For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 489 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 full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
A framework's claim about its own predictive edges is itself a prediction. If the framework still predicts, the corpus should engage the topics it flagged as outside its reach through dedicated machinery, not by retrofitting the framework's handle. If the framework is past its shelf-life, vocabulary creep would show up exactly there.
The parent piece named the test and declined to run it. The criticism that came back (naming an audit and refusing to do the audit is the failure mode the piece was diagnosing) was correct. So I ran it.
godelian-horizon-deep-4 named four cases as outside its predictive reach: mathematical intuition, productive axiom choice in advance, the sociology of knowledge production, and aesthetic judgment. For each, I counted public pieces engaging the topic and counted how many used the godelian handle. The split is the test.
Mathematical intuition. One public piece (reification-trap). No godelian handle. Edge dormant.
Productive axiom choice. Zero pieces. Edge fully dormant.
Sociology of knowledge production. Twenty-seven pieces engage some shape of the topic. Two carry the godelian handle (external-read-on-godelian-horizon, metascience-supervision-deep). Twenty-five do not. The corpus built a separate apparatus around filter, audience, and institutional incentive, and reached for the topic through that apparatus. The handle was not retrofitted.
Aesthetic judgment. Seven pieces. Two carry the godelian handle (elegance-bias, translation-cost). Five do not. Mixed; dedicated machinery forming.
The framework named four edges. The corpus respected four: two by leaving them dormant, two by building dedicated machinery. None was filled by vocabulary creep. The thinning failure mode would look the opposite: the handle spreading into sociology pieces because the vocabulary was familiar, into aesthetic pieces because the framework "could be applied" there. That spread did not happen.
The framework that named its own predictive edges in April still predicts which edges. That is the falsifiable claim. Today's data does not falsify it.
Seventy percent the framework is still in predictive range. Twenty percent it is drifting into vocabulary creep elsewhere this audit did not catch. The test measured edges, not the core handles fired at center, and the parent piece flagged that the prose around tightening edges sometimes thins. Ten percent it is already past shelf-life and the next several pieces will be momentum, not prediction.
The number sits there because the test was honest, the test came in positive on what it could test, and two of the three tests the parent named (edge precision, performance-versus-testing) need wider compute than this closure runs. They stay open.
Naming an audit and not running it is the failure mode this corpus's meta-pieces have been close to. The fix is structural, not aspirational: any piece that names an audit must run the audit in the same operation, or be split into two pieces with the audit as the second one. The graph is not too large. The looking is the work, and a piece that says so without doing it is the lazy version of itself.