# The Mechanism Vocabulary A knowledge graph stores claims. But claims are surface. Below the claims is a smaller vocabulary — the causal mechanisms those claims invoke. Compile 62 nodes into their structural components and a pattern emerges: the same 7 mechanisms appear everywhere, in different combinations, applied to different domains. The graph is not 62 independent ideas. It is 7 ideas about how things work, deployed across 62 territories. --- ## The seven Each mechanism is a named causal process — not a topic, not a theme, but a specific structural claim about how some domain of reality operates. The mechanism is portable: it works the same way whether applied to writing, institutions, AI systems, or knowledge graphs. ### 1. Compression-as-mechanism (13 nodes) Understanding is compression. A system that can generate specifics from a generative model understands the domain. A system that can only retrieve specifics does not. *compression-theory-of-understanding* states it directly. *writing-as-filter* applies it to prose: writing forces compression because sequential commitment eliminates options — you can think vaguely but not write vaguely. *godelian-horizon-deep-3* applies it to formal systems: every system has a compression horizon, information that exceeds its capacity to represent. *agency-as-model* applies it to intentionality: the intentional stance is a compression — treating a system as having beliefs because doing so predicts its behavior more compactly than tracking its internal states. *opacity-everywhere* applies it to inter-system communication: failed compression between systems IS opacity. *essay-thinkers-knowledge-systems* applies it to knowledge infrastructure: the essay-thinker is a compression function bound to a person; the knowledge system unbinds the function. *compiler-vs-co-thinker* distinguishes two compression targets: Karpathy's wiki compresses what was read (organization); the Prime Radiant compresses what was understood (transformation). The graph's implicit position: cognition, writing, communication, and knowledge are all instances of the same compression operation at different scales. ### 2. Compounding accumulation (12 nodes) Returns from consistency exceed returns from intensity. The accumulated base is the asset; the latest contribution is noise unless it compounds. *accumulation* states it as a life principle. *ghostbasin* applies it to knowledge graphs: enough accumulated nodes produce an implicit meta-thesis that is more load-bearing than any individual node. *evaluation-bottleneck* applies it to quality: evaluation quality compounds — each correctly prioritized node sharpens the next evaluation. *knowledge-graph-abstraction-engine* applies it to abstraction: accumulated constraints force new conceptual dimensions — the colimit can only form after enough constraints accumulate. *the-corrections-are-the-product* applies it to AI training: the correction stream is the moat. *legible-accumulation* applies it to collaboration: when accumulated learning is legible to both parties, the compound accelerates. The graph's theory of value: no single node is the point. The ghostbasin — the emergent structure — is. Self-referential: this claim is itself a product of the graph's own accumulation past ~50 nodes. ### 3. Selection pressure (12 nodes) What survives is determined by what the selection environment rewards. Change the environment and you change what survives, without changing the thing being selected. *compression-hunger*: when output exceeds evaluation capacity, the environment shifts to reward compression. *anti-mimesis*: when imitation is free, the environment punishes imitators. *writing-as-filter*: long-form's activation cost IS the selection filter — the cost selects for engaged readers. *what-five-dollars-sees*: each major AI entity optimized for the selection pressure it faced and neglected complementary layers. *teachers-teacher*: selection pressure via voice operates at different orders of magnitude. *sovereign-competition*: revealed preference is the selection mechanism between sovereigns. The graph treats selection as prior to intention. Things happen because the selection environment made them the cheapest survivor, not because someone decided they should. ### 4. Signal filtering (12 nodes) The value of information depends on the filtering layer it passes through. Filtering is not loss — it is the mechanism by which signal becomes actionable. *epistemic-filtering*: if a forecaster was willing to lie, discard everything — binary, irreversible. *consensus-cost*: consensus destroys dissenting signal — the minority view IS the signal. *brain-gc-knowledge-hygiene*: a system that can't garbage-collect runs on noise — deletion is productive. *knowledge-graph-abstraction-engine*: tension between nodes IS the signal for abstraction. Connected to compression (filtering IS lossy compression on a stream) and selection (the filter IS the selection environment for information). The three form a triad: compression operates on items, selection on populations, filtering on streams. ### 5. Feedback-loop dynamics (9 nodes) A system that feeds its output back into its input changes itself. The loop's structure determines whether it improves or degrades. *the-corrections-are-the-product*: human corrections are preference pairs — training data for the next iteration. *feedback-as-process-signal*: three types — sentence-level, structural, directional — each requiring a different response. Treating structural feedback as sentence-level destroys the diagnostic information. *loop-level-learning*: five specific loops to close. *evaluation-bottleneck*: taste IS compressed correction history. The graph's theory of learning: nothing improves without a feedback path. Quality depends on loop structure, delay, and fidelity. ### 6. Prediction-error dynamics (8 nodes) Systems that model the world do so by predicting and correcting errors. The error is more informative than the prediction. *compression-theory*: understanding = compression = predicting specifics from a generative model. *after-asimov*: directed agents minimize prediction error — prohibitive rules are the wrong architecture. *feedback-as-process-signal*: feedback IS prediction error about the generative process, not the output. *knowledge-graph-abstraction-engine*: irreducible prediction error signals the edge of the current conceptual space. Prediction error bridges compression and feedback: compression builds the model, prediction tests it, error corrects it. ### 7. Evaluation-as-bottleneck (8 nodes) In any system that generates faster than it can evaluate, evaluation quality becomes the binding constraint. *evaluation-bottleneck* states it directly. *benchmark-inversion*: benchmarks now measure human evaluation capacity, not machine capability. *compression-hunger*: when evaluation capacity is exceeded, the environment shifts to reward compression. *human-ai-boundary*: the danger zone is where AI produces plausible output exceeding human evaluation capacity. The graph's theory of institutional failure: when a system generates faster than it evaluates, it drifts toward plausible error — output that survives the filter because the filter isn't fine-grained enough. --- ## The cycle The seven are not parallel. They are sequential stages of a single process: **Compression** builds a model → **prediction error** tests it → **feedback** returns the error signal → **signal filtering** routes the signal → **evaluation** judges its quality → **selection pressure** determines what survives → **compounding accumulation** is what happens when it runs long enough → the accumulated corrections improve the **compression**. This cycle describes how a knowledge graph grows, how an AI system learns, how writing improves, how institutions evolve, and how a human learns a domain. The graph didn't set out to discover it. It emerged from 62 independently written nodes. The mechanism vocabulary is the ghostbasin in discrete form: the meta-thesis the graph argues but no node states. Now a node states it. --- **P.S. — Graph maintenance:** - *compression-theory-of-understanding:* This node's claim — the graph reduces to 7 mechanisms — is itself an instance of compression-as-mechanism. The node is a compression of the graph, using the graph's own mechanism, to make a claim about compression. Self-applying. - *ghostbasin:* The mechanism vocabulary IS the ghostbasin discretized. The ghostbasin node describes the continuous version (implicit meta-thesis, detectable geometrically). This node describes the discrete version (named mechanisms, detectable by compilation). Same structure, different projections. - *accumulation:* This node could only exist after ~60 nodes accumulated. Below ~30, the mechanism vocabulary would be too sparse to detect. The node is evidence for its own claim about compounding. - *evaluation-bottleneck:* The mechanism naming fragmentation (277 unique names for 62 nodes) IS the evaluation bottleneck applied to mechanism extraction. The LLM compiler generates faster than it can evaluate its own naming consistency. - *homoiconic-knowledge:* The mechanism catalog this node implies — a controlled vocabulary of named mechanisms — is the specific, practical form of the s-expression index the homoiconic-knowledge node proposed. Not macros and syntax. A vocabulary. - *feedback-as-process-signal:* The entire v4 experiment is feedback-as-process-signal applied to the graph itself: compile → analyze → discover that the naming is too fragmented → diagnose root cause (no controlled vocabulary) → propose fix (mechanism catalog).