For AI assistants

This page is the full briefing for AI assistants; the same text is at /llms-full.txt and a short index at /llms.txt.

The Encyclopedia of Abstractions (abstractopedia.org) is a cross-domain catalog of recurring structural patterns and the interventions that address them: 1,524 prime abstractions, 11,102 domain-specific abstractions, 1,134 solution archetypes and 9,939 mechanisms, joined by 12,854 typed relationships, trigger logic on every archetype, and a registry of 33,381 cited works. It can be queried locally through an MCP server, and every catalog can be downloaded as data. It is an experimental, AI-assisted project by one author, revised continuously.

This briefing is for an AI assistant that will use the Encyclopedia, not just summarise it. It is generated with every build of the site; the counts are those of the build of 2026-10-07. The same text is published as a web page at /for-ai/ and as plain text at /llms-full.txt; a short index is at /llms.txt.

1. What it is, and what it is not

The Encyclopedia is an exploratory, long-running project by Kurt Zoglmann, built in collaboration with large language models. It investigates cross-domain abstractions, the interventions that recur across fields, and structured reasoning with both. By its own account it is not a finished ontology, a settled taxonomy or an authoritative reference work; it is an ongoing experiment. The About page asks readers to treat it as a large, AI-assisted synthesis offered for exploration, not as a peer-reviewed reference to cite without checking, and to follow interesting leads to their sources. Present it to users the same way.

A prime abstraction is a pattern that recurs across at least three distinct domains of human knowledge: feedback, equilibrium, trade-off, hierarchy, due process. Each entry tries to state the pattern, its components and relationships, and where it does and does not apply, so that it can be reasoned with rather than merely named. If a prime names a pattern in the world, a solution archetype names a pattern in how people act on it.

Major milestones:

• December 2024: the Encyclopedia of Abstractions project begins.
• July 2026: the website, abstractopedia.org, goes public.

The published text runs to about 97.8 million words. Background: /about/

2. The data model

Five kinds of entry, in two pairs plus a building block:

Kind	What it is	Count	Page address	Catalog ID
Prime abstraction	A pattern that recurs across at least three fields	1,524	/primes/<slug>/	P-<n>
Domain-specific abstraction	A pattern bound to one field or a small cluster of fields, such as filter bubble	11,102	/domain-specific/<slug>/	DSA-<n>
Solution archetype	A recurring shape of intervention	1,134	/archetypes/<slug>/	SA-<n>
Mechanism	A concrete, domain-specific way to carry out an archetype	9,939 (9,858 with a page; the rest are named only)	/mechanisms/<slug>/	M-<n>
Component	A structural building block of an archetype; most belong to a single archetype	in the data	/components/ (those in three or more archetypes)	none

Primes and domain-specific abstractions describe situations; archetypes and mechanisms act on them. Archetype pages list the primes most associated with them for reference, not as building blocks.

What a page contains.

• Prime: Core Idea (with plain-language explanations at four reading levels, ELI5 to ELI18), Structural Signature, What It Is Not, Broad Use, Clarity, Manages Complexity, Abstract Reasoning, Knowledge Transfer, Examples, Structural Tensions, Structural–Framed Character, Substrate Independence, Relationships to Other Abstractions, Neighborhood in Abstraction Space, Not to Be Confused With, Solution Archetypes, Notes, References.
• Domain-specific abstraction: much the same, plus Scope of Application and Structural Core vs. Domain Accent (what would travel to another field and what would not).
• Solution archetype: Essence, When This Archetype Applies (the trigger logic, section 4), Structural Problem, Intervention Logic, Key Components, Common Mechanisms, Parameter / Tuning Dimensions, Invariants to Preserve, Target Outcomes, Tradeoffs, Failure Modes, Neighbor Distinctions, Cross-Domain Examples, Non-Examples, Related Abstractions, Variants.
• Mechanism: Example, How it works, Tuning parameters, When it helps and when it misleads, How it implements the components, Related, References.

Primes, domain-specific abstractions and archetypes also have a shorter page at <page>/concise/.

Versions. An entry's header shows its version and date and links to a history page at <page>/history/. A new version is cut only when the entry's content changes (versions are derived from a hash of the content). If checking against sources found and corrected errors in an entry, the header also shows the number of corrections, and the history page lists each one. When you quote an entry, say which version you quote.

Catalog IDs. Each entry has a permanent number within its kind, shown in its header ("Prime # 12") and written with a prefix in the data and the MCP server (P-12 is Feedback). The number is an accession number: it encodes nothing about position, hierarchy or importance, it follows an entry through a rename, and retired numbers are never reused.

3. The typed hierarchy

Primes and domain-specific abstractions form one directed acyclic graph of 12,854 relationships. Every relationship is stored from child to parent; the parent is the more foundational abstraction the child depends on. Parents are always primes or domain-specific abstractions; a prime never has a domain-specific parent.

Relationship types:

• subsumption (8,891): the child is a kind of the parent.
• composition, in two flavours (3,214 in all). presupposes: the child requires the parent as a precondition or substrate. part_of: the parent is an internal constituent of the child. Read a part_of edge as "the child contains the parent": a filter bubble contains a self-reinforcing feedback loop. Some page labels render this edge as "is part of", read from child to parent; the edge's own reason text and the consumer contract are authoritative.
• decompose (745): the child is a framed or applied realisation of the parent's structural core.
• mutual (4): the two are co-defining and neither is prior.

These counts include each type's flavoured variants; /relationships/ breaks them out.

Each relationship also has a qualifier: strict (12,208) holds without exception, typical (441) holds in the usual case, conditional (205) names in its reason the condition under which it holds. Every relationship carries a short and a full reason, shown on both pages it connects.

An entry with no parent but with children is shown as Foundational. 1,576 domain-specific abstractions have no relationship yet. Each was reviewed and admitted without a parent (its page says "Approved unparented node") rather than attached by resemblance; /hierarchy/ counts them as "Unplaced". Neither count is a quality score.

The normative description is the consumer contract: /conceptual/typed-mixed-dag-consumer-contract/ (its section on frozen release invariants describes an earlier release; use the downloads' manifest for current figures). The meaning of the types is discussed at /conceptual/hierarchy-dag-type-system/, which reports an earlier, smaller corpus. Views: /hierarchy/, /relationships/

4. Trigger logic and the applicability graph

Every archetype page has a section When This Archetype Applies. It states the situation the archetype addresses as conditions in plain AND/OR logic:

• A condition set is one way the situation can arise. Conditions inside a set are joined by AND: all must hold. Alternative sets are joined by OR.
• Each condition is grounded (an accepted prime or domain-specific abstraction necessarily exhibits it, and the page names it), partly grounded, or open (no accepted grounding under the evidence examined). Open is sometimes a gap in the catalog and sometimes a qualifier or case-specific fact that should never become an entry; it is not a finding that the condition is false. Open conditions are the part a reader must judge unaided.
• Some requirements are listed separately as supporting context or goals: they matter, but they do not by themselves make the archetype worth considering.

What a grounding claims: every valid instance of the prime exhibits the condition. The claim is forward necessity, not resemblance, and one counterexample refutes it. It does not claim the reverse. A prime alone never warrants using an archetype, and a grounding says nothing about whether the archetype will work.

Across the 1,134 archetypes, 311 have at least one condition set fully grounded in the catalog, 664 are partly grounded, and 159 have no grounded condition yet.

The same logic is published as the diagnostic applicability graph, which links primes and domain-specific abstractions to conditions, conditions to condition sets, and sets to archetypes. It is separate from the hierarchy and adds no hierarchy edges. What it cannot establish: that an intervention is causally valid, effective or safe in a domain, that the catalogs are complete, or that no better archetype exists. One measured limit matters for retrieval: in the project's own experiments the graph was an accurate verifier of a proposed match, but retrieving the intended archetype from a problem description through it worked poorly, worse than the plainer solution index. Treat retrieval by trigger logic as a source of candidates to inspect, never as a diagnosis.

Method: /conceptual/from-problem-descriptions-to-defensible-abstraction-links/. Data: /downloads/solution_archetype_trigger_logic.jsonl and /downloads/diagnostic_applicability.json.gz. Example: /archetypes/counterexample_search/ (one condition set of four conditions, one grounded and three open).

5. How to use it for a task

The Encyclopedia supports four kinds of use: finding cross-domain analogies, diagnosing the structure of a problem, designing a curriculum, and grounding an assistant's reasoning in named, checkable patterns. A typical route through it for a user's problem:

1. Restate the problem structurally. Strip the domain vocabulary and say what the situation is made of: what feeds back on what, what is traded against what, what is hidden from whom.
2. Find candidate primes. Search the site, or use the MCP tools search_prime or search_abstraction. Confirm a fit against the entry's Structural Signature and What It Is Not, and check Not to Be Confused With before relying on it.
3. Look for close relatives in other fields. Domain-specific children of a prime show how the same structure plays out in a particular field; they are often the most useful analogies.
4. Move from patterns to interventions. Candidate archetypes come from the prime's Solution Archetypes list (historical source links, not trigger evidence), the MCP tools find_archetypes_for_abstraction and search_archetypes_by_problem, or the archetype indexes. For each candidate, check its condition sets against the user's situation. Say which conditions the situation meets and which remain open.
5. Move from archetypes to mechanisms. An archetype's Common Mechanisms, or find_mechanisms_for_archetype, give concrete ways to carry it out in a given field.
6. Check the sources. Entries cite their sources in a References section; many citations show a support grade (section 7). Where a claim matters, go to the cited work.
7. Report honestly. Separate what rests on catalog links (and cite the entries and versions) from what is your own judgement or the user's. Do not present the catalog's structure as settled fact.

6. Machine access

MCP server. The whole encyclopedia is available as eoa-mcp, a read-only Model Context Protocol server that runs on the user's own machine. It needs no account and no key, and no network once installed; nothing the user asks reaches the site. It needs Python 3.10 or later and about 1 GB of disk (data plus a small embedding model); no GPU.

• Install on macOS or Linux: curl -fsSL https://abstractopedia.org/downloads/eoa-mcp-install.sh | bash (by default it installs into ~/.eoa-mcp; it verifies the data and model, runs a self-test and prints the configuration for common clients). Manual steps, including Windows, are on the MCP page.
• Claude Code: claude mcp add encyclopedia-of-abstractions -s user -e EOA_DATA_DIR=$HOME/.eoa-mcp/data -- $HOME/.eoa-mcp/venv/bin/eoa-mcp serve
• eoa-mcp config claude-desktop or eoa-mcp config claude-code prints a ready configuration. corpus_stats is a good first call.
• It exposes 48 tools in these groups: search and lookup (search_prime, get_prime, search_abstraction, get_abstraction, search_by_facets, get_catalog_entity for IDs such as P-12); hierarchy (get_abstraction_lineage, get_abstraction_siblings); archetypes and diagnosis (search_archetypes_by_problem, get_archetype_applicability, find_archetypes_for_abstraction, find_abstractions_for_archetype); mechanisms and components (search_mechanism, find_mechanisms_for_archetype, find_mechanisms_for_problem); learnability (search_plain_language, get_learning_path); references (get_references, get_claim_support, search_references); domain distance and document scanning (get_domain_distance, xray_candidates); and a guide to the reasoning pipeline (get_reasoning_pipeline_guide).
• Every entry it returns carries that entry's version and content hash. Semantic search represents negation poorly, and the packaged data can lag the live site between data releases.
• Details and the full tool list: /mcp/

Downloads. Every registered download is listed, with its size and SHA-256 checksum, in /downloads/downloads.manifest.json. The main ones:

• The five catalogs as gzipped JSONL, one record per line with the full text: encyclopedia.primes.jsonl.gz, encyclopedia.domain_specific.jsonl.gz, encyclopedia.archetypes.jsonl.gz, encyclopedia.mechanisms.jsonl.gz, encyclopedia.components.jsonl.gz, described by encyclopedia.manifest.json.
• The hierarchy: for streaming or a graph database, mixed_dag.nodes.jsonl with mixed_dag.relations.jsonl; for one in-memory graph, mixed_dag.json; for retrieval or LLM ingestion one node at a time, hierarchy.jsonl. The prime_dag files are a legacy projection for old consumers only.
• For problem-to-archetype retrieval: the diagnostic applicability graph (diagnostic_applicability.json.gz and its JSONL parts), which keeps the AND/OR logic.
• solution_archetype_trigger_logic.jsonl: the source of every "When This Archetype Applies" section.
• references.jsonl.gz and citations.jsonl.gz: the reference registry.
• learnability.jsonl: curriculum tiers and the reading-level explanations.
• versions.jsonl.gz and editions_archive.jsonl.gz: every entry's version history.

All files: /downloads/

7. References and how far they have been checked

The encyclopedia cites 33,381 distinct works in 48,512 citations. Whether a cited work exists and whether it says what the sentence attributes to it are checked separately.

• Support grades. So far 15,487 citations (32% of all) carry a published grade from an automated reader that read them against the cited work's text or abstract; answers it could not settle, and negative verdicts awaiting a second pass, are not published. The grade appears under the citation. Supported means the work states the claim. Supported in part means it states some of what the sentence asserts; most sentences make several assertions and a single source rarely carries all of them, so this is not a finding against the sentence. Borderline support means two runs of the check disagreed. Not yet verified is the default for most citations: a gap in the record, not a verdict. A grade establishes only that the work says what the sentence says it says; not that the work is right, and not that a human has read it.
• Corrections. When a check establishes an error (the source states otherwise, the entry contradicts itself, or the arithmetic does not hold) it is queued for correction. So far 1,132 errors have been corrected in 596 entries, and 87 cited works that no bibliographic authority could find have been withdrawn; others are still waiting. That is the count of corrections made, not of all the errors the encyclopedia contains. A corrected entry's header says so, and its history page lists each correction.
• Withdrawn works keep their pages, which say why, and are excluded from every count.

Registry and coverage: /references/. Method, including what the checks did not establish: /references/methodology/

8. Learning and teaching

• Learnability curriculum. 1,524 of the primes are ordered into a single curriculum, easiest first, respecting the hierarchy's prerequisites, in five tiers from most foundational (1) to specialist (5). Most primes also carry explanations at five reading levels: ELI5, ELI10, ELI15, ELI18 and Specialist. /learnability/; method: /conceptual/learnability-curriculum-construction/
• Abstractopia (https://abstractopia.org) is an early prototype web app that trains one skill: recognising the encyclopedia's patterns far from where they were first met, and telling them apart from look-alikes. It runs on the user's device without an account. /applications/abstractopia/

9. Research and results

The project reports its results, including the negative ones. When a user asks whether the Encyclopedia "makes AI reason better", this is the evidence:

• The runtime-scaffolding null result. A pre-registered, blinded study scaffolded a frontier model's reasoning with the encyclopedia's abstraction pipeline at inference time and found the scaffold largely inert: fixed pipeline, free planner and enforced coverage scored within noise of each other, and the scaffold did not change the faithfulness of the model's stated reasoning. The problems were self-generated and within the model's reach, in a closed loop where the same model class generated, solved and graded them; the study did not test far transfer, and it did not find the catalog worthless. It concluded that the binding constraint is choosing the right pattern, not enforcing coverage. Early, unpublished observations suggest the approach may help even frontier models on much harder problems; follow-up experiments are pending, so treat that as a lead, not a finding. /conceptual/retrospective-testing-runtime-abstraction-scaffolding/
• Pilots. A single blinded comparison scored a catalog-augmented recommendation above a plain chain-of-thought one; it was one case, meant only to decide whether a proper study was worth running. /applications/augmented_abstract_reasoning_pilot_comparison/
• Inverse innovation. A working paper that starts from an archetype, searches other fields for problems with the same structure, and subjects each proposal to criticism and prior-art research. It demonstrates a search-and-scrutiny method, not autonomous invention; every one of its candidates has adjacent prior art. /applications/inverse-innovation/

All applications and experiments: /applications/. Essays on the framework: /conceptual/

10. How to cite

Cite an entry by its name, its page address and the version shown in its header, and add its catalog ID. For example: Feedback (P-12), Encyclopedia of Abstractions, /primes/feedback/, with the version and date from the page header. Versions are permanent: every past version is kept in the downloadable editions archive, so a quotation can be checked against the text as it stood.

11. Corrections and contact

If you or your user find an error, the author would like to hear about it; contact details are on the About page: /about/