From compounding wiki to traceable signal.

01 / Executive verdict

Karpathy’s formulation. Our evidence-aware extension.

Andrej Karpathy gave the LLM Wiki pattern a timely and influential formulation: instead of retrieving raw chunks and rebuilding understanding for every question, let an LLM incrementally compile immutable sources into a persistent, interlinked Wiki.

Persistent personal knowledge, associative trails, source preservation, provenance, and human review all predate this formulation. Karpathy himself relates the pattern to Vannevar Bush’s Memex. Wiki Engine claims neither the broader paradigm nor “evidence discipline” as an invention. Its contribution is a particular implementation question:

The target answer is an evidence-aware pipeline with distinct authority layers: exact Source, source-bound Evidence, one-source Signal, cross-source Wiki, human Review, and operational Receipt.

02 / The original insight

Knowledge should be compiled once—and kept current.

Karpathy’s April 2026 LLM Wiki document is deliberately an idea file, designed to be pasted into an agent and instantiated with the user. Its architecture has three layers:

The paradigm moves knowledge work from:

That insight is significant because it transforms the output of AI research from an ephemeral answer into a durable artifact. Cross-references can already exist. Contradictions can already be marked. Valuable answers can be filed back into the Wiki. The system compounds.

Karpathy’s idea file offers the clearest expression of the pattern: Obsidian is the IDE; the LLM is the programmer; the Wiki is the codebase. It is attributed here as conceptual prior art; the observed gist does not provide a verified software license.

03 / What Wiki Engine changes

A compiler needs an intermediate representation.

Many implementations observed for this report move from a raw source into source summaries, entities, concepts, and graph links. Alpha.1 already inserts a one-source Signal reading layer; the target architecture further separates Source, Evidence, Signal, Wiki, Review, and Receipt.

In the target contract, Review and Receipt sit across that flow. Review records what a person accepted or rejected against an exact projection. Receipt records what the system did. Alpha.1 has read-only Review and operation receipts, but not the completed schema-2 contract claimed as stable.

The authority split

• The kernel owns bytes, identity, bounds, paths, lineage, state, and receipts.
• A human or model owns fallible interpretation and editorial judgment.
• Review owns local acceptance, not universal truth.
• The site owns explanation only; it cannot upgrade a claim beyond retained evidence.

This makes Wiki Engine less like an automatic note generator and more like a local research compiler: deterministic admission and publication around explicitly fallible semantic work.

04 / Side-by-side comparison

Same lineage. Different operating contract.

Wiki Engine is not a replacement for the original. It is one opinionated branch of the lineage: narrower, heavier, and intended for users who value traceability more than effortless automatic graph growth.

05 / The product wedge

Signal returns attention before the graph takes it.

A folder of saved sources preserves information but not attention. A fully automatic Wiki can create the opposite problem: dozens of synthetic pages whose maintenance and verification cost more than reading the source.

Signal is designed to answer the practical question:

A strong Signal report should expose:

One source should normally produce zero or one Signal. Wiki pages should earn cross-source life. That principle restrains generated page inflation and keeps the daily reading surface legible.

06 / Separation from the crowd

Not broader capture. An intended distinction in custody.

The Karpathy ecosystem is already active. Current public repositories package the formulation as agent skills, Obsidian templates, graph interfaces, multi-platform installers, extractors, and local workbenches. Adjacent hosted products compete for capture, summary, connection, chat, and learning jobs. The matrix below records public declarations and documentation observed for this report; competitor runtime behavior was not independently tested.

Wiki Engine should not claim superiority across these categories. Its proposed separation is the combination:

Relative to template-first LLM Wikis

Skill and template projects commonly optimize for setup simplicity, adaptation, and prompt-level portability. Wiki Engine accepts more machinery in pursuit of deterministic identity, correction history, migration discipline, and publication constraints. Whether that trade is better has not yet been tested with users.

Relative to graph-heavy systems

Graphs can encode typed provenance, evidence, review state, and much more than connection alone. Wiki Engine’s intended distinction is not “graph versus evidence”; it is that evidence custody and Signal review are the primary workflow, while visualization is a downstream projection.

Relative to hosted knowledge products

Recall’s public page declares distribution and convenience surfaces Wiki Engine does not offer. Wiki Engine’s proposed counter-position is custody: local files, replaceable model producers, no hosted account, no core telemetry, and explicit boundaries around external transfer.

Relative to capture tools

Obsidian Web Clipper’s official documentation demonstrates a mature durable-file capture surface. Wiki Engine should accept that donation rather than recreate it. Clipper preserves the page; Wiki Engine’s proposed job is deciding what becomes Signal and what eventually deserves cross-source Wiki life.

07 / All-angle assessment

Where the architecture is strongest—and where complexity accumulates.

08 / Honest boundary

A serious architecture is not yet a finished product.

The current system has meaningful local implementation work, tests, content-addressed source admission, Signal/Wiki publication, bounded video routing, linting, and transaction recovery. It does not yet possess the external evidence required for a stable significance claim.

• No meaningful public adoption or unassisted-user evidence.
• No stable semantic-quality benchmark across a held-out corpus.
• No mature GUI, mobile experience, or consumer-grade onboarding.
• No completed long-duration field soak or 100-source release witness.
• No final rights-safe promise for generalized YouTube acquisition.
• No stable Windows, multi-writer, network-filesystem, or active cloud-sync contract.
• No claim of defense against a hostile same-user process that coherently replaces all state.
• No guarantee that a citation is semantically sufficient merely because it is structurally valid.

The current Phase 1 candidate also reached a structural stop around referential lineage, explicit Wiki-to-Signal attribution, complete alpha video migration, and traversal-closed public schemas. Those findings are not hidden by this brand asset; they are precisely the kind of evidence the system is designed to preserve.

09 / Significance assessment

Influence, demonstration, and potential are different claims.

Karpathy’s formulation has visible influence: it supplied memorable language and a portable pattern that many public projects now explicitly cite. That does not erase earlier PKM, Memex, Wiki, provenance, or human-review work.

What would make the potential real?

1. Close the referential lineage and migration architecture.
2. Demonstrate Signal quality on licensed, representative mixed-source batches.
3. Prove correction, backup, restore, upgrade, and uninstall paths.
4. Run external users through the workflow without founder intervention.
5. Publish actual failure rates, citation errors, unsupported slices, and correction receipts.
6. Make first-run use simple enough that rigor does not become abandonment.

10 / The next system boundary

Wiki Engine should feed a second brain—not become the whole brain.

A second brain usually owns a wider human context: projects, identity, decisions, commitments, journal, plans, relationships, and long-lived knowledge. Wiki Engine should remain the research intake and evidence layer at its edge.

The bridge should be manual and disposition-bound at first. Proximity, keywords, or model confidence must not silently promote a Signal into personal authority. A second brain can cite or ingest a reviewed Wiki artifact, but it should preserve the Source/Evidence chain and the provisional status of interpretation.

Why this boundary matters

If Wiki Engine owns everything, it becomes another sprawling life-management system. If the second brain directly absorbs every raw capture, it recreates capture overload. The useful architecture is a selective membrane:

This report establishes the boundary, not the implementation. The second-brain integration is the next design chapter and should receive its own authority, routing, privacy, backup, and promotion contract.

11 / Position and moat

From saved sources to traceable signal.

The product should not claim to be the best second brain, the most automatic Wiki, or a universal memory system. It should not promise “save anything,” automatic truth, or hours saved without measured evidence.

Its strongest position is narrower:

The folder structure, prompts, and six nouns are all copyable. A defensible moat would come from:

• better Signal reports than generic summaries;
• reliable migrations and correction history;
• rights-safe mixed-media acquisition;
• cross-model portability;
• external proof that review is understandable and useful;
• a real corpus of preserved sources, failures, corrections, and successful restores.