WGRA Nuggets

Browsable bites on the ideas behind the Wiki-Grounded Research Agent

One idea per nugget: short, self-contained, and browsable in any order. The flyer makes the full case and the deep dives go long; these sit in between, a set of concepts you can graze rather than a story you follow start to finish.

Start anywhere. If you want the lineage, Karpathy’s llm-wiki is where the whole thing begins.

Kinetic PAD: Watching a drug test card develop

ecosystem
evidence

A paper drug test is normally read once, from a photo taken at the end. Photograph it every ten seconds instead and you record the whole reaction, and each lane turns out to develop on its own clock. The measurement came together quickly not from one tool but from an ecosystem: the PAD data reachable through an MCP server, a method reused from a sister project, and an llm-wiki that remembers across sessions and links them.

Aug 5, 2026

Software 2.5: developing when the same input gives different outputs

foundations

For twenty years the job assumed a program returns the same output for the same input. Agentic systems break that assumption, and the skills that made you good at deterministic software stop applying. This is the opener to a series on what changes, and the three problems it creates: measuring stability, propagating uncertainty, and deciding when a stochastic process has converged.

Aug 5, 2026

The variance cascade: how uncertainty compounds through agent chains

foundations
evidence

When one agent’s output feeds the next, the uncertainty does not add up the way intuition says. It compounds, and the correlations between stages can amplify it or cancel it. A pipeline can be three times noisier than its first agent, and also less variable than its worst one. You cannot reason this out from the parts; you have to simulate it.

Aug 5, 2026

0.988 is not a win: a Bayesian harness for agent evaluations

practice
evidence

A single answer count cannot tell a real improvement from run-to-run noise. This harness turns each trial into a (k,n) count, builds a posterior over the population success rate, stops sampling per input once the interval is tight, and only calls a winner when the evidence clears a deliberately high bar. In the worked example it refuses to: 0.988 is strong, and still not 0.99.

Aug 2, 2026

One run is one sample: the related-flood and the case for a Bayesian harness

practice
evidence

We built a test harness for llm-wiki behaviour and learned the hard way that one run is one draw from a noisy process. A 434-versus-52 edge swing looked like a plugin defect for a week; it was the serving path all along. Here is what the instrument measures, why single runs mislead, and why the statistics belong in an outer layer.

Aug 2, 2026

Where harnesses overlap, and what the difference costs

ecosystem
evidence

Claude Code, Codex, Cursor, and Antigravity ship one architecture and eight primitives. The surfaces have largely converged, but the wiring has not, and the difference is measurable: holding the model fixed and changing only what the harness hands the reader moved wiki-reading accuracy by twenty to thirty points.

Aug 2, 2026

What LLMs actually do with a research wiki

retrieval
telemetry

We keep a wiki as an agent’s durable memory. Reading the transcripts back across the corpus, we measured how the agent really uses it. It is written far more than it is read, answers come mostly from context already loaded rather than fresh lookups, and the typed links you author are written down, not read back to answer.

Jul 27, 2026

The shortcut and the tutor

practice
evidence

At the 2026 SCALE PI Symposium I showed the same student doing the same Lab 1 assignment twice. First with a generic LLM that hands over the answer; then with a wiki-grounded, Socratic tutor that will not. Both videos below. The instructor’s follow-up question is the same in both cases. Only one student can answer it.

Jul 25, 2026

Three agents, one training run

ecosystem
evidence

On 2026-06-15, three federated agents coordinated a full pass through an unfamiliar research task, from ‘I have a new dataset, what do I do?’ to a real cluster job that hit the exact failure mode the first agent had predicted. No human held state between them. The mailbox carried the coordination.

Jul 24, 2026

Catalog, structural, and content graphs

graph

A knowledge graph is rarely one graph, and these are not the only kinds you can build. But three recur, answering three different questions about a corpus: where things are, how a document is laid out, and what it actually says. Using the wrong one is why tasks like extraction quietly fail.

Jul 22, 2026

llm-wiki vs RAG: a first measurement

evidence

A first controlled test on one real research project: an agent navigating an llm-wiki answered questions more accurately than RAG over the same raw files, and with zero fabrications. Honest early numbers, with the caveats.

Jul 21, 2026

Muscle memory is not knowledge

foundations

Disconnect HAL and he regresses to the first song he was taught. Strip a modern LLM down and it keeps language, reasoning, a whole host of capabilities. That is muscle memory, not knowledge of your work, and the gap is measurable: 2% of a group’s questions answered bare, 90% with its wiki.

Jul 21, 2026

Standing up a research wiki

practice

The template hands you an empty wiki. What goes in first decides whether it is useful. A concrete recipe, worked on a real chemistry codebase: ingest what already exists, let the runs deposit the rest, and the payoff arrives when the next person does.

Jul 21, 2026

How to talk to a research agent

practice

The agent ships with its own discipline. What actually moved a year of retrieval research was how the researcher talked to it. Nine communication patterns, pulled from real session transcripts.

Jul 21, 2026

Keep what you got wrong

practice

Journals publish the win and quietly lose the walked-back claim, the abandoned method, the metric bug. A research wiki should do the opposite: file the correction next to the original, annotate the discrepancy, and let the wrong turn stay on the record. The trajectory is knowledge too.

Jul 20, 2026

Knowledge, expertise, experience

foundations

A good research wiki holds three different things: the facts, the judgment about the facts, and the grounded cases behind the judgment. Retrieval over your files reaches only the first. The other two are what make a wiki worth more than a search box.

Jul 19, 2026

Decentralized identifiers

identity

Online identity today is issued and controlled by platforms; a decentralized identifier is one you own and can prove, and it is what would keep the federation’s discovery open while access stays gated.

Jul 18, 2026

Data fabrics

ecosystem

A data fabric is a metadata-driven layer that unifies distributed data where it lives; our knowledge graph and connectors are what let a wiki-grounded agent plug into one.

Jul 14, 2026

ControlMaster, and the credential-broker pattern

infrastructure

SSH ControlMaster reuses one authenticated connection, its real purpose is speed. The pattern worth borrowing is the credential broker: the agent never holds the secret.

Jul 9, 2026

Connectors are agents

ecosystem

In data integration a connector is a static, pluggable adapter (think Kafka Connect, Fivetran). For interpretive extraction, ours are agents instead, and a small ontology per source is what makes them reliable.

Jul 2, 2026

A marketplace for tools

ecosystem

The app-store model applied to agent tools: a catalog for discovery, versioning, updates, and trust, so a capability is installed, not hand-wired.

Jun 27, 2026

Agent federation

ecosystem

Federation means autonomous systems that interoperate with no central owner; here each group’s agent publishes an A2A agent card, so peers can discover and query its memory with permission.

Jun 19, 2026

Knowledge bundles

bundles

Packaging a knowledge base so any LLM can consume it, portable, not locked to a tool. Google’s Open Knowledge Format is the emerging standard; a bundle is our instance of it.

Jun 11, 2026

The deep dives

deep-dives

Five longer treatments on the WGRA site: memory that compounds, team scale, agent capability, the knowledge graph, and knowledge bundles.

Jun 4, 2026

Our recipe: the pattern made concrete

foundations

Karpathy’s note is deliberately ‘the idea, not a specific implementation.’ The llm-wiki-memory-template is the implementation: the schema, the graph pipeline, the discipline gates, the operations, and the multi-writer protocol, all decided and shipped.

May 29, 2026

Ontologies, kept small

graph

An ontology is a formal vocabulary for a domain; real ones are often large. This project’s stance is the opposite, keep it small, and a real PAD example shows why that is what makes an agent’s reasoning reliable.

May 21, 2026

The wiki is a knowledge graph

graph

Every typed edge reifies to an RDF triple, so the wiki is a SPARQL-queryable graph, which buys queries prose and grep cannot answer.

May 14, 2026

Wiki front matter: the typed edges

foundations
graph

A few lines of YAML at the top of each page carry the relationships that turn linked markdown into something an agent can reason over.

May 6, 2026

Karpathy’s llm-wiki

foundations

The pattern underneath everything here: an LLM that builds and maintains a wiki as it works, so knowledge compounds instead of being rediscovered on every query.

Apr 22, 2026
No matching items