ideonomy

Source-available under Harvest Commercial 1.0. Commercial use of new protected contributions requires a separate paid license; no donation or self-assessed zero substitutes for an agreement. Prior MIT grants and third-party rights remain intact. Commercial licensing: contact@exopriors.com, attention Xyra Sinclair.

Patrick Gunkel's ideonomy — "the pure and applied science of ideas and their laws, and of the use of same to describe, generate, investigate, or otherwise treat all possible ideas related to any subject, problem, thing, or other idea" — rebuilt as inference-time machinery.

Gunkel (1947–2017, MIT) pursued it by hand and with 1980s combinatorics programs: hundreds of systematic lists, 236 named divisions of the field, cross-products yielding tens of thousands of candidate ideas (230 elementary shapes × 74 types of order = 17,020 "shapes of order"), each a question someone could take seriously. The method's bottleneck was always evaluation — a human cannot read 84,496 two-word psychological states. Models can. Ideonomy is an LLM-era project that arrived forty years early.

Where the information lives

A model's default generations sample the mode of its training distribution: ask for ideas and you get the densest neighborhood, fluently. Ideonomic work is the deliberate traversal of the rest of the space, and everything in this repo stores its information in one of five places:

• The denominator. A list is a bounded claim about a region of idea-space: which universe it enumerates and how much of that universe it has actually covered. Coverage is labeled — covered, named-gap, ruled-out — never implied.
• The partition. An induced typology is a structural claim, tested by whether the types exclude each other and whether naming a neglected type yields real new members. The neglected types are the gradient the growth climbs.
• The order. A seriation axis or relational map asserts geometry the underlying embedding does not certify; axes and verdicts stay labeled as interpretation, and canon order is never overwritten.
• The gate. Generation is cheap and discrimination scarce, so value concentrates in what was refused and why. Every drop is recorded as residue with its reason, and residue seeds the next pass.
• The ratchet. Structure is kept only while it pays for itself in description length; a cycle that does not hold the compression ratio is reverted, so the corpus cannot bloat its way to the appearance of progress.

Gunkel could work the first three by hand. The last two are what models make affordable, and they are where this repo differs from its siblings: gates and accumulation.

Before and after

Ask an agent whether a product's landing copy is in the right register.

Default: "It's solid, 8/10."

With triangulate-without-oracle loaded: the agent names that no computable answer exists, splits the judgment into axes — austerity, exactness, structural clarity — and gets two independent reads per axis. The reads agree on two axes and disagree on whether one line overclaims. That disagreement is the actual review finding, and it goes to the human who owns the call, with the grounds named. No fabricated number.

What's here

Three layers, each usable alone:

1. skills/ — agent skills. Each is one page, self-contained, with its trigger in the frontmatter and its gate in the body. Start with practice-deep-ideonomy for list-making that discovers its own grammar, or the oracle trio — triangulate-without-oracle, build-the-oracle-before-the-answer, audit-the-oracle-coverage — for judgment with and without a computable check. The original three-persona distillation and its independent adversarial reviews are in docs/premier-skills.md.

2. ORGANON.md — 37 inference-time cognitive primitives in one tension-metabolizing loop (SENSE → ORIENT → GENERATE → JUDGE → ACT → PERSIST), synthesized from three streams that converge on it: Gunkel's ideonomy, automatic taxonomy induction, and automatic codebase improvement. CYCLES.md adds the orthogonal expansion↔compression axis, ratcheted on minimum description length, that turns the catalog into an engine.

3. ideonomy/ — the engines and Gunkel's corpus as data. Stdlib-only Python 3.9+, zero dependencies, fully offline-testable. Engines: cycles.py (MDL-ratcheted expansion/compression), triangulate.py (independent judgments per axis when no oracle exists), trial.py (advocate vs adversary, swap-balanced, independent bench), parley.py (multi-party constraint solving, maximin at impasse), residue.py (cross-session ledger), lists.py (typed, provenanced list algebra), seriate.py (spectral ordering with an explicit smoothness objective). Corpus: Gunkel's divisions (divisions.py) and generative operators (operators.py) machine-usable, and his lists recovered verbatim with per-URL provenance in the canon layer (canon.py + data/). Any CLI or callable is a model (models.py).

Quick start

git clone https://github.com/XyraSinclair/ideonomy && cd ideonomy
python3 -m ideonomy.cycles_demo     # the MDL engine over its own catalog — offline
python3 -m unittest discover tests  # the whole suite, no network, no deps

The demo prints a cycle log: grp is how many groups the compression found, codelen the cost of writing that structure down plus encoding the corpus through it, and raw the cost of the corpus with no structure at all. ratio is raw over codelen — above 1.0 the structure pays for itself, and a cycle is kept only if the ratio holds or rises — while resid counts the items the structure fails to explain, which seed the next cycle.

Install the skills

Claude Code, one command:

/plugin marketplace add XyraSinclair/ideonomy

(then install the ideonomy plugin; skills appear as /ideonomy:<skill-name>.)

Any other agent that reads SKILL.md files:

./install.sh                        # copies skills/ into ~/.claude/skills/

Or load any single skill by putting its SKILL.md in context — each is one page, self-contained (the P…/M… keys inside are provenance pointers into ORGANON.md, not prerequisites). Start with the top three: triangulate-without-oracle, build-the-oracle-before-the-answer, reframe-until-it-dissolves — or install route-to-the-right-move and let it dispatch.

Use the library

from ideonomy import operators, divisions, primitives, cycles

# Ideocombinatorics (P12): cross two lists, read the product for live cells.
qs = operators.combine(
    ["recurrence", "symmetry", "cascade"],
    ["grief", "negotiation", "metabolism"],
    template="Can there be {a} of {b}?",
)

# Gunkel's divisions (236 recovered), each a fault-model for thought.
divisions.DIVISIONS["ANALOGIES"]          # -> 'Icelology'
print(divisions.lens_prompt("ANOMALIES", "the git commit graph"))

# The organon, machine-readable.
[p.key for p in primitives.PRIMITIVES if p.phase == "JUDGE"]

# The respiratory engine: expand -> judge -> compress over any corpus of text.
state = cycles.seed(["symmetry of grief", "cascade of negotiation"])
cycles.run(state, cycles=5)               # MDL-ratcheted

Model-backed, with any CLIs you have (heterogeneous panels are the point):

python3 -m ideonomy.demo --model 'claude -p {prompt}' --model 'ollama run llama3.3'
python3 -m ideonomy.triangulate "Is this landing copy in the right register?" \
    --axis austerity --axis exactness \
    --judge 'claude -p {prompt}' --judge 'codex exec {prompt}'
python3 -m ideonomy.trial "this API should be deprecated" \
    --advocate 'claude -p {prompt}' --adversary 'codex exec {prompt}' \
    --judge 'ollama run llama3.3'                 # adversarial trial, swap-balanced
python3 -m ideonomy.parley "name the release" \
    --party ops='claude -p {prompt}' --party brand='codex exec {prompt}' \
    --constraint 'ops:must be greppable' --constraint 'brand:must not be generic'
python3 -m ideonomy.residue --topic mywork open   # cross-session residue ledger
python3 -m ideonomy.draw "your problem" --n 3     # forced non-default lenses, offline
python3 -m ideonomy.registers "the launch post" --n 2   # forced register mixes, offline
python3 -m ideonomy.lists ls                      # the cross-chat list store

(The residue ledger stores its state in ./.residue/<topic>.json, created on first use; override with --store.)

The database

Browse the published catalog atlas or its offline copy to search names, types, and items, then read a list in its stored or canon-sidecar order. Coverage counts are derived from the data, and a stored projection is shown only while its item fingerprints match — growing a list hides its stale point rather than pretending the old map still measures it. Rebuild without network access: python3 -m ideonomy.atlas.

Two provenance tiers, never confused (ideonomy/data/, load via python3 -m ideonomy.canon ls and ... --tier grown ls):

• canon — Gunkel's own lists, recovered verbatim with per-source provenance: the archived pre-redesign ideonomy.mit.edu text (canon-wayback.jsonl), all 403 of his photographed charts transcribed by vision models with legibility labels (canon-charts.jsonl), and monograph pages stitched from page scans (canon-monographs.jsonl). Canon text is never edited; even its seriation orders live in a sidecar (data/seriations.jsonl).
• grown — the machine-extended edge, produced by the hill-climb (corpus/climb.py): grow → induce the list's own typology → name the types it neglects → gap-fill → gate every candidate for genuine-category, distinctness, and combinatorial phrasing → ratchet, with drops recorded as residue and a plateau flagged when the keep-rate falls. Grown lists carry source.tier == "grown" and can never masquerade as canon.

Grown-tier conventions all ride in source with zero schema change: registers (each list written in one named voice, a deliberate lean rather than an accident), maps (source.kind == "map", relations with exact item endpoints so reordering cannot silently change them), openings (source.priorities names a near-term lead and a wild branch — attention choices, not confidence scores), seriation (model-direct or spectral via corpus/seriate_drive.py; the axis, smoothness, and seriability score are stored in source.seriation, and smoothness alone is not treated as proof of a one-dimensional spectrum), and boundary claims (a discovered edge of the universe is recorded and the excluded universe gets a sibling list). The operative quality bar is itself a grown list — list-excellences — and every authored breath self-gates against it, recording residue.

The per-list ledger (residue, keep-rates, by: run labels) is in corpus/climb-ledger/. The growth, widening, and seriation drivers in corpus/ require GEMINI_API_KEY and make billable model calls; the atlas and everything in ideonomy/ need neither. Dated fieldwork records — specimen admissions, a Gemini 3.8 Flash sketching experiment, growth checkpoints, each with its repairs and limits — are in docs/fieldnotes.md.

This is Gunkel's progressive loop (list → induce types → find missing items → refine) made executable, with the evaluation gate he lacked. The catalog applies to itself: the same gate-and-typology discipline the skills prescribe is what grows the database.

Status

The catalog, skills, and offline engines are substantive and tested. Committed empirical evidence: docs/breath-log-external.txt — the MDL engine run on two real external corpora, finding named structure where it exists (ratio 1.110) and reporting near-null where it doesn't (1.011) — and docs/heterogeneous-trial-r7.txt — a real three-vendor trial (Claude vs Gemini argued, sides swapped for balance, local Qwen judged) whose swap-stable verdict went against the maintainer's own publication decision and is committed unedited: the trial engine does not flatter its owner. Still undemonstrated: compression depth beyond token-level MDL (semantic grouping with model-backed judges). The full labeled denominator of what "canonical" means here — covered, named-gap, ruled-out, nothing unlabeled — is docs/canonicality.md.

Lineage and related work

Everything here descends from Patrick Gunkel (the primary source: ideonomy.mit.edu, five scanned volumes and ~400 charts) via Grace Kind's revival essays (intro), which named the field's three open problems: evaluating idea quality, contextualizing ideation strategies, and effective AI-driven ideation. This repo's center of gravity is the first two — gates and routing — which is also what distinguishes it from its siblings:

• latentwill/ideonomy-skill — a well-made pair of Claude skills built on Kind's essays: an external random picker (8 operators × 17 organons × 29 dimension-prompts) against ideation mode-collapse. Take its thesis seriously — draw.py here is that idea pointed at Gunkel's full catalog. What it deliberately lacks is what this repo is for: evaluation gates and cross-session accumulation.
• Morpheis/ideonomy-engine — a TypeScript CLI curating 28 of Gunkel's divisions into agent-usable lenses with chaining/synthesis verbs. Good curation; the composition is string templates, and there is no judgment layer.
• kindgracekind/ideonomy-legacy — Kind's own 2023 experiments, including the first stab at the idea-quality problem (discriminator.py).

Distilled research notes with full provenance: docs/gunkel.md.

License and compensation

Harvest Commercial 1.0 is in force for new protected contributions. Personal non-commercial study, qualifying research and teaching, independent public-interest scrutiny, and narrowly necessary emergency use have a no-fee grant. Commercial use requires a separately executed license with payment terms agreed before use. Those terms may use fixed fees, minimum guarantees, milestones, royalties, or value-sharing grounded in an agreed Shapley model. Attribution, a donation, or an honest zero is not commercial permission.

This is source-available, not OSI-approved open source. No license can create copyright over ideas, guarantee collection, or turn an undefined contribution score into a debt. The paid agreement defines the amount, scope, reporting, due dates, and dispute process.

The earlier MIT grant remains available for material already published through commit 52d5ca477c23ab160318a21ab1e4f2a5b030af38; it is not revoked or extended to later protected contributions. Private preparation history is not part of the public release.

Gunkel's recovered material is third-party work: source URLs establish provenance, not permission to relicense it. No new grant from its rightsholders was established in preparing this release; Harvest does not supply one.