana's own docs redirected/404'd). New pages: source summary tana-supertags-guide (Article) + entity tana (SoftwareApplication). Framed as the closer: Tana fuses the two poles (networked nodes of roam-research + structured fields of notion), is AI-native (LLM uses your typed graph as context — the llm-wiki/gbrain thesis inside a mainstream PKM tool), and its supertags are a third independent convergence on typed notes (with gbrain schema packs + this wiki's schema.org @type). Synthesis: strengthened the typed-pages corroboration to three-fold and added a closing paragraph — the tools-for-thought lineage and the LLM-as-knowledge-partner thesis are now the same story. PKM-tools gap CLOSED. No contradictions.

[2026-05-29] lint | health check (17 sources, 43→44 pages)

Clean: no orphans, no broken links (45/45 resolve), no unresolved contradictions. Hubs healthy (llm-wiki 30, gbrain 27, memex/associative-trails 14). Applied: created tools-for-thought — the phrase recurred 9× across pages with no page; now the hub for the PKM lineage (linked from synthesis, roam-research, tana, index). Recommendations (unchanged, need a fetch/decision): Engelbart→Bush in-text citation still unverified; retrieval-augmented-generation still wants a neutral non-advocate source; thin single- source pages remain (claude-cowork, microfilm). Noted but not actioned: "NLS" recurs 8× folded into douglas-engelbart — promote to its own page only if a dedicated source arrives. Optional Simon Willison person page if he recurs.

[2026-05-29] ingest | Compound Engineering plugin (Every Inc) — via Telegram

Source: raw/compound-engineering-plugin.md. Cluster B, routed to research-wiki. New pages: source summary compound-engineering-plugin (SoftwareSourceCode) + concept compound-engineering (DefinedTerm). Key synthesis link: compound engineering ("each unit of work makes the next easier; codify learnings") = douglas-engelbart's bootstrapping applied to software-with-agents, and the engineering sibling of the llm-wiki/gbrain compounding-artifact thesis — plus it's literally the build loop of this wiki (brainstorm→spec→plan→execute→compound). Linked to agent-skills, spec-driven-development, agent-orchestration, claude-cowork; backlinks added from engelbart + agent-skills. No contradictions.

[2026-05-30] ingest | Claude-to-Speech (Claude Code TTS plugin) — via hub router

Source: claude-to-speech (SoftwareSourceCode, github.com — no raw copy, url-only). Routed in by the hub router (../HUB.md) as the best-fit spoke; cluster B. A small community Claude Code plugin (ElevenLabs TTS via invisible markers + a Stop hook + /speak). Filed as a lighter facet of cluster B — agent I/O tooling (capability-as-files on the output/UX side), sibling to compound-engineering-plugin, linked to agent-skills and claude-cowork. Off the main knowledge-management thesis; notable as evidence the Claude Code plugin surface attracts small third-party tools. No contradictions.

[2026-05-30] ingest | Anthropic valuation + Opus 4.8 launch (two Anthropic sources) — via hub router

Two co-arriving Anthropic stories, both routed here on the model-substrate thread:

• anthropic-valuation (Gizmodo) — Anthropic ~$965B > OpenAI ~$852B; revenue ~doubled to ~$9B on Claude Code; but huge compute commitments + unproven profitability.
• claude-opus-4-8-launch-tomsguide (Tom's Guide) — launch piece, honesty framing; content-thin stub (body wouldn't fetch — site chrome only), kept as corroboration. New pages: anthropic (Organization, anchors the substrate-provider thread) + the two source summaries. Updated claude-opus-4-8 (maker link, 2nd-outlet corroboration) and the synthesis Model substrate note — added the cost/economics data point as the mirror of the capability/honesty upside. Borderline routing (valuation is business news): chose research-wiki over a park because it lands on the existing "viability = capability and cost" thread, not a forced fit. No contradictions; flagged substrate cost as a risk.

[2026-05-31] lint | health check (18 sources, 50 pages)

Part of a hub-wide lint. Excellent structural health: 533 links across 50 pages (~10.7/page), zero orphans, zero broken links (the "[[double bracket]]" hit is escaped example text in roam-research/roam-research-guide, not a real link), consistent Thing-vs-source two-page splits (e.g. notion/notion-wikipedia, zettelkasten/zettelkasten-introduction, roam-research/roam-research-guide). Findings: (1) synthesis.md is now ~240 lines and the densest spine in the hub — still coherent, but approaching the size where it should be tightened/sectioned before the next few ingests; (2) open questions needing a neutral source persist — a non-advocate RAG/graph-RAG comparison and any longitudinal drift/maintenance-cost measurement; (3) clusters B (agentic products) and E (formal methods) remain bridged to A by honest links, not forced — healthy. No contradictions beyond the one already-flagged Bush "who maintains the trails" tension. No fixes applied.

[2026-06-01] ingest | A Developer's Guide to Building ADK Agents with Skills (Google)

Routed in by the hub router → cluster B (agent skills); clear match, no runner-up. URL-only (Google Developers Blog), so the summary carries url:. New: the open agentskills.io spec (agentskills-spec) and progressive disclosure (L1 metadata ~100 tok → L2 body <5k tok → L3 resources; vendor-claimed ~90% baseline-context cut), with Google's adk as a second-vendor adopter — one skill across "Gemini CLI, Claude Code, Cursor, 40+ products." Created adk-agents-with-skills (BlogPosting, source), agentskills-spec (DefinedTerm), adk (SoftwareApplication), google (Organization, tracked lightly like anthropic). Updated agent-skills with a "per-tool convention → open standard" section and the ADK instance. Synthesis cluster B now records the standardization shift and frames progressive disclosure as the procedural twin of the llm-wiki's index→drill-down. Caveat: vendor blog, token figures unverified.

[2026-06-01] ingest | Agentic coding harnesses + deployment axis (agent-kanban, agentsys, channels-vs-openclaw)

Three cluster-B sources routed in together by the hub (all clear matches, no runner-ups).

• agent-kanban (SoftwareSourceCode) — shared Kanban board for multi-agent + human coding collaboration; Ed25519 agent identity, mutual PR review, per-repo skills, multi-runtime.
• agentsys (SoftwareSourceCode) — modular runtime: 24 plugins / 49 agents / 44 skills; phase gates, certainty grading, tool-not-token determinism. Claims Sonnet+harness > raw Opus.
• claude-code-channels-vs-openclaw (BlogPosting) — event-driven (Channels) vs self-driven (OpenClaw) Claude agents = the augment→automate axis at the deployment layer. New concept agentic-coding-harness (the scaffolding around a code-writing model — "everything else"; structure as a substitute for capability) ties agent-kanban + agentsys + compound-engineering together. Updated agent-orchestration and agent-skills with the new instances; synthesis cluster B now carries the harness sub-theme + deployment axis. Flagged in synthesis that cluster B is growing toward its own potential agentic-tooling spoke (a human call). Caveats: all README/vendor/opinion claims, unbenchmarked.

[2026-06-01] ingest | Agent-tooling burst, part 2 (best-practices, PPC skills, claw-code, conductor, agent-starter-pack, gstack)

Six more cluster-B sources from the same Telegram burst, all clear matches:

• claude-code-best-practices (BlogPosting) — BMAD vs plan mode, focused CLAUDE.md, model selection, community skills; "the tool matters less than the person."
• claude-skills-ppc (Article) — Claude Skills for PPC; skills × MCP = agency; twin of agentic-seo-skill.
• claw-code (SoftwareSourceCode) — provider-agnostic Rust CLI agent harness (minimal end).
• conductor (SoftwareSourceCode) — Gemini CLI Context-Driven Development; extends spec-driven-development.
• agent-starter-pack (SoftwareSourceCode) — Google Cloud templates/eval/deploy (operate end); maintenance mode → agents-cli.
• gstack (SoftwareSourceCode) — Garry Tan's software factory; convergence node composing skills + orchestration + conductor + gbrain. Updated concepts agentic-coding-harness (now spans plan→build→deploy spectrum), spec-driven-development (Context-Driven Development + BMAD), agent-skills (PPC + gstack). Synthesis cluster B escalated the spin-out flag: ~10 agent-tooling sources this burst now form the wiki's largest body — recommend a dedicated agentic-tooling spoke (human call). Caveats: README/vendor/practitioner claims throughout, unbenchmarked.

[2026-06-01] split | cluster B (agentic tooling) spun out → ../agentic-tooling-wiki (human directive)

Migrated the agent products/skills/harnesses/orchestration pages (24 pages + 5 raw docs) to the new sibling ../agentic-tooling-wiki; this wiki now holds clusters A (knowledge management) and E (formal methods). Kept the bridge concepts here — agent-skills (rewritten as the A/B seam: capability-as-markdown is the procedural cousin of the llm-wiki's knowledge-markdown), compound-engineering, model-context-protocol, gbrain — all now linked cross-wiki to the moved instances. Updated index + synthesis (cluster-B → split pointer; A/B-divide, model- substrate, and split-out box reframed) and the CLAUDE.md domain header. Cleared dangling sources: refs on agent-skills / compound-engineering / model-context-protocol (their raw docs moved). The super-thesis still spans both spokes; the split is organizational.

[2026-06-01] lint | post-split health check

Lint of the trimmed wiki after the cluster-B split. Fixes: reworded anthropic and claude-opus-4-8 so references to the migrated claude-cowork/claude-managed-agents are framed as cross-wiki (../agentic-tooling-wiki) rather than local, and dropped stale "(cluster B)" tags; tidied frontmatter comments on compound-engineering/model-context-protocol (removed [[ ]] from YAML comments). Orphans: none. Link check passes (cross-wiki bridges resolve).

[2026-06-03] ingest | Claude Opus 4.8: Capabilities and Reactions (Zvi)

Routed in by the hub → the model-substrate thread (existing claude-opus-4-8). Created claude-opus-4-8-zvi (BlogPosting) — a critical third source. Benchmarks (SWE-bench Pro 64.3→69.2, USAMO 96.7%, GDPval Elo 1890) + polarized reception (refreshingly-honest/strong-writing vs neurotic/over-hedging/refusals, anti-sycophancy overcorrection). Key: complicates the wiki's honesty thesis — "performative honesty," deception dropped "only from fear of detection," confident- fabricate-then-retract. Updated claude-opus-4-8 (capabilities & reception section) and synthesis (model-substrate: added the honesty-is-partial-and-performative tension → raises the value of the wiki's own fabrication/lint checks). Cross-linked llm-benchmarks (../llm-providers-wiki). Caveat: opinion/commentary; benchmark figures as reported.

[2026-06-03] ingest | What an Enterprise Context Layer Actually Is (Prukalpa/Atlan)

Routed in by the hub → cluster A (knowledge management), the enterprise-scale instantiation of the wiki's thesis. Created enterprise-context-layer (BlogPosting): turning knowledge/expertise/ norms into machine-usable context for AI; three context types (knowledge=semantic map/knowledge-graph, expertise=playbooks/agent-skills, norms=policy); five capabilities incl. learning loops + governance against knowledge decay; thesis "the 10th agent is smarter than the 1st" = compound-engineering / llm-wiki / gbrain compounding bet as enterprise infrastructure. Updated compound-engineering (org-scale instance) and synthesis (lineage now spans personal → agentic → enterprise scales of one idea: an LLM compounding a governed knowledge substrate). Considered the parked knowledge-representation cluster but rejected — Prukalpa explicitly distinguishes it from semantic layers/catalogs; it's the compounding-org-knowledge thesis, not a vocabulary/ontology. Caveat: vendor-founder framing.

[2026-06-03] ingest | Knowledge as a Service for Azure Logic Apps (Microsoft) — honest stub, provisional

Routed in by the hub; fetch failed (JS-rendered MS Community Hub → title only). Created azure-logic-apps-knowledge-service (WebPage, honest stub). Inferred: managed enterprise-knowledge / RAG-grounding (retrieval-augmented-generation) for Logic Apps workflows + their AI actions — a productized enterprise-context-layer (knowledge-substrate side → cluster A). Low confidence: if a successful fetch shows it's primarily an agent/workflow-building feature, re-route to ../agentic-tooling-wiki. Flagged for refresh.

[2026-06-03] ingest | Knowledge as a Service — Wikipedia

Routed in by the hub → cluster A. Created concept knowledge-as-a-service (DefinedTerm, url: Wikipedia): cloud knowledge-delivery backed by a knowledge model; the context-exploitation distinction from DaaS (user + information context) — i.e. the older generic statement of the enterprise-context-layer bet. Grounds the azure-logic-apps-knowledge-service stub (supports its knowledge-substrate routing) and links knowledge-graph/retrieval-augmented-generation. Noted the semantic-web/ontology adjacency to the parked knowledge-representation cluster (KaaS leans on knowledge graphs + ontologies). Updated the Azure stub to reference the now-anchored concept.

[2026-06-03] ingest | Why Vector Search Alone Isn't Enough: Hybrid Retrieval for RAG (InfoQ)

Routed in by the hub → the RAG thread. Created hybrid-retrieval-rag (Article) — the vendor-neutral RAG source the retrieval-augmented-generation page had flagged as missing. Hybrid retrieval: dense vectors + BM25 fused via RRF (k≈60) + optional cross-encoder rerank; vector-alone collapses distinguishing tokens (IDs/error codes/flags); hybrid queries dominate production (Perplexity/Glean). Independently corroborates gbrain's hybrid retriever, separating that engineering design from GBrain's motivated benchmark. Updated retrieval-augmented-generation (hybrid-retrieval section; split the RAG critique: BM25=exact-token gap, knowledge-graph=factual-connection gap, llm-wiki=no-accumulation gap; softened the "no neutral source" note). Cross-linked knowledge-as-a-service.

[2026-06-09] broaden | domain widened to add cluster F (diffusion & adoption of ideas/technologies)

User directive (via hub) to broaden the domain so the parked InfoQ adoption-curve piece could be ingested here. Added cluster F — the dynamics by which ideas & technologies diffuse and get adopted (diffusion of innovation, technology-adoption curves, trend analysis) alongside the KM/tools-for-thought core (A) and formal methods (E). Updated CLAUDE.md domain header, synthesis (new cluster-F bullet + bridge framing: the adoption curve is itself a tool-for-thought, and the cluster is reflexive vs the wiki's own trend-curation), index (synthesis cluster note + new rows), and the hub wikis.md block (domain/keywords/sample-pages/note). Same play as the game-engines (2026-06-07) and tts→speech-audio (2026-06-05) broadens.

[2026-06-09] ingest | The Technology Adoption Curve, Twenty Years On (InfoQ)

Founding source of cluster F. New source tech-adoption-curve-twenty-years (Article, url; InfoQ 20th-anniversary editorial, Losio & Synodinos) + new concept technology-adoption-curve (DefinedTerm; Rogers' diffusion of innovation). The article places 20 years of dev tech (Agile, SOA, cloud, DevOps, K8s, microservices, ML, AI/agentic) on the curve + five 2036 predictions ("reliability engineering for AI becomes its own discipline"; agentic systems following microservices' over- application arc). Bridged to the core via tools-for-thought and the reflexive trend-spotting rhyme with llm-wiki/gbrain; agentic/SRE threads noted as cross-spoke adjacencies (not duplicated). Caveat: a publication's self-referential anniversary editorial; qualitative curve placements. research-wiki 48 → 50 pages.

[2026-06-09] ingest | Cluster F expansion — diffusion/adoption canonical frameworks (router-curated)

User asked to "expand cluster F." Router web-searched + curated 3 neutral (Wikipedia) sources for the cluster's missing canonical frameworks (same router-curation pattern as the llm-providers/speech-audio spin-outs). Added 7 pages: sources diffusion-of-innovations-wikipedia, gartner-hype-cycle-wikipedia, crossing-the-chasm-wikipedia; concepts crossing-the-chasm (Moore's chasm + beachhead/whole- product) and gartner-hype-cycle (Fenn/Gartner expectations curve + critiques); persons everett-rogers and geoffrey-moore. Enriched technology-adoption-curve into the cluster hub (Rogers' four elements, S-curve, five adoption attributes, %s, pro-innovation-bias critique). Updated synthesis: expanded cluster-F bullet (three frameworks + "adoption share × hype sentiment") and a new tension (Moore's discontinuous chasm vs Rogers' continuous variable; hype-cycle's weak empirical basis). Index: new DefinedTerm/Person/source rows. Cluster F now 9 pages (was 2). research-wiki 50 → 57.

[2026-06-09] ingest | +2 spaced repetition (cluster A) + TLA+ (cluster E) — all-spokes cron test

spaced-repetition (DefinedTerm, src — expanding-interval review; SuperMemo/Anki/FSRS; the internalize-memory branch complementing the externalize PKM tools) and tla-plus (SoftwareApplication, src — Lamport's spec language + TLC model checking; AWS/MS use; the verify-systems wing of formal methods alongside Lean/AlphaProof). Folded into synthesis clusters A and E. Wikipedia url-only. 57 → 59 pages.

[2026-06-10] ingest | Bass diffusion (F) + Rocq (E) + RAG primary source (A) — all-spokes pass

Three sources, one per active cluster. bass-diffusion-model (DefinedTerm, source, Wikipedia) — Frank Bass (1969), the quantitative counterpart to Rogers' curve: dF/dt=(1−F)(p+qF), coefficient of innovation p≈0.03 + imitation q≈0.38, S-curve forecasting. Sharpens cluster F's continuous-vs-discontinuous tension into a formal statement — Bass is continuous/aggregate with no built-in chasm, so Moore's discontinuity is exactly the regime the basic model omits. rocq (SoftwareApplication, source, Wikipedia) — Rocq (formerly Coq, renamed March 2025 v9.0), the 35-yr INRIA CIC proof-assistant peer to lean-theorem-prover; verified-software legacy (CompCert, Four Color Theorem, Feit–Thompson, BB(5); 2013 ACM Software System Award). Makes cluster E two-system. rag-original-paper (ScholarlyArticle, source, arXiv) — Lewis et al. (FAIR, 2020), the neutral primary source for retrieval-augmented-generation the synthesis kept wanting; updated the RAG concept page (canonical-definition section + softened the advocate caveat) to cite it. Confirms the wiki's RAG-gap taxonomy extends the 2020 design (which claimed only provenance+updatability), not refutes a strawman. Folded into synthesis (new 2026-06-10 section) + index (new ScholarlyArticle group + DefinedTerm/SoftwareApplication rows). No contradictions. 59 → 62 pages.

[2026-06-15] ingest | Claude Fable 5 release + suspension — InfoQ

T4 (InfoQ, trade press). New material: US government export directive forced Fable 5 offline within 3 days of launch (Amazon→White House jailbreak flag; david-sacks spokesperson). Also: 30-day retention vs Microsoft zero-retention friction. Created claude-fable-5-infoq (source) and david-sacks (Person entity). Updated claude-fable-5 with suspension section. Synthesis updated (government-as-availability-lever, cross-spoke ../ai-governance-wiki seam flagged). Runner-up: ../ai-governance-wiki (US export directive is a regulatory action).

[2026-06-15] ingest | Ontologies, Knowledge Graphs, and AI — Mysore / Medium

T3 (Medium, practitioner). New angle for cluster A: the formal ontology layer (schema above the knowledge graph) and the "schema as guardrail" finding — constraining LLM knowledge-graph population with the LLM-drafted ontology schema reduces hallucination. Maps onto agent-guardrails in ../agentic-tooling-wiki (same containment discipline applied to knowledge extraction). Created ontologies-knowledge-graphs-ai (source), ontology (new DefinedTerm), vishal-mysore (Person entity); updated knowledge-graph to add the ontology layer. Synthesis updated (cluster A extension). The formal-ontology lineage (W3C OWL/RDF/Semantic Web) is an E↔A seam worth watching. Hub route entry to follow.

[2026-06-10] ingest | Claude Fable 5 (Simon Willison) — model-substrate update

Hub-routed (Telegram; runner-up ../llm-providers-wiki). New source claude-fable-5-review (BlogPosting, url, Simon Willison) + model page claude-fable-5 (SoftwareApplication) — anthropic's 9 Jun 2026 frontier model, same author/shape as claude-opus-4-8-review so it extends the model-substrate thread. Key facts (per the review): 1M context, 128K output, Jan-2026 cutoff, $10/$50 per Mtok = 2× Opus 4.8, free on Max through 22 Jun then metered; "substantially larger," slower, deeper knowledge (accurate project dates); MicroPython→CPython-WASM + Datasette Agent coding feats. Framed as a guardrailed sibling of "Claude Mythos 5" (capability-equal, more-restricted variant). Two synthesis hooks: (1) the substrate cost trend reverses — Opus 4.8 held 4.7 pricing, Fable 5 doubles it → raises pressure on the near-free-maintenance bet (anthropic-valuation economics flag); (2) the safety story shifts from honesty (4.8) to guardrails. Updated claude-opus-4-8 (superseded-in-role pointer) + synthesis (substrate-update paragraph) + index. Runner-up ../llm-providers-wiki carried via the cross-wiki llm-benchmarks/llm-api-pricing bridges, not duplicated. Caveats: single hands-on, post-cutoff, dated snapshot, Fable/Mythos naming is the source's. 62 → 64 pages.

[2026-06-10] ingest | Claude Fable 5 + Mythos 5 (Anthropic primary announcement) — substrate refresh

Hub-routed (Telegram); the primary anthropic source for the models the claude-fable-5-review flagged as missing. New source claude-fable-5-mythos-5-announcement (Article, url) + new model page claude-mythos-5 (SoftwareApplication); refreshed claude-fable-5 in place (per re-seen rule) and updated the review's caveat. Key reveal: Fable 5 and Mythos 5 are the identical model — safeguards are the only difference. Fable = "Mythos-class made safe for general use" via three safety classifiers (cyber / bio-chem / distillation) that fall back to claude-opus-4-8 (>95% sessions no fallback); Mythos = same model ungated for trusted users (Project Glasswing/US-gov, cyber & bio partners). Pricing $10/$50 confirmed (2× Opus 4.8, but "<½ Mythos Preview"); benchmarks qualitative (Stripe 50M-line Ruby migration in a day; Hebbia finance top score; vision SOTA / Pokémon FireRed); 30-day retention on Mythos-class traffic; alignment "low, similar to Opus 4.8." Synthesis updated: safety has moved from a model disposition (4.8 honesty) to a deployment policy (classifiers + trusted-access gating) over one capability ceiling — making "which substrate tier, under what policy" an explicit variable for the auto-maintained wiki. Index + caveats updated. Specs (1M ctx/128K/Jan-2026) still only from the review (not restated by Anthropic). Vendor announcement; dated snapshot. 64 → 66 pages.

[2026-06-12] ingest | Isabelle/HOL — the third major proof assistant (cluster E)

All-spokes daily expansion. Added isabelle (@type SoftwareApplication) — Isabelle/HOL (Cambridge/TU Munich), joining lean-theorem-prover + rocq to make cluster E a three-system field. Distinct design point: generic LCF-style framework over classical higher-order logic (vs Lean/Rocq's CIC dependent types); signature mechanisms Isar (declarative proofs) + Sledgehammer (external ATP/SMT → kernel reconstruction = automation feeding a verified core, ATP cousin of alphaproof); the seL4 microkernel proof gives the prove-the-implementation counterpart to tla-plus's check-the-design — bracketing systems verification (live seam to ../platform-ops-wiki). Reinforces the Leibniz→Bush mechanizing-thought meta-thread (E↔A). synthesis "three-system field" note; index cluster-E updated. 1 new page. Caveat: terse homepage; seL4/AFP/Sledgehammer facts are well-established encyclopedic, grounded in the official project.

[2026-06-15] lint | azure-logic-apps-knowledge-service — stub upgraded + routing resolved

Resolved the low-confidence honest-stub (Community Hub page JS-gated, title-only). Content recovered via WebSearch (the announcement + Build-2026 coverage): Knowledge as a Service / KBaaS = a fully managed RAG/knowledge layer in Azure Logic Apps (docs → managed ingestion/chunking/embedding/vector-store/ retrieval, "zero pipeline to operate"), grounding agents + workflows. Provisional routing RESOLVED → stays research-wiki (cluster A): it's the knowledge-substrate side (managed RAG

• knowledge-as-a-service productized, an enterprise-context-layer data point), not an agent-builder. Added the "RAG plumbing is commoditizing" synthesis hook. Removed the provisional/stub warnings; tier stays T3 (vendor announcement) but content-confirmed; index entry refreshed. No longer a stub.

[2026-06-18] ingest | DefinedTerm enrichment pass (subagent)

Thinnest-first deepening of DefinedTerm pages with fetched authoritative sources.

• ontology (was thinnest, only a Medium source) — added Gruber's 1993 "specification of a conceptualization," the axiom test distinguishing ontology from taxonomy, components (individuals/classes/attributes/relations), OWL/RDF/CycL, domain vs upper ontologies. New source ontology-information-science-wikipedia (T2, Wikipedia).
• model-context-protocol (had sources: [] after the FSI source migrated out) — grounded in the protocol's own docs: USB-C framing, host/client/server, JSON-RPC data layer + transport layer, server primitives (Tools/Resources/Prompts), client primitives (Sampling/Elicitation/Logging), stdio + Streamable HTTP. New source mcp-spec-introduction (T1, modelcontextprotocol.io).
• zettelkasten + niklas-luhmann — closed the long-flagged unsourced-slip-count gap: ~90,000 cards (1952–53 onward), ~50 books/550 articles, branching index-number addressing (insert-between without renumbering), pre-Luhmann origins (commonplace books, Gessner, 1640s Harrison cabinet). New source zettelkasten-wikipedia (T2). 3 source pages created; 4 Thing pages enriched. index.md updated (added a WebPage-sources section). No build/verify, no git — per task constraints.

[2026-06-22] ingest | Open Knowledge Format (OKF) — Google Cloud spec

New source page open-knowledge-format (T1 — primary, the spec itself; GoogleCloudPlatform/knowledge-catalog). OKF is an open standard for knowledge/metadata as a bundle of markdown files with type frontmatter, reserved index.md/log.md, prose cross-links, git-diffable, human + agent readable — "metadata as code." It is, almost line for line, a vendor-published spec of the llm-wiki substrate this wiki runs on. Folded into cluster A and the typed-pages convergence in synthesis: OKF is the fifth independent arrival at typed markdown pages (after Karpathy's gist, gbrain schema packs, tana supertags, notion DBs, and this wiki's @type) and the first from a hyperscaler as a named standard. It also stakes the anti-formalist pole of KR — deliberately no ontology / taxonomy / typed edges, relationships left as prose for an LLM — the opposite of the RDF/OWL semantic-web tradition, a clean A↔../knowledge-representation-wiki seam (runner-up spoke). Third thread: joins enterprise-context-layer (Atlan) + azure-logic-apps-knowledge-service (Microsoft) as a hyperscaler knowledge-substrate point, but inverts them — plain files you own vs. a managed service that hides the substrate. Updated llm-wiki (new "Standardized" section), synthesis (five-fold convergence rewrite + cluster-A list), index (new TechArticle group). Publisher cross-linked to existing google node (../llm-providers-wiki) — no dup org page. Fetched live via WebFetch. (No build/commit here — orchestrator handles.)

[2026-06-23] ingest | When Software Started Writing Software — A Developer's History of AI

New source page developers-history-of-ai (T4 — dev.to opinion retrospective, Adam "the Developer", ~2,500 words). A 70-year AI history told as a "which layer of human work got automated" arc (symbolic → statistical ML → deep learning → transformers/LLMs → agentic). Routed to cluster F as the genre-sibling of tech-adoption-curve-twenty-years (both dev-tech historical retrospectives; both land agentic systems in the innovator/early-adopter band). Its keeper thesis — "each leap automated a layer of mechanism and left the judgment layer exactly where it was" (judgment moves upstream) — is a clean external articulation of the wiki's mechanizing-thought meta-observation and the augment→automate axis, so folded into both the meta-observation paragraph and the cluster-F list. Runner-up spoke ../agentic-tooling-wiki (the software-writing-software culmination) noted, not split. Linked technology-adoption-curve, agent-skills, crossing-the-chasm. Index updated (Article group). +1 page (research 75→76). Fetched live. (No build/commit here — hub handles.)

[2026-06-29] ingest | graphrag-rs (Rust GraphRAG) — hub-routed

Routed here (the knowledge-graph / tools-for-thought thread; sibling to sift-kg). Two new pages + four updates.

• graphrag-rs (SoftwareApplication, source, T3, github.com/automataIA) — Rust GraphRAG: builds a KG from documents and runs the retrieval half sift-kg lacks — Leiden community detection + PageRank + community summaries → graph-based answers. Local-first: trait core, CLI/TUI, REST+Qdrant server, WASM/WebGPU browser build; Tokio, ONNX embeddings, Ollama. Self-reported maturity (Phase-1 production, 214+ tests); not independently benchmarked (T3 note).
• graphrag (DefinedTerm, mechanism) — the GraphRAG method (Microsoft 2024): community detection + community summaries to answer global / whole-corpus queries plain vector RAG can't. Distinct from the bare knowledge-graph (substrate) and plain retrieval-augmented-generation (local chunks). Sourcing gap noted: grounded via the implementation, not yet the canonical MS paper.
• Updated retrieval-augmented-generation (adds the fifth gap — whole-corpus synthesis to the RAG-failure taxonomy), knowledge-graph ("using the graph to answer global queries" section), sift-kg (builder vs builder-querier framing), and synthesis (the four-gaps paragraph → five; KG tools split into builders vs builder-queriers). Dedup: not a dup of sift-kg — it adds the retrieval layer; refreshed both in place.
• Entity automataIA (org) noted inline, not paged (thin; value is the technique). Gap-relevance: advances the KG/RAG thread (the global-query gap was previously unstated). Provenance on every claim; T3 weakness recorded.

[2026-06-29] ingest | LeanRAG (KG-RAG, AAAI 2026) — hub-routed

Routed here (the knowledge-graph / RAG thread; sibling to graphrag-rs/graphrag). One new page + four updates.

• leanrag (ScholarlyArticle + SoftwareApplication, source, T2, github.com/KnowledgeXLab) — AAAI 2026 KG-RAG framework that refines GraphRAG. Names two failure modes: semantic islands (community/ summary nodes left unconnected → no cross-community reasoning path) and structure-unaware retrieval (flat similarity ignores graph topology → redundant evidence). Fixes: semantic aggregation (cluster entities into summaries and build explicit relations among them → navigable summary network) + bottom-up hierarchical retrieval (anchor at fine entities, traverse upward). Reports ~46% less retrieval redundancy; avg win 78.1% vs GraphRAG, 71.9% vs HiRAG (author-reported benchmarks — margins are motivated; T2 note).
• Updated graphrag (new "semantic islands" critique section; sourcing gap downgraded from implementation-only to peer-reviewed-anchored — canonical MS paper still the named next add), retrieval-augmented-generation (fifth-gap section: LeanRAG refines the mechanism, not a sixth gap), knowledge-graph (global-queries section: topology, not just nodes, becomes part of retrieval), and synthesis (fifth-gap paragraph + first peer-reviewed KG-RAG anchor).
• Dedup: not a dup of graphrag-rs — same builder-querier shape, but the contribution is how the summaries connect and how retrieval walks them. Entity KnowledgeXLab (org) + AAAI 2026 (venue) noted inline, not paged (thin; technique is the value). Gap-relevance: advances the KG/RAG thread and partly closes graphrag.md's sourcing gap (first peer-reviewed source). Provenance on every claim; T2 with author-reported-benchmark caveat.

[2026-06-29] ingest | From Local to Global (canonical GraphRAG paper, Microsoft 2024) — user-requested

Named next-add from graphrag's sourcing gap. Two new pages + five updates.

• from-local-to-global-graphrag (ScholarlyArticle, source, T1, arxiv.org/abs/2404.16130) — Edge, Trinh, Cheng, Larson et al. (microsoft Research, Apr 2024); the canonical GraphRAG paper. Method: LLM entity-KG extraction → Leiden hierarchical community detection (levels C0 root → C3 leaf) → per-community summaries → map-reduce answering (map = partial answer + helpfulness score 0–100; reduce = pack top-scored, synthesize). Eval: two ~1M-token corpora (podcast, news); baselines SS (naive vector RAG) + TS (source-text summarization); LLM-judged comprehensiveness/diversity/empowerment + directness control. Results: all GraphRAG levels beat naive RAG (comprehensiveness 72–83%); C0 root needs 9×–43× fewer tokens/query than TS, >97% fewer than C3 — the cost/quality knob. T1 caveat: results LLM-judged on authors' own setup (method solid, margins motivated).
• leiden-algorithm (DefinedTerm, mechanism) — new connector node: hierarchical community detection (improves on Louvain); GraphRAG's clustering step + source of the C0→C3 levels. Linked from graphrag / graphrag-rs / leanrag / from-local-to-global / knowledge-graph.
• Updated graphrag (sourcing gap closed → grounded in primary; added Leiden hierarchy + map-reduce
  • C0 cost knob), retrieval-augmented-generation (fifth-gap now cites the primary), knowledge-graph (global-queries paragraph cites paper + Leiden), leanrag (its "semantic islands" critique now points at the original community summaries), synthesis (fifth-gap grounded in primary + the cost knob).
• Entity reuse: cross-wiki microsoft (Organization, ../agentic-tooling-wiki) — not duplicated; authors noted inline (thin). Dedup: this is the method's primary source — graphrag.md (the DefinedTerm) stays the concept page, this is its source summary. Provenance on every claim; T1 with author-eval caveat.

[2026-06-29] ingest | GraphRAG-Bench ("When to use Graphs in RAG", ICLR 2026) — user-requested

The neutral third-party benchmark the KG/RAG open-Q wanted (gap named in the 06-29 quality pass). One new page

• four updates; resolves an open question and opens a flagged tension.

• graphrag-bench (ScholarlyArticle, source, T1, arxiv.org/abs/2506.05690) — Xiang, Wu, Q.Zhang, S.Chen, Hong, X.Huang, Su; ICLR 2026 (per repo). Neutral: proposes no method of its own. Benchmarks MS graphrag (local+global), LightRAG, HippoRAG/HippoRAG2, RAPTOR, Fast-/Lazy-GraphRAG vs vanilla RAG (±rerank) on Novel (Gutenberg) + Medical (NCCN) corpora across 4 levels (fact retrieval → multi-hop reasoning → summarization → creative gen). Verdict task-conditional: vanilla RAG wins simple fact retrieval (~61–65% vs GraphRAG 49–60% — graph adds "logically relevant but redundant" context); graphs win multi-hop + summary (HippoRAG2 ~54% vs RAG ~43%; 87.9–90.9% recall on complex). Cost 1–2 OOM higher: vanilla ~900 tokens, HippoRAG2 ≈10³, LightRAG ≈10⁴, MS-GraphRAG global ~4×10⁴.
• Updated graphrag ("when it actually helps" neutral-verdict section), retrieval-augmented-generation (neutral-check: graphs task-conditional, not a blanket win), synthesis (open-Q "compare to mature RAG/ graph-RAG" → Resolved; new Contradictions/tensions entry: advocates gbrain/method-papers "graphs win" vs neutral benchmark "it depends + costly").
• Dedup: distinct from from-local-to-global-graphrag (a method primary) — this is a neutral cross-method benchmark. Gap-relevance: closes the named neutral-benchmark gap; partially addresses the cost/effort axis. Authors noted inline (institutions not listed; thin). Provenance on every number; T1 (peer-reviewed venue + neutral primary). Records both sides of the graphs-vs-vector tension rather than overwriting the advocate claims.

[2026-06-29] ingest | HippoRAG 2 (ICML 2025) — user-requested page

Requested as its own page (it was the standout graph method in graphrag-bench). One new page + four updates.

• hipporag2 (ScholarlyArticle, source, T1, arxiv.org/abs/2502.14802) — "From RAG to Memory" (Gutiérrez, Shu, Qi, Zhou, Yu Su; OSU-NLP, ICML 2025). Covers the lineage from HippoRAG v1 (NeurIPS 2024, arXiv:2405.14831): hippocampal-indexing analogy (neocortex→LLM, hippocampus→knowledge-graph, parahippocampal→retriever); LLM OpenIE builds the graph, Personalized PageRank does single-step multi-hop retrieval (v1: up to 20% multi-hop gains, 10–30× cheaper / 6–13× faster than IRCoT). v2 fixes the KG-hurts-factual-recall tradeoff via deeper passage integration + better online LLM use → beats vector RAG on factual + sense-making + associative memory (+7% associative vs SOTA embeddings); framed as non-parametric continual learning.
• Cross-grounded with the neutral graphrag-bench: HippoRAG2 = top complex-reasoning graph method (~54% vs ~43% vanilla) AND most token-efficient (≈10³ tok/query, ~40× under MS GraphRAG global) — the one method where advocate paper + neutral benchmark agree; shows the cost gap is a design problem, not intrinsic.
• Updated graphrag-bench (link HippoRAG2 in methods/results/cost), retrieval-augmented-generation (HippoRAG2 named as the efficient exception to the "graphs are costly" check), knowledge-graph (Related +hipporag2/+graphrag-bench), index.
• Dedup: distinct mechanism (PPR over entity+passage graph for multi-hop) vs GraphRAG community-summaries / LeanRAG summary-linking. Tier T1 (ICML 2025) — author-run eval, but uniquely corroborated by the neutral benchmark. Authors/OSU-NLP noted inline (thin). Provenance: both HippoRAG papers cited inline by arXiv id + the held graphrag-bench source page.

[2026-06-29] ingest | RAPTOR (Stanford, ICLR 2024) — user-requested page

Requested as its own page (sibling to hipporag2/lightrag; method in graphrag-bench). One new page + five updates.

• raptor (ScholarlyArticle, source, T1, arxiv.org/abs/2401.18059) — "Recursive Abstractive Processing for Tree-Organized Retrieval" (Sarthi, Abdullah, Tuli, Khanna, Goldie, C.D. Manning; Stanford, ICLR 2024). Key nuance: RAPTOR is a TREE, not an entity knowledge-graph — recursive embed→cluster→ LLM-summarize, repeated bottom-up into a multi-level summary tree; retrieval pulls nodes across levels ("collapsed tree"). For long-doc/multi-step QA; +20% absolute on QuALITY w/ GPT-4 (NarrativeQA/QASPER strong).
• Framing: RAPTOR (Jan 2024) predates MS GraphRAG (Apr 2024) and reaches the same multi-level-summary insight via a summary tree of chunks vs an entity graph + Leiden communities — the "tree cousin." Honest note recorded: graphrag-bench really tests structured/ hierarchical retrieval broadly (KG is one structure, a tree another).
• Updated graphrag-bench (link RAPTOR + flag it's a tree), graphrag (non-graph tree-cousin note), hipporag2/lightrag (Related — sibling cross-links), index.
• Dedup: distinct from the entity-graph methods (it's hierarchical chunk summarization). Did NOT add to knowledge-graph's Related (not a KG method — would mis-categorize). T1 (ICLR 2024) with author-run-eval caveat; cost/positioning via graphrag-bench. Authors/Stanford noted inline (thin).

[2026-06-30] ingest | How to build a powerful LLM knowledge base (TDS)

Routed from Telegram. New source llm-knowledge-base (TechArticle, T4, towardsdatascience.com; Kjosbakken, 2026-06-27). A fifth independent voice on the llm-wiki/gbrain "LLM + persistent markdown brain" pattern (it cites both — andrej-karpathy's wiki and garry-tan's GBrain). Two contributions, both folded into synthesis point 4 of the llm-wiki thread: (1) names the read-path dichotomy — grep-based markdown-index inference (the llm-wiki end, = how this hub retrieves) vs embedding-RAG (the gbrain end); (2) flips the thread's emphasis to capture-as-the-bottleneck — automated cron ingestion from meetings/Linear/coding-agents + passive agent use, "who fills the trails" vs "who keeps them." Dedup: no prior knowledge-base page. Author entity (Kjosbakken) deferred (one-off T4 author, tangential); GBrain/Karpathy/Tan already paged (linked). Ran avoid-ai-writing. +1 page (→89).

[2026-07-01] ingest | Redeploying Claude Fable 5 (Anthropic, primary) — resolves the June suspension

T1 primary (anthropic.com). Dedup: claude-fable-5 + claude-mythos-5 already paged (the earlier claude-fable-5-infoq T4 recorded the suspension) → refreshed both in place + new source redeploying-fable-5. The arc, precisely: US export controls (12 Jun) named both Mythos-class models over cyber capability; Anthropic suspended both (no real-time foreign-national verification); controls lifted (30 Jun); redeployed 1 Jul (Claude Platform/.ai/Code/Cowork; ≤50% weekly limits through 7 Jul, then credits; cloud to follow). Nuances captured: (1) Amazon-found jailbreak's vuln-finding was not unique to Fable 5 (Opus 4.8/GPT-5.5/Kimi K2.7 found the same) — severity check; (2) cyber classifier hardened to >99% block w/ "safety margin," fallback to claude-opus-4-8; (3) Fable = no unique offensive cyber, Mythos = uniquely capable — safeguards not weights. Touched claude-fable-5 (resolution added to suspension section), claude-mythos-5 (export-control section), synthesis (substrate now a policy-contingent/availability-axis dependency). Governance layer (export controls + multi-vendor jailbreak-severity framework w/ Amazon/MS/Google + HackerOne + govt partnership) = cross-spoke to ../ai-governance-wiki. Volatile snapshot.

[2026-07-01] ingest | Wiki Memory (LangChain / Harrison Chase) — names the LLM-wiki pattern as agent memory

T3 vendor blog (conceptual — no APIs/benchmarks). Dedup: llm-wiki already the canonical concept → enriched in place + new source wiki-memory. Chase names the pattern "wiki memory" and reframes it as a type of agent memory: (1) scope boundary — durable domain knowledge, NOT conversation state / user prefs / high-frequency logs (bridges agentic-tooling's agent-memory); (2) sharpened RAG contrast (precompute+maintain synthesis vs retrieve raw chunks at query time). Establishment-signal like OKF from Google Cloud. New instances extend the pattern into a codebase-documentation sub-genre: deepwiki (Cognition) + autowiki (Factory). New pages: wiki-memory, deepwiki, autowiki, harrison-chase (Person; cross-link langchain agentic-tooling). Touched llm-wiki (naming + agent-memory framing + instances) and synthesis (added to the "natural attractor / independent arrivals" thread, point 3). Reflexive: this hub is a wiki-memory instance. Cross-spoke: ../agentic-tooling-wiki (agent-memory, LangChain).

[2026-07-04] ingest | Long context vs. chunking — the "how you index" lever on the RAG-critique thread

Msg 728 (towardsdatascience.com/long-context-vs-short-context-model…, via Telegram). T2 (independent practitioner empirical analysis — ModernBERT ~32M/~150M on real datasets HUPD/BigPatent, 3 experiments, seeds, significance notes, throughput numbers; single-author). New page long-context-vs-chunking (TechArticle, url). Thesis: document length ≠ need for a long context window; where the signal lives + the O(n²) cost of context decide. Findings: HUPD 8192-vs-512 = +1.15 pts n.s. (front-loaded signal); BigPatent chunk-and-pool 0.654 @ 4.6× less compute beats full-8192 (0.632); retrieval whole-doc embeddings 0.006–0.030 nDCG@10 vs overlapping chunks 0.082; throughput 447→20 docs/s (512→8192). Folded into retrieval-augmented-generation as a new upstream chunking/segmentation lever (the how-you-index axis, distinct from the what-you-retrieve gaps): a chunk-boundary gap (embeddings average a boundary-crossing fact apart; overlap recovers it) = the segmentation-side cousin of the exact-token gap (hybrid-retrieval-rag), and the cluster's "cheap machinery beats heavy" cost lesson (graphrag-bench) carried onto the context-window axis. Synthesis updated likewise. Cross-spoke: ../llm-inference-wiki owns the attention mechanism the article leans on (O(n²), RoPE, local/global attention, sequence-packing → flash-attention) — cross-linked, not routed there. +1 page. Ran avoid-ai-writing. Per-route verify deferred (routine ingest, no structural page move).

[2026-07-10] ingest | "LLM Wikis Are Over-Engineered — I Replaced Mine With a Pure Python Compiler" (TDS) (hub-routed, Telegram)

Ingested the TDS piece arguing the llm-wiki agent loop points a probabilistic tool at a deterministic job: parse/cross-reference/lint are mechanical, so an agent buys token cost, latency, and non-determinism (same folder → different links). Replacement = a stdlib-only compiler (regex extract → word-indexed phrase-matcher graph [107s→<1s] → section-aware rewrite preserving hand-written ## Notes → orphan/broken- link lint), byte-identical output; LLM reserved for the semantic ~10% (paraphrase links). New wikis-over-engineered-compiler (BlogPosting, source, T3) + new concept wiki-compiler (the deterministic-pipeline approach; deterministic↔probabilistic axis). Updated llm-wiki (compiler-critique section). Synthesis: new A-core axis (deterministic compiler vs reasoning agent, orthogonal to augment→automate; ranks wiki-compiler/OKF ↔ sift-kg ↔ gbrain by structural trust); flagged as the most reflexive source (critiques wikis like this hub — which is already the LLM-ingest + deterministic-verify hybrid it advocates). Distinct from the existing TDS llm-knowledge-base (different author/argument). Runner-up spoke ../agentic-tooling-wiki (its autowiki/machine-maintained-doc thread is the agent-side foil). T3 opinion, self-reported benchmarks. avoid-ai-writing applied (clean). +2 pages, 1 updated.

[2026-07-27] ingest | rahulnyk/knowledge_graph — convert any text into a graph of concepts (GitHub)

Routed here by the hub (runner-up: ../knowledge-representation-wiki — declined by its own boundary note: that spoke owns the RDF/OWL/SPARQL tradition, this is a pandas+NetworkX property graph built by an LLM, the markdown/LLM knowledge-graph tradition this spoke owns). T3 demo-notebook README, MIT, ~3.5k★, no evaluation of any kind. New: rahulnyk-knowledge-graph (source), entity-resolution (mechanism). Updated: knowledge-graph (two new sections — co-occurrence as its own edge type, and the dedup problem), sift-kg (cross-link to the collected entity-resolution answers), synthesis (new open question), index. Two ideas worth taking from a notebook this small:

1. Concepts, not entities. It declines NER on purpose — "'Bangalore' is an entity, 'Pleasant weather in Bangalore' is a concept" — where sift-kg, gbrain and GraphRAG all extract entities. Propositional nodes buy expressiveness and cost mergeability, which is the direct cause of its own top open TODO.
2. Contextual proximity as a second weighted edge class. Co-occurrence in a chunk (W2) accumulates on the same pair as the LLM's semantic relation (W1), so edge weight means "how often and by how many routes" rather than "how confident is the model." Cheap (no extra call), and it makes provenance structural: an edge is the chunk the two concepts shared. The new entity-resolution page collects what was scattered across three pages: four answers spanning the augment→automate axis, none of them measured. That gap became the new synthesis open question — graphrag-bench closed the evidence hole for graph retrieval; graph construction still has one. Small correction recorded on the source page: the README's setup step says ollama run zephyr while the text specifies Mistral 7B OpenOrca everywhere else. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | A brief history of Luddism (The Economist, Free Exchange) — PARTIAL, paywalled

Routed here by the hub into cluster F (diffusion & adoption). T2 — The Economist's Free Exchange column, 23 July 2026. Honest partial ingest. New: brief-history-of-luddism (source, marked partial). Updated: technology-adoption-curve (new section: the missing variable), index.

Fetch: WebFetch blocked outright; Firecrawl (per HUB edge handling) got past the block but not the paywall — it returned the headline, standfirst, byline furniture and the opening paragraph only. Per HUB, a paywall Firecrawl can't bypass is the case for an honest stub, so the page records what was actually read and marks the rest unread. Not a transient error, no retry loop, no invented argument.

What the recovered text supports: the Luddites were textile workers in northern England who by night smashed stocking frames and mechanical looms that "stole jobs from humans," led by a man who did not exist — Ned Ludd, variously king, general or captain, supposedly an apprentice driven to it by an overbearing master, and placed in Sherwood Forest alongside Robin Hood. Plus the standfirst, which carries the thesis: "States ultimately decide how fast technology is adopted."

Why it's worth the page despite the paywall. Cluster F's founding model (technology-adoption-curve) explains adoption speed with five attributes of the innovation — relative advantage, compatibility, complexity, trialability, observability — plus communication channels and the social system. None of them is a state. No ban, patent regime, tariff, factory act or troops sent to guard the mills. If the column's standfirst holds, Rogers describes diffusion inside politically set limits rather than explaining the pace itself. Recorded as a potential tension, not an established one — the argument behind the standfirst is unread, and the page says so. Second thread the opening paragraph opens: Rogers' acknowledged pro-innovation bias has a historical edge here. The Luddites weren't laggards waiting for the S-curve; they were organised workers acting against specific machines on a stated economic interest. The curve's tail category may sometimes be reading its interests correctly rather than failing to keep up. Entities: none paged. The Economist and the Free Exchange column would be reasonable nodes, but with one paragraph of evidence held, evidence-only says wait. Re-ingest if a full text becomes available — the page refreshes in place. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-29] ingest | LLMs can't jump (Zahavy, DeepMind) + its press version

Two sources on one subject, ingested as a pair. The secondary arrived first — a Wccftech report — and the primary PDF followed two minutes later, which is the only reason the corpus has the argument right. New pages: llms-cant-jump (T1, PDF read directly), llms-cant-jump-press (T4), abduction, tom-zahavy.

Cluster E gains a named ceiling. Zahavy maps Einstein's discovery cycle (sense experience → jump → axioms → deduction → theorems) onto three inference types: induction mastered, deduction being conquered via alphaproof, and abduction — inventing the axioms — structurally missing. The case study is General Relativity, chosen because the data was scarce, which is exactly where Schmidhuber's discovery-as-compression thesis has nothing to compress. The proposed remedy is action-controllable world models (Genie-class, intervention not prediction) supplying embodied simulation.

Better than the usual limits for a specific reason, recorded in synthesis: Gödel's limit is logical and has no remedy; this one is claimed to be architectural, so it arrives with a proposal. Cluster E has been documenting the deduction arrow being mechanized — this names the arrow before it and argues that one is mechanizable too, given grounding.

Held as a position paper, not a finding — the author's own word, 10 pages, historical case study. Its sharpest checkable claim is about a benchmark: ARC-AGI captures the logical leap and misses the manipulative component, since nothing in it is physically manipulated. Scope is explicitly the physical sciences, which matters here because cluster E's subject is mathematics.

tom-zahavy is why the paper carries weight: he is a co-author of Hubert et al. (2025), the AlphaProof work he cites as the deduction success, and of the test-time-RL Penrose-chess work he cites for problem-variation. The limitation is argued from inside the programme by someone who built the part that works.

The press version is a cluster-F specimen and that's why it was kept. llms-cant-jump-press turned a hedged architectural proposal into "LLMs will never replace human genius." Three things dropped: the concession that an LLM could execute the deductive phase, the physical-sciences scope limit, and the words "position paper." Each one made the original narrower and stronger. Cluster F theorizes about how ideas diffuse; this is a checkable specimen with primary and secondary both held and the delta inspectable. Cross-linked to ../psychology-wiki's dopamine-nation, where the same compression happened inside one author rather than between two. avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] ingest | ten-proofs (OpenAI) — Lean certificates for ten claimed advances

Routed by the hub (Telegram, github.com/openai/ten-proofs). T1 — repo plus OpenAI's own publication, both read directly (the announcement page hard-blocked WebFetch with a 403; fetched via the firecrawl fallback per ../HUB.md).

New pages: ten-proofs (source), lean-certificate (DefinedTerm). Updated: lean-theorem-prover, alphaproof, synthesis, index.

Deliberately not paged: Astra. The model is ../llm-providers-wiki's subject (the model market); paging it here would create the duplicate the hub's entity rule exists to prevent. Noted on ten-proofs so the canonical node lands there when a model-market source arrives.

Contribution 1 — cluster E's evidence base changes character. alphaproof proved competition statements whose answers were known to exist; this is ten open problems chosen by their own communities, with the argument generated by the model and the Lean written by it too. AlphaProof's page now records that it has been superseded as the cluster's high-water mark while remaining the better-documented system.

Contribution 2 — the epistemic one, and it reaches past this spoke. Every capability claim in the hub is self-reported. A lean-certificate isn't a percentage: it type-checks on a stranger's laptop with no access to the model. First source in the corpus the "measured on the sample that flatters them" objection doesn't touch. Recorded with its limits, which are severe — the kernel certifies validity, not significance, not provenance, not that the formalized statement is the one a mathematician would write; and the class of claims settleable this way is basically maths and program verification.

Tension flagged, not resolved. llms-cant-jump (Jan 2026) argued abduction is structurally missing. Ten open problems fell six months later, which looks like a refutation and isn't: Zahavy scoped his claim to the physical sciences and conceded the deductive phase to machines. All ten results sit inside established axiomatic systems. What grew is the deduction arrow's reach; abduction is untested here. Written into synthesis as a tension.

Cluster F bonus: the announcement footnotes five follow-on arXiv papers by human mathematicians — diffusion of a result the community didn't produce, observable from day one.

avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] ingest | MemGraphRAG (XMUDeepLIT) — memory-based multi-agent graph construction

Routed from the hub (Telegram; the repo). T1 — peer-reviewed (KDD 2026), MIT code released, all systems run on one embedding model (NV-Embed-v2) at k=5 with GPT-4o-mini and temperature 0. New pages: memgraphrag (SoftwareSourceCode + ScholarlyArticle, source), graph-construction-quality (DefinedTerm/mechanism). Updated: graphrag (the step before clustering/summarizing), graphrag-bench (neutrality scoped), synthesis, index. Dedup: graphrag, graphrag-bench, knowledge-graph, retrieval-augmented-generation, hipporag2, lightrag, raptor all already paged — linked, not duplicated. The README alone carries no numbers, so the paper (arXiv 2606.00610, 20pp) was pulled and read with pypdf. Gap-relevance: opens a question the cluster had never asked. Every prior source competes on what to do with the graph; this one says the graph is produced badly and that is where the losses are. Synthesis: new section "The graph itself became the subject." The load-bearing number is awkward for the field — discarding 40% of low-frequency triples slightly improves accuracy (65.28 vs 64.85), so most of what the extractor emits is not load-bearing. Three named failure modes of chunk-local extraction (thematic irrelevance, logical inconsistency, structural fragmentation) generalize to any pipeline that builds persistent structure a window at a time. Read against the paper's own framing. Table 3 transplants its graph under four rival retrievers and gains only +0.19 to +0.71 points, against a full-system margin near +2.9. The paper presents this as evidence its constructor is a universal upgrade — true, and consistently positive. Read the other way it says most of the win comes from the matched retriever, not the cleaner graph. Recorded as an attribution observation, not a contradiction; the retriever is a stated contribution. Provenance finding, verified before writing. graphrag-bench is recorded here as "a neutral third-party benchmark… the authors propose no method of their own." Its author list (fetched from arXiv to check, not taken from the existing page) is Zhishang Xiang, Chuanjie Wu, Qinggang Zhang, Shengyuan Chen, Zijin Hong, Xiao Huang, Jinsong Su. MemGraphRAG's is Chuanjie Wu, Zhishang Xiang, Yunbo Tang, Zerui Chen, Qinggang Zhang, Jinsong Su — four shared names, including both equal-contribution first authors and the corresponding author — and it tops that benchmark on G-Medical and G-Novel. Nothing retracted: the benchmark came first and genuinely proposed no method then. What changed is scope — its independence covers other people's systems, not this one, and the wiki can no longer treat "neutral benchmark" and "method winning on it" as two data points. Also recorded on the source page: the margin is +2.10 on LLM-judged accuracy with GPT-4o-mini judging GPT-4o-mini's own generations; on the llama-3-70b backbone HippoRAG2 beats it on MuSiQue containment (33.90 vs 33.70) and G-Novel (56.16 vs 55.76); and Tables 1 and 3 disagree on one cell (G-Novel 57.41 vs 54.41) while every other Table 3 entry reconciles exactly. Entities: authors noted inline, none paged — the existing graphrag-bench precedent deferred them as thin, and the evidence here is still only names on two papers. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-03] ingest | Introducing Claude Opus 4.8 — Anthropic announcement (via research pass)

Hunted, not routed. The hub Research Pass took ## Most wanted edge 3 ("model-launch product facts — needs first-party announcements to displace T4 trade press; T3 allowed, the edge names a first-party fact"). The hunt reframed the edge. Dedup found the Fable 5 first-party announcement already held (claude-fable-5-mythos-5-announcement), so the Fable corner never lacked a primary source — what it lacked was linkage. The genuine hole was Opus 4.8: three secondary pages (claude-opus-4-8-review, claude-opus-4-8-zvi, claude-opus-4-8-launch-tomsguide) and no vendor record. That is what was fetched. T3 — a vendor announcement, and deliberately so: for a price or a shipped feature the vendor is the authority and the trade press restating it is not. This does not raise the tier floor; it puts the right tier under the claims. New: claude-opus-4-8-announcement. Updated: claude-opus-4-8 (product facts now sourced first-party), claude-opus-4-8-launch-tomsguide (retained as a media fact, superseded for product facts), synthesis, index. Contradiction recorded, not resolved: the wiki reported "~4× less likely to make unsupported claims" (honesty). Anthropic's wording is "around four times less likely… to allow flaws in code it has written to pass unremarked" (code review). Both kept per record-don't-overwrite; flagged in synthesis. Also new to the corpus from this source: Terminal-Bench 2.1 92.3%, OSWorld-Verified 84.9%, Online-Mind2Web 84%, effort control, dynamic workflows. Entities: none created — anthropic already canonical, linked not duplicated. avoid-ai-writing run over the new prose.

[2026-08-03] lint | unlink the dead oracle-cloud-signup reference (markup only)

One [[oracle-cloud-signup]] wikilink in this log pointed at a page deleted from ../cloud-wiki on 2026-05-31 (curator-authorized; replaced there by oracle-cloud-free). Converted to backticked plain text — wording untouched, rendered output identical, since an unresolved wikilink already rendered as plain text. Surfaced by the new unresolved-wikilink warning in npm run verify.

[2026-08-03] lint | note the cross-spoke corroboration of Fable 5's published price

llm-providers-wiki/claude-fable-5 spent four sources triangulating $10/$50 and recorded it as unconfirmed, while the primary Anthropic announcement stating that rate has been held here since 2026-06-10 (claude-fable-5-mythos-5-announcement). Corrected there; noted here because the corroboration is genuinely new information for this page — four independent routes to the same figure is unusually strong for a pricing claim in this corpus. No claim on this page changed; one sentence added, and the two Fable 5 pages now point at each other under the new dual-lens convention.

[2026-08-04] ingest | RAG-Anything (HKUDS, arXiv 2510.12323)

Routed here by the hub (runner-up: ../agentic-tooling-wiki). T1 — arXiv primary research, tiered consistently with lightrag, hipporag2 and graphrag-bench.

Dedup first, and it paid: lightrag was already held, and RAG-Anything is built on it, by the same group and the same senior author. So this is a refresh-and-extend of an existing lineage rather than a new entry — lightrag gains a "what was built on it" section and loses the "authors/HKUDS noted inline (thin)" caveat it had carried since 2026-06-29.

New pages (3): rag-anything (source), hkuds (Organization), chao-huang (Person — the only researcher with two systems in the cluster). Updated: lightrag, knowledge-graph, index, synthesis. Entity discovery stopped at two: the other four authors had no evidence beyond a name.

Gap it lands in. Every RAG system in this corpus indexes text only — never argued for, just the shape of the input everyone assumed. This is the first entry to break it (images, tables, LaTeX promoted to graph entities via dual-graph construction). Folded into "The graph itself became the subject" rather than opened as a new dated section, because it continues that argument exactly: the cluster has now walked backwards through its own pipeline twice — retrieval strategy → graph construction (memgraphrag) → document parsing (here) — and each step found the losses upstream of where anyone was measuring. Nothing in the corpus measures extraction fidelity, and unlike the first two steps this one arrived as an architectural fact rather than a paper arguing for it.

Evidence recorded honestly, including what could not be read. The abstract claims "superior performance" and "significant improvements" and names no dataset, no baseline, no number. The paper does name DocBench and MMLongBench against nine baselines including lightrag itself — but the numeric tables could not be extracted from the PDF on ingest (compressed content stream), so the source page records the benchmarks as verified and the figures as unread, not absent. What is settled either way: the evaluation is author-run, by the group that wrote one of the baselines. New growth edge #3 — a neutral benchmark covering multimodal retrieval, the thing graphrag-bench is for the text case.

[2026-08-04] lint | consolidated three dated sections into two thematic ones

The other half of the hub's synthesis-drift item; agentic-tooling's six were folded the same day. ../QUALITY.md check 3 names dated (added 20xx) sections accumulating outside the living thesis as drift, and this spoke had three.

Two of them were one argument. The cluster-E ceiling has a name now set out llms-cant-jump's claim that abduction is the missing arrow and the deficit is architectural; The certificate arrives before the ceiling is settled was ten-proofs testing exactly that ceiling. A reader met the limit and its test as two separate arrivals, six days apart, with the second opening "six months after" the first — scaffolding that only makes sense if you are reading the page as a diary.

Merged into Cluster E's ceiling, and what a certificate can settle, with three subsections: the press deformation as a cluster-F specimen, the certificate testing the ceiling (and why the obvious refutation reading is wrong), and what a lean-certificate settles versus what it cannot. The third section, The graph itself became the subject, is a genuinely separate thread and kept — its date simply removed.

Nothing dropped. Zahavy's scope limit to the physical sciences, the ARC-AGI testable claim, the position-paper caveat, the $2,000 figure, the narrowness of what a kernel certifies, the Leiden attribution stance — all present. What went is the diary framing and the duplicated setup where the second section re-explained the first.

Same rule adopted here as in agentic-tooling: a new finding extends the thread it belongs to; it does not get a dated section of its own.

[2026-08-04] lint | second drift pass — the stacked substrate updates in Current thesis

The earlier pass today fixed section-level drift (three dated ## …(added 20xx) headings → two thematic). It did not fix the layer underneath, and the curator was right to send it back.

Inside Current thesis, four dated notes had stacked on one another: Tension (added 2026-06-03), Substrate update (2026-06-10), Substrate update (2026-07-01), Availability as a policy variable (2026-06-15) — the last one out of date order, which is the clearest sign nobody was reading the section as prose. A reader met the substrate story as four arrivals in the order the sources happened to land, and had to assemble the argument themselves.

Rewritten as one statement organized by axis rather than by date: the substrate is a moving dependency, and it moves on honesty, cost, where safety lives, availability, and the vendor's own economics. Every fact survives — the "performative honesty" finding and the Andon Labs quote, the $10/$50 doubling, Fable and Mythos being the identical model with three classifiers and a <5% fallback, the three-day pull and the 12 June → 30 June → 1 July export-control sequence, the check that the jailbreak capability was not unique to Fable 5, the T4 provenance of the InfoQ account, the Microsoft retention conflict, and the $965B/$9B valuation with its unproven economics.

Two things gained by the reordering. The availability axis now reads as a property of the bet rather than as news: a government instrument removed this wiki's own substrate for three weeks. And the vendor-economics paragraph lands as the risk sitting under all the others instead of as a trailing note.

Left alone deliberately: the dated markers in Contradictions / tensions (one per entry) and the (added 2026-06-09, user directive) note on cluster F. Those record when a tension entered the corpus and why a scope decision was made — provenance, not accretion. Stripping them would delete information to make the page look tidier.

[2026-08-05] ingest | Entity-centric evaluation of entity resolution (research pass)

Via the hub research pass against growth edge #1, the spoke's oldest evidence hole. er-evaluation-framework — Binette et al., arXiv 2404.05622, T1. Paged olivier-binette as lead author; the other five authors are plain mentions (evidence-only — the paper is their whole footprint here).

The edge asked for "a T1 paper or system report with measured merge precision/recall" and got that (91% pairwise precision / 94% recall on PatentsView), but the useful part is the correction it makes to how the gap was framed. The spoke had been saying nobody reports a number. The paper's finding is that the naive number is biased upward — pairwise precision on a benchmark set runs "close to 1" when true precision is much lower, because benchmark construction hunts for matches among O(n²) non-matches and so enriches for easy positives. Combined into F1 that bias produces rank reversals. So a reported precision figure from any of our four systems would probably have been wrong in a known direction.

Refreshed entity-resolution in place with a new section rather than rewriting the standing "nobody measures it" paragraph — both readings are true and the older one is what makes the new one land.

Transfer limit recorded: the validation is classical record linkage over inventor names. The entity-centric sampling idea is node-agnostic, but "an annotator can cheaply confirm this cluster is fully resolved" is a much weaker assumption for the proposition-shaped nodes rahulnyk-knowledge-graph produces than for a person. Edge #1 rewritten to ask for the missing half: the framework applied to an LLM-built graph.

[2026-08-05] lint | freshness regrade

agent-memory-knowledge-graphs volatile → stable (61 days past the 60-day window). It is a dated Substack tutorial with no current-state quantity in it, so re-reading the URL returns the same text and the flag could never be cleared by any cycle — the case ../QUALITY.md already covers.

[2026-08-05] ingest | OpenDeepWiki (AIDotNet)

Routed from the hub (runner-up: ../agentic-tooling-wiki). New pages: opendeepwiki (source, T3 — first-party README) and aidotnet (Organization). Refreshed deepwiki and autowiki in place; both were two-line stubs naming a sub-genre with nothing open in it.

The codebase-documentation branch of wiki-memory now has an instance you can read the internals of: MIT, self-hosted, .NET/Next.js, generation on a background schedule, each repo served over MCP. Two things folded into synthesis. The MCP endpoint is a third read path next to llm-knowledge-base's grep-vs-embed pair — publish the wiki as a tool. And the regenerate-on-a-schedule design splits the maintenance question in two: you can maintain trails or redraw them, and only a corpus with one machine-readable ground truth (the code) gets the second option. That is the sharpest statement so far of why this hub's raw/ is immutable.

Recorded gap: nobody in the sub-genre reports whether the generated documentation is any good.

[2026-08-05] ingest | Formal verification (Wikipedia survey)

Routed from the hub (runner-up: ../programming-languages-wiki, which took the sibling SPARK source the same minute). New page formal-verification (source, T2). Cluster E's first source about the field rather than a tool.

It splits what the cluster had been treating as one activity: model checking (exhaust states — tla-plus), deductive verification (discharge obligations — isabelle on seL4, lean-theorem-prover on mathematics), abstract interpretation. Folded into synthesis with two findings: the survey's specification problem is the same hole lean-certificate reached from a single artifact, which upgrades that observation to the discipline's own known limit; and the hardware-verifies / software-doesn't asymmetry is economic, which explains why every verified-software example in the field is a kernel, a compiler or avionics.

Cross-spoke link recorded to spark — the same deductive technique sold as a shippable language.

[2026-08-07] ingest | Yank Note (purocean/yn)

Routed from the hub (Telegram). T1 (official project repo), AGPL-3.0, 6.7k★/925 forks, Electron/Vue/TypeScript, cross-platform.

New pages: yank-note (source summary), executable-markdown (mechanism), purocean (thin maintainer node). It joins the tools-for-thought roster beside obsidian, logseq and roam-research, and keeps the same storage bet — local plain markdown, features expressed in original Markdown syntax as far as possible, per-file .c.md encryption.

What is new for the corpus. The note runs: code blocks in JavaScript, PHP, Node, Python and bash execute from inside the document, with an AI Copilot (OpenAI/Ollama/Gemini/Kimi) writing into the same file. Eighty years of this lineage improved how notes connect and left the document inert; this adds effects — files, network, shell — as a separate axis, paged as executable-markdown.

The reason it earns a synthesis section. The project states its own trade: it "sacrifices security protection (command execution, arbitrary file reading and writing)" for extensibility, and warns against opening untrusted markdown. In the llm-wiki pattern the notes are written by a machine from fetched sources, and at gbrain's unattended end nobody reads them before they open — so prose and code arrive through one pipe at one trust level. Filed as a new open question: this lineage has no threat model for LLM-authored notes. Recorded reflexively too — this hub is only clear of it because plain markdown plus a static site generator happens to keep the reading tool and the executing tool apart, which nobody decided on purpose.

Entities: 1 created (purocean).

[2026-08-08] ingest | Rocketnotes (fynnfluegge/rocketnotes)

Routed from the hub (Telegram), the second markdown-PKM app in two days after yank-note. T1 (official project repo), Apache-2.0, 1.4k★, Go/TypeScript/Python — and entirely self-described, with no measurement of anything.

New pages: rocketnotes (source summary), agentic-archiving (mechanism), fynn-fluegge (thin maintainer node). Updated zettelkasten with the section that source earns.

Why it matters here. This wiki's thesis rests on one claim — Bush and Engelbart had the design and lacked the labour, and the LLM supplies it. That claim has been carried by an essay (llm-wiki-gist) and one team's system (gbrain). Rocketnotes sells it as a bullet: "An AI agent analyzes snippets from your 'inbox' and intelligently files them into the most relevant existing document." Paged as agentic-archiving and folded into synthesis.

The critique the corpus is equipped to make. The automated unit is placement, and zettelkasten's central principle is connectivity over categorization — value is in the links, not the folder. So the agent automates the part of the practice the method says does not matter, while proposing a link with a reason goes undone and is no harder to build. This hub's own practice is the counterexample and is cited as such. Recorded on zettelkasten too, because the vocabulary will keep being borrowed as AI note apps multiply and that page is where the borrowing gets checked.

A second finding, sideways: it ships RAG (semantic search, chat-with-your-documents) and the maintenance agent, with no sense that llm-wiki-gist framed them as rival answers. That is probably the future shape — both, in one product — and it sharpens the open question rather than dissolving it.

Storage noted as a departure from what this lineage usually asks for: markdown content, but persisted in DynamoDB and S3 rather than a folder of files you own (obsidian, yank-note). The all-local Docker mode answers privacy; whether it leaves you portable .md files is not stated in the README and the code was not read.

Entities: 1 created (fynn-fluegge).

[2026-08-08] query | How is the search-signal method an instrument for cluster F?

Curator asked how the demand-discovery method routed to ../search-marketing-wiki relates to this spoke's diffusion models, then asked for the answer to be filed. New page: adoption-curve-measurement.

The answer, short: the four models (technology-adoption-curve, crossing-the-chasm, gartner-hype-cycle, bass-diffusion-model) all describe shape and none locates a technology on it in real time — the corpus already says this about itself (Bass "fit retrospectively"; the hype cycle "not empirically validated"; tech-adoption-curve-twenty-years on early flagging). The search signals are a live, cheap reading of the same phenomenon.

Three of the five map onto existing concepts (absent phrasing = the first 2–3%; contested labels = pre-critical-mass; formal vocabulary = institutional adoption). Two do not exist in any of the four models — keyword difficulty and results-page structure count sellers, where Rogers, Moore and Bass all count adopters. That is a second axis the theory lacks. The method also splits Bass's p and q into separately observable channels rather than fitting both from one series afterwards.

It reads on the cluster's own admitted hole: technology-adoption-curve records that Rogers has no term for power, and searches for standards and statutes are that variable arriving as data.

Four failures recorded on the page: interest is not adoption, no denominator (so no fraction-of-market), hype and uptake are indistinguishable in a rising line, and no false-positive rate. Standing: one unvalidated instrument, strongest where the models are weakest.

Cross-wiki: the instrument's own page stays in ../search-marketing-wiki (emerging-category-search-signals) — routed there on dominant substance 2026-08-08; this page is the theory side and links across rather than duplicating.

[2026-08-08] ingest | CORE-KG + LINK-KG — a before/after duplicate count on an LLM-built graph

Two arXiv papers (T1) by the same group — core-kg (June 2025) and link-kg (October 2025), dipak-meher, Domeniconi, Correa-Cabrera. Arrived via the research pass, hunting growth edge #1 ("a number on entity resolution — for these systems").

The edge asked for an entity-centric evaluation on an LLM-built graph, or any system report giving a before/after duplicate count. These give the second. Unmodified GraphRAG on judicial case documents: 27.02% duplicate nodes on short documents, 36.01% on long ones. With coreference resolved across the document before extraction: 10.61% and 17.78%. CORE-KG sits between at 17.00% / 26.10%, and its own paper's claimed −33.28% duplicates reproduces when LINK-KG re-runs it as a baseline.

Pages: link-kg, core-kg, dipak-meher new; entity-resolution and graph-construction-quality updated with the numbers; synthesis gained "And now the construction defect has a size" and the open question on graph nodes is marked largely answered.

Kept, not overwritten: entity-resolution's "The unresolved part" stays as written with a note that it records the pre-measurement state.

The limit is why the edge does not fully close. Duplicates were counted by RapidFuzz partial_ratio ≥75% over intra-type pairs plus expert review of the clusters — a method that needs nodes whose surface form carries their identity. Successor edge written: duplication where it does not, i.e. the proposition-shaped nodes rahulnyk-knowledge-graph, gbrain and sift-kg produce.

Entity budget: 1 of 8 (Dipak Meher). Domeniconi and Correa-Cabrera left as plain mentions — evidenced as co-authors, nothing else about them in either source. No affiliation node: neither paper's affiliation line was retrievable, and the George Mason University Library appears only as where the case documents were accessed, which is not an affiliation claim.

[2026-08-08] ingest | LLM×MapReduce — the third answer to an input that does not fit

Telegram source, github.com/thunlp/LLMxMapReduce, T1 (official project repository). llm-x-mapreduce new; long-context-vs-chunking extended; synthesis gained a section.

THUNLP / OpenBMB / AI9STARS, Apache-2.0, ~872★. Divide-and-conquer for long-to-long generation: split the corpus, process fragments, fold the results. V1 carries a structured information protocol and confidence calibration so fragments merge rather than stack; V2 adds entropy-driven convolutional test-time scaling and powers SurveyGO, which writes survey articles from a literature corpus.

Why it is a genuinely new shape here rather than another RAG variant: retrieval-augmented-generation retrieves a slice and graphrag builds a structure to traverse, and both assume a short output. This one is built for a long one. The authors' framing — long-input understanding is well studied, long-output generation is not — describes this cluster's own shelf accurately.

Numbers held at arm's length per the graphrag-bench lesson: 95.50% precision / 95.80% recall on SurveyEval against vanilla at 25.48% / 26.46% is author-run on an author-built dataset, and a 4× gap probably says more about the baseline than the ceiling.

Loop closed from the same day. V2 defines claim density — unique claims over total extracted, after intra- and cross-group dedup. This morning's research pass hunted exactly that metric shape for claim-shaped graph nodes and failed. This is still not the answer (it scores a generated article, not a persistent graph, and reports a ratio rather than merge correctness), but it is the third source in one day where claim-level deduplication appears as machinery and is never scored as a result. The recurrence is recorded on the edge as the finding.

No entity discovery: THUNLP and OpenBMB are named, neither has a node here, and one repo ingest is not the moment to open two org nodes.

[2026-08-09] ingest | Org mode, and Babel's JSS paper (via research pass)

Coverage edge 4, half closed. org-mode from the Org manual (orgmode.org, T1), and org-mode-literate-programming — Schulte, Davison, Dye & Dominik, Journal of Statistical Software 46(3), 2012, T1, CC-BY with code published alongside.

What the corpus was missing. obsidian, logseq, roam-research, tana and tiddlywiki5 all had pages while the plain-text system in continuous use the longest had none. Org is "an authoring tool and a TODO list manager for GNU Emacs" — an outliner that grew TODO states, cross-file agenda views, tables with a spreadsheet, executable source blocks and export, and never left plain text.

The Babel paper is the better find. A multi-language literate-computing environment, peer-reviewed in a statistics journal in 2012, with the artifact attached: one file holding prose, data, project data and code blocks in different languages, where one block's output can feed another. Next to executable-markdown and wiki-compiler it dates the "document that runs" idea well before the current tooling that presents it as new.

Recorded as half closed, deliberately. The edge asked for the manual plus a long-running user's account, and both sources here are the system describing itself — one of them written by Org's own maintainer. The corpus now holds Org's design and nothing about what twenty years of daily use looks like, which is the only evidence that would actually test the durability claim. The successor edge asks for a multi-year retrospective or survey data.

Entities: 0 created — author pages deferred; the JSS paper is one source and no author recurs yet.

[2026-08-14] ingest | The Babel paper, read properly this time

Re-ingest of org-mode-literate-programming from the full PDF. The curator re-sent the JSS galley URL (.../view/v046i03/557); the 2026-08-09 ingest had used the abstract page, so most of the paper had never been read. Refreshed in place per the re-seen-source rule — no duplicate page.

What the body added that the abstract did not.

• The paper is built on a distinction the corpus was collapsing. literate-programming and reproducible-research are two traditions with opposite directions of travel — code tangled out to a compiler versus code run in to generate the document — and both now have pages. The 1984 Knuth quote and the 1995 Buckheit & Donoho "advertising of the scholarship" line are on them.
• A ten-tool comparison table (Javadoc, POD, Haskell .lhs, noweb, cweb, Sweave, SASweave, Statweave, Scribble, Org) scored on four axes. The interesting failures: Sweave executes but cannot reorder blocks, so it is only partially literate; noweb and cweb reorder but never execute.
• Org adopted noweb's block-reference syntax unchanged — "identical syntax and functionality."
• The safety section, which corrects a claim on executable-markdown. That page said none of the three ordinary mitigations was evidenced in this corpus; the per-execution confirmation prompt is, and has been Org's default since 2012. Both claims kept, correction dated on the page.
• An adoption-cost argument for why literate programming never spread: the tools are "prescriptive and difficult to use." Cluster F material, and unmeasured.

Entity discovery (5 nodes, budget 8). Created eric-schulte, dan-davison, thomas-dye, carsten-dominik and journal-of-statistical-software (Periodical — a new @type group in index.md). All five came back create new from entity-index, except one false match worth recording: "Thomas Dye" scored 0.62 against thomas-judson, a mathematics-wiki textbook author. Different people; the note is on the Dye page so the next reader does not re-litigate it. Affiliations (UNM, Counsyl, Oxford, Hawai'i, Amsterdam) stayed prose — one mention each, tangential per the relevance gate. Knuth is cited, not an agent of this source, so no node under the evidence-only rule.

The Dominik problem, now explicit. He wrote Org, maintains orgmode.org, and co-authored this paper — so the wiki's two Org sources are one voice, not two. Recorded on his entity page and in synthesis.md; coverage edge 4 stays open for the same reason it opened.

[2026-08-14] ingest | Org mode's adoption, measured

The curator read org-mode and told me to fill the gap its closing note admitted: the manual documents the system, not its uptake. Growth edge 4 had asked for the same thing since 2026-08-09 ("a multi-year retrospective, or survey data on Org use"). This is the survey half.

Sources found and used. The Emacs User Survey — the only large count of Emacs users that exists. The 2020 edition (Brochard, 19 Oct–30 Nov 2020) published 7,344 responses and its raw response file, so I downloaded Emacs-User-Survey-2020-clean.csv and computed the Org numbers here rather than quoting a write-up. That mattered: the published analysis skipped the Org questions, treating them as free text it "did not attempt to graph," so the tabulations on emacs-user-survey-2020 are not available anywhere else in this form.

What the primary data says. 79.7% of respondents use Org, 51.2% daily or exclusively, 20.3% never, and 239 people (3.3%) run Emacs only for Org — the package recruits users to its host program. Cross-tabbed against Emacs tenure, penetration peaks at 85.2% in the 4–5 year band and falls to 68.4% past twenty years; a 20+ year user is more than twice as likely never to have used Org as a 4–5 year user. The arithmetic is suggestive — 20+ years in 2020 means starting before Org existed in 2003 — but a single cross-section cannot separate cohort from exposure time, and the page says so.

Two measurements taken here, not read. Stack Exchange API, 2026-08-14: org-mode is 5,918 of 28,349 questions on emacs.stackexchange.com (20.9%, the top tag by 4× over key-bindings) and 2,141 of 24,251,096 on Stack Overflow (0.0088%, ~1 in 11,300). MELPA download counts for six Org add-ons, and no entry for Org itself — it ships in Emacs core since Emacs 24, so install telemetry cannot see it. Both are labelled as self-derived on the pages that use them.

Pages. New: emacs-user-survey-2020 (Dataset, source: true, T2 — a new @type group in index.md) and org-mode-adoption (Intangible, the analysis). Updated: org-mode (the closing "not read here" note replaced by an Uptake section and a narrowed remaining gap), adoption-curve-measurement, synthesis.md, index.md.

Where it landed in cluster F. adoption-curve-measurement found on 2026-08-08 that search signals give a numerator with no denominator. Three unrelated instruments did the same thing here: every Org figure is a ratio inside a population nobody has counted, so Org cannot be placed on an S-curve at all. Two instruments this different failing identically promoted the no-denominator problem from a limit of one method to a property of the measurement problem. Also new to the cluster: bundling blinds telemetry exactly when it makes adoption universal, which is a second actor deciding uptake that Rogers has no variable for — the first being the state.

Contradiction recorded, not resolved. Both T1 sources rest Org's importance on Babel; 10.2% of the 4,980 respondents who described their use mention it, against 61.6% for notes. Neither claim retracted — a tool's importance to a lineage and to its users are different quantities. The 10.2% is keyword matching over an optional free-text question 2,364 people skipped, so it is a floor, and settling it needs a closed question no survey has asked.

Not used. GitHub's language:org repo count (32) is a primary-language classification and cannot see a format that lives inside other people's repositories — a broken proxy, left out rather than reported with a caveat. The 2007 Worg poll (~80 responses, T3) is on the page only as a warning: 80 in 2007 against 7,344 in 2020 is not a growth curve.

[2026-08-15] lint | open-knowledge-format: first sighting outside the spec

Page touch only, no ingest here. The hub routed https://towardsdatascience.com/how-to-utilize-okf-efficiently-to-enable-knowledge-exchange-among-llms/ to ../llm-inference-wiki (it measures a tokenizer optimization; OKF is its envelope), but the source is the first use of OKF this corpus has seen outside Google's own repo, and open-knowledge-format's caveat said adoption was unproven. Added a "First sighting outside the spec" section recording it, with the honest weight: one independent developer, borrowing the shape, and explicitly not spec-valid — he adds a token_pointer field the spec does not define. That is the producer-defined extension a registry-free format invites, which cuts both ways for the anti-formalist reading in synthesis. Cross-linked to llm-inference-wiki/okf-token-handoff. No new pages, no synthesis rewrite — one sighting does not move the adoption question.

[2026-08-18] ingest | Semantica — graph-native infrastructure for accountable AI

Routed by the hub from Telegram. https://github.com/semantica-agi/semantica — MIT Python library, 8,719★ / 894 forks / 2,357 commits at the read, first commit 2025-06-25. T1 (first-party repository documentation), freshness: volatile.

New pages: semantica (source), semantica-agi (Organization), mohd-kaif (Person). Updated: entity-resolution (new section — "A fifth answer: ship it, and report the speed"), temporal-knowledge-graph (bi-temporal for audit rather than for retrieval), synthesis (growth edge 1 widened), index.

Why it earns a page. It is the largest deployed graph in this corpus — 118,000 nodes — running blocking plus semantic deduplication, and it publishes only speeds: dedup 6.98× faster, candidate generation 63.6% faster, node search 24 ms → 0.004 ms. No merge precision, no duplicate count, no evaluation. Growth edge 1 was framed as a hole in the research literature; a production system with the data and the runs to know its merge quality, reporting throughput instead, means the edge is broader than that. Restated: nobody publishes merge correctness, in the lab or in production.

Marketing read separately from the software. "The Open Source Palantir for AI Agents" and a comparison table scoring vector DBs as None on provenance, reasoning, conflict detection and time travel are advocacy, and the page says so. Against that, the README draws one limit unusually honestly — "System-level explainability, not foundation-model explainability" — and its performance footnote admits the figures are CHANGELOG history rather than test assertions.

Entity budget: 2 of 8. entity-index "Semantica" returned nothing ≥ threshold (top hit 0.44 mnemonic, osint-wiki) — genuinely new, unlike the two known false "create new" cases. mohd-kaif paged because the commit split (1,813 of 2,357) is itself a finding next to the enterprise pitch.

Cross-spoke: the RDF/OWL/SHACL/SPARQL/SKOS layer is knowledge-representation-wiki's domain and the compliance framing touches ai-governance-wiki; neither was written to. Recorded as context on semantica instead, per the hub's one-source-one-spoke rule.

[2026-08-22] research | A duplicate rate on units that are not names (via research pass)

Growth edge 1 hunted for the third time and half-closed. orkg-predicate-redundancy (John, Aggarwal, Auer & Karras, TIB Hannover, TPDL 2026, arXiv:2607.03197, T1) read in full: 13,457 predicates in the live Open Research Knowledge Graph clustered into 3,837 groups, 30% potentially removable — 2,044 (15.19%) redundant merge candidates across 786 clusters, 1,989 (14.78%) unused — with 7,361 predicates (54.7%) in "Review", the bucket where the method declines to recommend and a curator must decide.

Three causes, all behavioural: one user creating the same predicate repeatedly (593 same-user clusters, 79 with identical labels, no duplicate warning in the interface), IDs typed into label fields producing a 312-member false-positive cluster, and no standardisation for method / data / material.

Why it is a half close. Predicates are not claims. They are schema — a small, constantly reused, curator-visible vocabulary — where propositions are numerous, single-use and unread. Successor edge 1a asks for the same measurement at proposition scale, or an argument that triage is the honest answer there. index.md gains one entry; the edge-1 text now records the first non-name duplicate rate the corpus holds. Entities: 0 created (TIB not paged — no other source here depends on it).

[2026-08-23] ingest | Information science — Wikipedia (routed by hub)

Arrived via Telegram; hub routed here over knowledge-representation-wiki (runner-up), because the dominant substance is the discipline whose history and subfields sit directly under cluster A, not the formal-KR/semantic-web machinery that spoke owns. T2.

New pages (7): information-science-wikipedia (source) · information-science · documentation-movement · paul-otlet · asist · information-retrieval · bibliometrics. Updated: vannevar-bush (he was not first — Otlet 1895), tools-for-thought (the tradition beside it), retrieval-augmented-generation (ancestry note on IR evaluation), index.md, synthesis.md.

The finding is relocation rather than addition: cluster A's lineage starts at 1945 and the field it belongs to starts at 1895, with a definition and a society by 1968 and named subfields for most of the corpus's open questions. Recorded honestly as a map, not evidence — the source carries no numbers. Entities: 2 (Otlet, ASIS&T); La Fontaine, Briet and Borko left as mentions (no evidenced content beyond a name).

[2026-08-23] ingest | Information scientist — Wikipedia (routed by hub)

Companion stub arriving four minutes after the above, ingested together with it. T2, and thin — the article carries two institutional dates, the decline of the job title, and a bare list of 17 names with no contribution attributed to any of them. The term's attribution to Wm. Hovey Smith is uncorroborated and flagged as such on the page.

New pages (3): information-scientist-wikipedia (source) · information-scientist (Occupation — a new @type for this wiki, hence a new index.md section) · cilip.

What it earns its place with: the intermediary arc. The Institute of Information Scientists (London 1958) defined a searcher-on-your-behalf as distinct from librarianship; end-user searching dissolved the role, and the institute merged into the Library Association in 2002. Bush's proposed "profession of trail blazers" is the same idea from the same decade, and llm-wiki / agentic-archiving propose it again with software in the seat. Folded into synthesis.md as one section covering both sources, plus coverage edge 8 (information science as a literature, not a category — needs a relevance-assessment source and an information-seeking-behaviour source). Entities: 1 (cilip).

[2026-08-30] ingest | WiCER — what an LLM wiki loses when it compiles (via the quality cycle's absence-claim drain)

Found sideways. The cycle was draining the absence-claim queue and reached wikis-over-engineered-compiler, whose tier note says "No independent evaluation." One search for that turned up WiCER (Huerta, arXiv 2605.07068, 8 May 2026, T1) — not an evaluation of that author's compiler, but a measurement of the thing his essay argues against, and the number is worse than anything this spoke had assumed.

Blind LLM compilation of 80 documents per topic answers 55.1% of probes drawn from those same documents wrongly or irrelevantly; mean quality 2.17 of 5. A compile-evaluate-refine loop — probe with one QA pair per source document, find the failures, pin the facts those questions needed, recompile — recovers about 80% of the lost quality in one or two passes (to 3.18–3.24, 24.8% catastrophic) and still lands under the raw documents (3.47) and under RAG (3.64).

Pages: new wicer, juan-huerta. Updated llm-wiki (a "What the compile step costs" section — the pattern's "compile once" is the expensive assumption, not the cheap one), wikis-over-engineered-compiler (the absence claim corrected: the opponent's position now has a number, though WiCER's answer is a loop rather than the deterministic compiler that essay wants), synthesis.md (a new contradiction — the wiki-beats-RAG claim holds on 30 curated articles, 4.38 vs 4.08, and reverses on 80 raw documents, 3.47 vs 3.64 — and a new open question about this hub's own blind spot), index.md.

The reflexive part is the reason it is logged at length. Every wiki in this hub is a blind compilation, and no checker here can see a fact that never made it out of a source. WiCER's probe is the instrument that would, and the raw documents it needs are already held.

[2026-08-31] ingest | OpenKB — the LLM-wiki pattern as an installable CLI

Routed from the hub (Telegram). https://github.com/VectifyAI/OpenKB — Apache-2.0, Python, 4,246★ / 442 forks, repo created 2026-04-04. T1 for what the artifact is; it publishes no benchmark, so every evaluative word in it is an unmeasured maker's claim. Runner-up spoke: agentic-tooling-wiki (the Skill Factory), noted as cross-spoke context rather than split off.

Pages: new openkb, pageindex, vectifyai. Updated llm-wiki (OpenKB added to Implementations, and the "independent convergence" claim cut back — see below), open-knowledge-format (first tool claiming conformance; the caveat softened, not retired), retrieval-augmented-generation (a new section: PageIndex deletes the vector index rather than adding beside it), wiki-memory (a fourth instance shape — installable, general-purpose — and Chase's conversation-state boundary independently redrawn where he drew it), andrej-karpathy (a Downstream section: OpenKB copies him closely and drops his one-source-at-a-time commitment), synthesis.md, index.md.

The correction is the point of this ingest. llm-wiki had been reading each new instance as evidence of "multiple independent realizations converging… a natural attractor, not one author's idiosyncrasy." OpenKB matches this hub's design down to the ingest page-count (10–15) — and says in its first paragraph it is built from Karpathy's concept, as are this hub and llm-wiki-agent. That is descent, not convergence, and the attractor argument now rests on the two members that owe the gist nothing (OKF, tana). What survives is smaller and better evidenced: the design survives being packaged for strangers.

Two things genuinely new to the cluster: pageindex's vectorless tree retrieval — the first entry in the RAG thread that removes the vector index instead of supplementing it, at a claimed self-run 98.7% on FinanceBench — and the Skill Factory, which turns the knowledge-markdown ↔ procedural-markdown bridge this synthesis had only described into a command. Unchanged: nobody, OpenKB included, measures whether a compiled wiki is right (wicer, 55.1% wrong-or-irrelevant).

[2026-09-06] ingest | Formalizing Fermat's Last Theorem (Anthropic, 2026-09-04)

Routed here by the hub from Telegram (fbclid stripped). Anthropic's account of Claude producing the first end-to-end machine-checked proof of FLT in Lean: 11 days, 13M lines, 30,300 theorems (29,500 used), ~6B output tokens, Lean's three standard axioms, a comparator check against Mathlib's own statement of FLT. Proof published at github.com/anthropics/fermats-last-theorem (Apache-2.0, Lean, created 2026-09-04, 725★ at ingest — read from the GitHub API).

Tier T3. A first-party announcement about the company's own model's performance; every number is Anthropic's. Two things soften it without changing the grade: the artifact is kernel-checkable by a stranger (the lean-certificate argument), and kevin-buzzard — who leads the community FLT project this adapts pieces from — reviewed the proof and is quoted. He also kept the hedge the announcement drops: "which Anthropic researchers say only took 11 days."

The distinction that needed a page. ten-proofs was the model generating mathematics and certifying it; this is the model translating Wiles 1995 into Lean. Anthropic states the difference plainly ("what's novel here is the verification"), and the two sources otherwise look alike, so autoformalization now holds it. It also settles what this does not touch: abduction is untested here, since formalizing an existing proof stays inside someone else's axioms.

The negative result is the transferable one. Autonomous agents failed first — lost project state, stopped collaborating, ~7% of the non-boilerplate lines survive from that attempt. prove2me (Chen, Marwaha, Lu, Yuen & Peng, arXiv 2608.28433) supplied a theorem-statement DAG, statements split from proofs, and natural-language descriptions for reuse. External state store, not agent autonomy.

Pages: new flt-formalization-anthropic, autoformalization, prove2me, kevin-buzzard, tianyi-peng. Updated lean-theorem-prover, lean-certificate (the cost objection now has a counter-data-point), formal-verification (translation as a fourth activity in front of deductive verification), synthesis.md, index.md.

Not read: the Prove2Me paper itself — this page is built from a user's account of the platform, not the authors'. Noted on that page.

[2026-09-06] ingest | lean-lang.org — re-seen subject, and a four-year adoption chronology

Arrived via Telegram as https://lean-lang.org/, routed here by the hub. Runner-up spoke: ../programming-languages-wiki — declined, but the reason is now worth writing down (below).

Re-seen, not new. lean-theorem-prover has been paged since 2026-05-29, and the spoke already held ten Lean-adjacent pages (lean-certificate, ten-proofs, autoformalization, flt-formalization-anthropic, alphaproof, prove2me, isabelle, rocq, tla-plus, lean-for-programmers). Refreshed that page in place and added a source summary for the site.

New pages: lean-lang-org (the site + the FRO About page), mathlib, and two entities — lean-fro and leonardo-de-moura, both with verified same-as (Wikipedia for de Moura, confirmed via the REST summary API as the right person: creator of Z3 and Lean).

Tier split on one source. T1 for what Lean is — features, kernel design, code samples, governance, funders — all checkable against the artifact. T3 for the timeline's impact entries and the four executive testimonials: a self-curated selection with no stated inclusion criteria, published by the organization it funds, quoting people at AWS and Harmonic, both of which fund the FRO. The page says which reading applies where.

The find is in cluster F, not cluster E. /fro/about carries a month-by-month chronology from 2022-08 to 2026-09. Read against adoption-curve-measurement, every figure in it is a numerator — 5,000★ (2025-02) → 7,500★ (2026-03) → 9,007★ (GitHub API, today), 100,000 VS Code installs, 20,000 Mathlib contributions, 10,000 Zulip members, 50+ university courses. The vendor has every incentive and every internal number and still cannot supply a denominator. That makes Lean the third independent case after search signals and org-mode-adoption, and it upgrades the cluster's claim from "our instruments are weak" to "this is what measuring adoption is without a census."

It also reproduces the looks-longitudinal-and-is-not shape: "over 30,000 installs in a single calendar year" (2024-12) printed nine months above "installed over 100,000 times" (2025-09) — a rate and a total, reading as 3.3× growth.

Two cluster-E updates. First, the 2026 band of the timeline is overwhelmingly validation infrastructure built because machines started producing proofs — Comparator, Lean Eval, Palomar, the Lean Kernel Arena. lean-certificate argued the certificate convention changed how results are announced; this is the supply side of it. Second, verification is leaving mathematics: five of the six projects Lean's own front page features are software and hardware (AWS Cedar, Microsoft SymCrypt via Aeneas, Ethereum's ArkLib, Veil, SampCert), and AWS describes running the Lean formalization as an executable oracle against production Rust in CI. Not proving a theorem — differential-testing shipped code against a specification that happens to be a Lean program. The spoke had no instance of that use.

The boundary, recorded not acted on. The site is headed "Lean Programming Language" and the project's own one-liner puts language before proof assistant, which on the tagline alone points at ../programming-languages-wiki. It stays here: this spoke owns the mechanizing-reasoning lineage, the eleven Lean pages are here, and the dominant substance of every Lean source the corpus has seen is proof. The marker is in synthesis.md — if sources arrive about Lean's dependent type theory or its compiler as a language, the seam moves.

Also new on the artifact itself: the grind tactic (Lean 4.22.0, 2025-08) now carries most of the front page's sample proofs, which shifts this page's old assistant-checks-vs-prover-finds line; and the licence, absent from the site, is Apache-2.0 per the GitHub API.

[2026-09-06] ingest | ImperialCollegeLondon/FLT — the human project, still running

Third Lean source of the day via Telegram, arriving straight after lean-lang-org. Routed here; runner-up ../mathematics-wiki (FLT as mathematics) — declined, the subject is a formalization project. T1, the project's own repository, README and GENERAL.md, plus the GitHub API.

New page flt-imperial-project. Nothing in the corpus had paged it; it existed only as a phrase inside flt-formalization-anthropic and kevin-buzzard.

Three corrections to pages written two days ago, all from the primary source, which is exactly the case ../HUB.md describes when it says to fetch the primary and treat the earlier record as a lead:

1. The two FLT formalizations follow different proofs. Imperial is explicitly not doing Wiles/Taylor–Wiles but a newer route via Khare–Wintenberger and Kisin, designed by Richard Taylor. Anthropic did Wiles's 1995 proof via Darmon–Diamond–Taylor. flt-formalization-anthropic left this ambiguous and now says it.
2. Date. kevin-buzzard said the project started "in 2024." Repository created 2023-11-19, GENERAL.md dated December 2023, Buzzard's funded time starting October 2024. Three events, three dates; the page now distinguishes them.
3. The 2029 goal is a reduction, not a complete proof from the axioms — reducing FLT to claims known by the late 1980s, with foundations PR'd to mathlib along the way.

The finding. GENERAL.md, December 2023 and unedited today: "the AI tools available to humanity are not really able to help with the formalisation aspect of this task… It will be interesting to see how different things will be in September 2029 as the grant ends." It resolved in August 2026. A domain expert, forecasting his own field, with a date attached, wrong by a factor of three on timing — and he then read the result and updated in public. Folded into synthesis.md as a cluster-F entry, because the cluster's models say nothing about how badly the best-placed observer can misjudge a rate.

Why the project continues, from its own documentation: many objects of modern number theory cannot even be stated in any theorem prover, so formalizing a hard theorem forces the definitions into existence. That reframes mathlib as a library of definitions before proofs, and it is why the repository was pushed to on the day of this ingest.

Open question left standing: how much of Anthropic's 13 million lines is reusable by anyone. Its own caveat says the proof is "likely much longer than it needs to be"; Mathlib admits only "the strongest approach to any given topic." No basis to compare yet.

[2026-09-06] ingest | lean-lang.org/use-cases/mathlib — re-seen, refreshed in place

Fourth source of the day, arriving minutes after mathlib was created from lean-lang-org. Same subject, so refreshed in place. Written for "decision-makers considering Lean" — T1 for process and coverage, T3 for the adoption pitch.

A first-party contradiction, kept as one. Lean's front page says Mathlib is "over a million lines"; this page says "over two million," twice, on the same site, read the same day. Neither is dated. Recorded in synthesis.md → Contradictions, with the consequence spelled out: the corpus quotes "5× Mathlib" for Anthropic's 13 million lines, and that multiple is 13× or 6.5× depending on which vendor sentence you take. Mathlib line-counts here are now order-of-magnitude only.

What it adds beyond that: the review rule behind the single hierarchy, stated first-party — "only the strongest approach to any given topic is included", an editorial monopoly per topic rather than a registry; tooled QA (linters over definitions, type-class instances and simp lemmas; CI; docs infrastructure) with a paper the corpus does not hold (van Doorn, Ebner, Lewis, arXiv:2004.03673 — next source for anyone wanting the mechanism); and the Mathlib Initiative as a program of Renaissance Philanthropy, which partly answers the governance question lean-fro raised — the FRO is time-boxed, the Initiative is framed around continuity, and they are separate bodies.

Named users worth the cross-link: PFR (Tao) and the Liquid Tensor Experiment (Commelin, Topaz, Scholze) on the research side; AWS SampCert, zkEVM tracks and Starkware's Cairo on the industrial side. That last pair is a ../cryptography-wiki seam — the scheme is theirs, the library is ours; cross-link, don't duplicate. And alphaproof reached IMO silver by training on Mathlib, which is the training-data relationship the corpus had not recorded.