Find conflicting facts INSIDE an rlat knowledge model (.rlat) — two documents that disagree about the same thing — using cheap corpus geometry to surface the candidates and YOUR own reading to judge them, with no external API key and no metered cost. Trigger when the user asks "are there contradictions in my corpus?", "does my knowledge base disagree with itself?", "find conflicting facts", "is my documentation inconsistent?", "which docs contradict each other?", "audit my .rlat for conflicts", or wants the corpus to check its own consistency. This is the contradiction half of the corpus self-audit (the gap half is `rlat-gap-scan`). NOT for: answering a question (use the `rlat` search skill); finding what's MISSING (use `rlat-gap-scan`); auto-fixing a conflict (this v1 SURFACES for review, it does not edit the corpus).
Human-in-the-loop curation of an rlat knowledge model (.rlat): surface the self-audit findings (gaps, contradictions, stale external facts), then let the USER decide each one — approve a drafted fill, hand-provide an authoritative source, or skip — so nothing the user hasn't sanctioned ever lands. Trigger when the user says "curate my knowledge base", "let me review before adding", "I'll approve what gets added", "human in the loop", "I'll provide the source", "review the findings with me", or wants final say over what enters the corpus's earned layer. This is the approval/authority front-end over `rlat-gap-scan`, `rlat-contradictions`, and `rlat-refresh-facts`. NOT for: autonomous filling (the loop does that unattended); answering questions (use the `rlat` skill).
Find the gaps in an rlat knowledge model (.rlat) — questions the corpus cannot actually answer — using YOUR own reading of the retrieved context, with no external API key and no metered cost. Trigger when the user asks "what is my corpus missing?", "where are the gaps?", "what can't this .rlat answer?", "scan my knowledge base for holes", "what should I add to the docs?", or wants the .rlat to self-improve / notice what it lacks. Also the detection half of the closed learning loop: a confirmed, recurring gap is what earns a cloud-authored fill. NOT for: answering a question (use the `rlat` search skill); multi-hop research (use `deep-research`); gaps judged by retrieval score alone (proven blind — see below).
Teach an rlat knowledge model (.rlat) the world it serves — from inside the session, on the user's subscription, no API key. Extract durable WORLD facts, standing constraints, and tried-and-falsified findings from what the USER said, apply the four capture gates (incl. the privacy gate: facts about the person never land), confirm with the user, then land each via the key-free `rlat capture-attribute`. Trigger when the user says "remember this", "rlat should know this", "teach the knowledge model", "capture that as a constraint", "record that X didn't work", "save what we learned about the project/tenant/garden", or at a session's natural end when durable world facts surfaced. NOT for: personal preferences or facts about the user themself (the privacy gate drops them); corpus content that already lives in the documents (build/refresh handles those); session-transient state.
Check whether the EXTERNAL (web-fetched) facts inside an rlat knowledge model (.rlat) are still true in the live world, by re-fetching each fact's source and re-reading it yourself — no API key, no metered cost. Trigger when the user asks "are my fetched facts still current?", "did anything I added from the web go stale?", "re-check my external facts", "refresh the facts in my .rlat", "is the external info in my knowledge base out of date?", or wants the corpus's outside facts re-verified against the live world. This is the staleness `rlat refresh` CANNOT fix — refresh only re-syncs corpus files to disk; an external fact's truth lives on the web. NOT for: corpus drift (use `rlat refresh` / `rlat reverify`); finding contradictions (use `rlat-contradictions`); finding gaps (`rlat-gap-scan`).
Build large rlat knowledge models (.rlat files) on Kaggle's free T4 GPU when the local machine has no GPU or encoding would take too long on CPU. Walks the user through Kaggle account + CLI setup, builds a kernel script that installs rlat[build,ann], sparse-clones (or uploads) source, encodes the corpus on T4, writes a remote-mode .rlat back, and pulls the artefact home. Trigger when the user mentions Kaggle, "free GPU", "build remotely", "no GPU here", a corpus too large to encode locally (~10K+ passages on CPU), or wants to rebuild several .rlat files in one batch. Not for: small corpora that finish in minutes on CPU; users who already have a CUDA GPU (just `pip install rlat[build]` and run `rlat build` locally).
Orchestrate rlat CLI workflows on the user's behalf for any task involving an rlat knowledge model (.rlat file). Triggers cover: setting up rlat for a fresh project (init), building or refreshing a knowledge model on drift, running grounded retrieval to ground an answer in citations, generating skill-context for an Anthropic skill !command block, recalling past decisions from memory, comparing two corpora, switching storage modes (bundled / local / remote), and the opt-in MRL optimise pipeline. Trigger when the user mentions rlat by name, references a .rlat file, asks to build/query/refresh/sync/watch a knowledge model, asks "what does the project know about X" on a project with a .rlat file, or describes any task that maps onto an rlat workflow. This skill executes the workflow on the user's behalf — it does not just document the CLI. For multi-hop research with cross-file synthesis (plan → retrieve → refine → synthesize loop), defer to the deep-research skill instead. Not for: Pinecone, Chroma, FAISS, or ot
Interactive multi-hop research over an rlat knowledge model, run by Claude in this conversation against `rlat search`. Plan → retrieve → refine → retrieve → synthesize loop, driven by the assistant turn-by-turn so the user can redirect at any hop. Bench-validated: 92.2% answerable accuracy at 0% hallucination on the Microsoft Fabric documentation (63 questions, Sonnet 4.6, relaxed rubric) — same loop the `rlat deep-search` CLI implements, exposed here as a free Claude Code workflow that uses the user's existing subscription instead of the Anthropic API. Trigger for cross-codebase synthesis questions whose answer requires evidence from multiple files: rationale ("why was X chosen over Y"), trade-offs, contradictions across sources, historical "why" questions, long-horizon investigation, "summarise the X across the codebase", "what's the full story of Y". Trigger for memory-recall research across past sessions (drive the same loop on a memory knowledge model). Trigger when a single `rlat search` returned thin o