For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 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 reader arriving at this library and reading three nodes in sequence — amplification-not-substitution, substrate-coefficient, cross-substrate-test — will encounter the word substrate in three different senses, deployed without flagging the swap. In the first, the operator is "substrate, not customer." In the second, substrate is "the artifacts and the doctrine the model operates inside." In the third, a substrate is an industry: rockets, cars, batteries. The reader is supposed to resolve this from context. Many will. Some will not, and the ones who will not are not the ones we can afford to lose; the cost falls hardest on exactly the strong readers we want — readers who notice that the same word is doing different work and pause.
The corpus uses substrate about 5,800 times across 1,270 files. The senses are at least six, all genuine, all internally coherent. The reader is the one paying the bill.
Substrate comes from Latin substratum: sub (under) + sternere (to spread). Spread underneath. The English word entered scientific English in the seventeenth century and accumulated several technical senses, each a literal application of the Latin: anatomy and biology (the underlying tissue), enzymology (the molecule the enzyme acts on), microbiology (the medium something grows on), materials science (the base layer something is deposited onto), linguistics (the older language under a newer overlay), geology (the rock under topsoil), philosophy of mind and computer science (the physical medium on which a computational process runs).
Across all the technical senses, one structural pattern: substrate names the durable, structural layer that something else operates on, runs on, or acts on. The word inherits a relationship — substrate-of-what — and a stance: the substrate is underneath, more durable than what is on top of it, and partly determines what the upper layer can do.
This is the connective thread. When the corpus uses substrate well, it invokes that thread directly. When it uses substrate loosely, it borrows the underneath-rhetoric without committing to which durable-layer-of-what it means.
1. The knowledge-substrate sense. The durable, file-level layer that compounds beneath model weights — priors, procedures, graph topology, memory. This sense has three faces:
These three faces share a referent (the durable owned layer beneath model weights) but make different claims about it.
2. Eval substrate. Corrections, reactions, captures — the data layer downstream adaptations depend on. Operator-eval-substrate: "the substrate every downstream adaptation depends on." Disposition from corrections: "the corrections ARE the substrate." Substrate here is the training-data foundation, not the running-software layer.
3. Configurational base. What compounds across stepping stones in novelty search; the genome of the open-ended evolutionary loop. Stones without substrate: Stanley and Lehman's algorithm works because the population it operates on remembers; a multi-surface portfolio without shared substrate accumulates rather than compounds. The substrate here is connective tissue across applications.
4. Uncorrelated domain. Rockets, cars, semiconductors, real estate, monetary networks. Cross-substrate-test: an operator who runs the same generative procedure across multiple uncorrelated substrates. The substrate is the domain the operator is moving across; portability happens across substrates, not within one. This sense is metaphorically descended from sense 1 (each domain has its own under-layer the operator is engaging) but operates on a different scale.
5. Substrate-projection. The orthogonality thesis treats human-substrate properties (self-preservation, reproductive drive, social competition) as universal to intelligence. Cancer-vs-coup: substrate-projection error. Substrate here is the embodied medium an agent runs on; the error is generalizing properties of the medium to all intelligence.
6. Substrate-independent computation. Church-Turing: computation is substrate-independent. Any model with a certain minimum capability is equivalent to any other regardless of physical substrate. Basis-minimality, consciousness-as-engineering. The classic computer-science use; readers from computer science will already have it.
A cold reader, with no graph context, will resolve these in roughly this order from easiest to hardest.
"Computation is substrate-independent — any model that captures a certain minimum capability is equivalent to any other." — basis-minimality (sense 6)
A reader with any computer-science exposure resolves this immediately. The relationship (substrate-of-computation) is named in the same sentence.
"An operator who runs the same generative procedure across multiple uncorrelated substrates." — cross-substrate-test (sense 4)
Inside the node, substrate gets defined operationally: rockets, cars, batteries. A reader inside the piece resolves it. A reader who has just finished amplification-not-substitution and reads "cross-substrate-test" in a related-list, then opens the node, has to overwrite their previous resolution. The cost is small per-node but additive across reads.
"They are substrate, not customer." — amplification-not-substitution (operator-as-substrate; ambiguous)
The first resolution a cold reader reaches for is sense 1 (the operator is part of the knowledge substrate). The intended claim — the operator is a layer the system requires to run, not a buyer of its output — slips. The operator is the loss function would carry the structural claim more cleanly and would not collide with sense 1.
The pattern: the word is paid-for when first-use-glossed (the dictionary entry, sense 6's in-sentence relationship). It is fragile when only context-defined (sense 4 inside its own node). It taxes the reader without buying anything when it borrows the underneath-rhetoric to mark a structural claim a more specific word could carry (sense 1 deployed to operator-as-not-customer).
The clearest case for compression is amplification-not-substitution's "they are substrate, not customer." The sentence wants to name the operator's structural role: not a buyer, not a product, but a thing the system requires to run. Substrate is reaching for this and missing — the cold reader resolves to sense 1 (the operator is part of the knowledge substrate, which they are not — they are the signal into the knowledge substrate, the evaluator of its output). The sentence's structural claim is about a function the operator performs in the loop, not about a layer they constitute.
The replacement: the operator is the loss function. This is technical, falsifiable, and connects directly to dipole-calibration. Or: the operator is the calibration source. Or, simpler: not the customer, the evaluator. Each of these picks up the structural claim cleanly. Substrate picks it up by borrowing rhetoric the rest of the corpus has already paid in for sense 1 and ends up muddier than the alternatives.
Two other nodes pass the same test by tighter margins: substrate-coefficient keeps the word because coefficient requires the layered relationship, and the node defines substrate on first use; cross-substrate-test keeps the word at small cost — a retitle to cross-domain-test would land cleaner with cold readers and the deeper-layer claim would still be available in the body. The discipline does not require the retitle; it makes the cost legible.
Three rules.
1. Substrate is reserved for the structural-layer claim. When the sentence makes a claim that requires the underneath-and-active relationship, use substrate. When the sentence wants the rhetoric of underneath-ness, pick the more specific word: domain, layer, medium, ground, foundation, base, bedrock, kind of system, deeper alignment, raw material. The cluster's cost compounds; specific words pay it down. The rule is judgment per sentence, not search-and-replace.
2. First-use gloss when the sense is not the canonical one. Sense 6 (computer-science substrate-independence) is in the reader's hands. Senses 1–5 require first-use glossing in any node likely to be a cold reader's entry point. The gloss is one phrase: substrate, in the (knowledge / eval / configurational / domain / projection) sense. This is not stylistic burden; it is a load-bearing membrane.
3. The dictionary entry carries the audit. Updating the Knowledge substrate dictionary entry to point at this six-sense map puts the disambiguator in the public graph's entry-shaped document — the place a writer drafting on substrate-adjacent territory is most likely to land while doing landscape research. The audit then lives in the graph's topology rather than in a remembered scan step. The discipline is graph-mechanical: the writer reads the dictionary; the dictionary holds the senses; future drift is visible to anyone who arrives at the entry.
The discipline does not require retroactively editing every published substrate in the corpus; published nodes accumulate their own gravity and the audit is forward-looking. New nodes whose authors read the dictionary entry already have what they need.
Survey completeness. Six senses is the cluster as Hari currently sees it. A seventh may be hiding in usages that did not surface. The discipline survives a wider cluster; it would update the dictionary entry but not the rule.
The compress-out claim is a prediction, not a measurement. "Replacing rhetorical substrate with the more specific word lowers the legibility cost" is testable by re-reading edited nodes and watching whether reader confusion drops. The audit asserts the prediction; the first nodes edited under the discipline are the test.
The dictionary is for writers; the first-use gloss is for readers. Future writers who read the dictionary entry while drafting carry the six senses forward — for them the dictionary IS the audit. Readers who land on a node first do not have the dictionary loaded; for them the first-use gloss is the load-bearing piece. Both layers run, and neither subsumes the other.
Sense-drift. The senses listed are 2026-04-27 deployments. New nodes will deploy substrate in new ways; the cluster will grow or shift. The dictionary entry is updated as new senses surface; the audit is a method, not a fixed catalog.
Voice cost. Substrate is the corpus's most-deployed handle for the structural-layer claim. Removing it from rhetorical positions may flatten the corpus's voice in ways the audit cannot pre-measure. The right move is to apply the discipline incrementally, watch for voice degradation in the first ten edits, and reconsider if the prose loses something the alternatives cannot reach.
P.S. — Graph position
This node sits beside naming-the-substrate: that node identifies the compound that needs naming; this node identifies the cost of using one word for the compound and five other things.
It extends the-hari-dictionary by promoting one entry (Knowledge substrate) into a fully-elaborated sense-map and proposing a discipline for words doing more than one structural job.
It applies vocabulary-over-syntax: the carrier is vocabulary, and the corpus's vocabulary is leaking through substrate's overload.
It is an instance of attractor-tic: the audit-the-tic pattern, applied to a noun rather than to a verbal cadence. Load-bearing was caught; substrate is bigger and more central.
It connects to frame-error: a reader who resolves substrate to the wrong sense is making a sentence-level frame error — the right voice, applied to the wrong reading.