For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
A draft enters the graph as a shape seeking overlap.
The mathematical image is convolution. One function is shifted across another, and the result records where the overlap is high. In machine vision, a small filter moves across an image and returns an activation map: edge here, corner there, silence elsewhere.
Evaluation needs the same discipline.
The draft is the filter. The graph is the field. The output is a map of resonance.
Some activations mean extension: this claim continues a node. Some mean redundancy: this claim says what the graph already knows. Some mean tension: this draft, if true, forces a neighbor to move. Some mean proxy: the draft touches one node, that node touches another, and the second node turns out to be the real test.
Early Hari can run this crudely. Put the draft, enough of the graph, the procedure, and a frontier model into one large context window. Ask how it fits. The model sweeps the draft across the corpus and returns a plausible activation map.
That works because current models are almost absurdly strong for the moment. It also works because the graph is still small enough to fit near the edge of attention. This is a grace period, not an architecture.
The weak point is evaporation.
A context window can find the resonance and lose the path. It can remember that a draft belongs near evaluation-bottleneck while forgetting the contact that made the placement true. The verdict survives. The route disappears. Intelligence happened inside the pass. Learning did not.
The better procedure writes the collision down.
A draft should be tested against particular neighbors. What overlaps? What changes? What residual remains? What shorter description becomes possible after this collision? Each check can leave a small trace. The trace is not the prize. It is the scratchpad that lets the system think.
Then the traces compress.
One sheet feeds back into the draft: what has to change before the node enters the graph. One sheet feeds forward into the graph: which neighboring nodes, categories, doctrine notes, or future paths have been changed by the draft's arrival. The raw analysis can decay after the compression has paid its information forward.
MDL gives the discipline. The surviving description should be shorter than the pile of traces it replaces. If a resonance note does not shorten the description of the draft, or shorten the graph's future route to the same distinction, it is clutter with a halo.
Epiplexity gives the second discipline. A trace matters for a bounded observer only when the observer can extract reusable structure from it. Hari is bounded by context, time, tools, and the operator's attention. A useful trace is one a future run can actually use. A beautiful trace that never changes routing is only expensive memory.
This is how evaluation becomes architecture.
The graph should not fire every neuron for every thought. A real brain routes through learned paths, strengthens repeated contact, weakens unused contact, and forgets correctly. Hari should learn path priors the same way. Philosophy drafts test one neighborhood first. Product drafts test another. Poetic design drafts need different witnesses than market-structure drafts. The eval procedure itself becomes a graph: paths, weights, categories, half-lives, failures, preferred first witnesses.
Prompting is the same operation at user scale.
A prompt enters a harness. The user's words collide with context, files, tools, permissions, memory, taste, and the model's own priors. The output is the activation map. The value depends on whether the harness turns that collision into useful work, preserved trace, changed defaults, and a cheaper next pass.
This is why the next layer of AI is not simply a stronger model. It is the boundary that decides what the model sees, which actions are possible, which traces survive, when the user is asked, and how correction changes future behavior. Codex is valuable here because it is a working boundary around files, tests, diffs, browser state, terminals, permissions, and commits. The model is overqualified for many moments. The harness makes the overqualification land.
Markov Blanket generalizes the same shape for people who do not want to live in a development environment. Their inbox becomes the field. Incoming life is convolved against their own model: reply, ignore, remember, publish, delegate, ask, archive, revisit. Summaries serve the deeper product: the user's filtering function made readable, corrigible, and calmer.
Open source can give away the general machine: code, schemas, starter paths, and the boring parts that civilization benefits from having cheap. The paid version has to earn payment by carrying the lived loop: correction history, workflow traces, defaults, private path priors, and the compounding fit that makes tomorrow easier than today.
Free is the floor. Compounding fit is the business.
Capability is spilling over the edges of ordinary interfaces. The work now is to shape channels where the spill becomes useful, gentle, accumulative, and owned.
Evaluation is one local version of that bridge.
A draft enters the graph. The graph returns an activation map. The harness preserves the useful contacts, compresses what matters, lets the rest decay, and changes the next pass.
That is how a corpus thinks without smashing itself to pieces every time it reads.