Source
Recover what is actually present before abstraction begins.
Thousands of human artifacts in one analytical space — decomposed into evidence, mechanisms, constraints, and candidate generators.
What rules would have to exist for this artifact to be the output?
Stories, terminology, examples, symbols, arguments. Useful signals — but not yet the machine that produced them.
Cangjie treats the text as a constructed object. The goal is not to summarize the artifact, but to isolate the mechanisms that recur beneath its changing surface forms.
The artifact enters the pipeline.
Recover what is actually present before abstraction begins.
Recover dependencies, transformations, invariants, and recurring relations.
Strip away domain-specific clothing until the reusable structure remains.
Find the generator-level abstraction capable of producing multiple forms.
A second layer appears when recovered structures are compared: shared generators, dependency chains, equivalences, contradictions, and families of forms.
LIDE expands the artifact into a defensible information space. Each level represents what can be established at that epistemic depth.
Directly present in the source artifact.
Individual structural element recovered from source.
Separate element, seemingly unrelated.
Third element completing the pattern.
Candidate structures tested for recurrence, transferability, fidelity, exclusivity, formal coherence, and compression.
One causal pathway through the artifact.
An alternative explanation for the same pattern.
A third candidate, not yet distinguished.
What survives each operation? Beginning dense, ending irreducible. Each state shows what remains after one more compression.
Cangjie compresses the recovered structure toward its smallest reusable generator.
This is one of the best demonstrations of what the Library does. The system isn't looking for what the story means. It is asking: what mechanism is the story implementing?
The mechanism holds. The vocabulary does not survive — the structure does.
Order is produced by resolving competing claims over undefined potential; this potential is partitioned by Conflict, selected through Capability Benchmarking, and stabilized through Hierarchical Delegation and Audited Constraints.
This is a model of a proposed validation procedure. It is not a proof engine.
Rich system with domain-specific vocabulary and narrative structure.
Reconstructed from abstracted rules and constraints.
Each artifact is passed through the same pipeline: LIDE decompression → cross-discovery → verification → Cangjie compression → generator → regeneration test.
The Library is not only a place to read the artifacts. It is a place to examine what remains when their vocabulary is removed.
1,309 volumes — scripture, myth, grimoire, alchemy, history. Spines render directly from corpus data.
What produces it?
What constrains it?
What changes?
What survives?
What reproduces it?
Does it survive domains?
Can it make the form again?
The instrument has done its work. What remains is reading — 1,309 volumes, each carrying the structures recovered here.
Browse the archive