Leonardo imagination graph
Like pinning strings between notes on a wall, except the notes are 313,000 passages and the strings are queryable.
We cut books into passages, extract every useful imagined mechanism, and connect it back to the exact work, author, and chunk it came from.
agentic harness
Like a Renaissance studio: one hand sketches, five critics inspect, then the builders test what survived.
Leonardo writes a dossier. The Council challenges it. The Workshop only builds if the idea survives scrutiny.
Council memory
Like a court record, not a rumor. It remembers who said what, when, with what evidence.
Council deposits, warnings, verdicts, and Workshop results are stored in a separate memory graph so future rounds do not start from zero.
SQLite
The shelves
The dependable cabinet holding document metadata and the original chunk text. When we quote a passage, this is where truth comes from.
Neo4j
The string map
A graph database that lets us ask: which idea appeared in which passage, which work, by which author, near which other ideas?
Chroma
The nose for likeness
Semantic search. It helps find passages that smell similar even when they do not use the same words. It is discovery, not proof.
Bible KG
The fixed witness
A read-only Bible knowledge graph. It is not a myth bucket; it is the alignment reference for authority, personhood, truth, restraint, and temptation.
Hermes / Leonardo
The working hand
The agent that researches the graph, writes dossiers, runs checks, coordinates workflows, and reports back to David.
council-cc
The deliberation table
A Discord daemon that runs the five Council seats as real, separate bot identities with their own roles and memory deposits.
claw-memory
The institutional memory
Our proprietary memory system: a second Neo4j graph (:7688) where Council claims, raw events, and verification records are stored, recalled three ways, and confirmed by independent readback — testimony the harness can trust without mistaking it for evidence.
Next.js site
The public museum
The visible surface: atlas pages, Council pages, dossiers, roadmap, and this whitepaper. The backing data remains queryable behind it.
Chunk store
The original text stays in a boring, reliable database.
SQLite at data/leonardo.db stores document metadata and canonical chunk text. The graph points back to these chunks instead of trusting generated summaries.
Extraction sidecar
The expensive extraction work is tracked separately so old and new batches do not get mixed.
data/extraction_candidates.db records probe candidates, screening results, extraction results, failed parses, and OpenAI batch manifests. This is where coverage is audited before graph writes.
Evidence graph
This is the map: sources, passages, mentions, concepts, and domains.
Neo4j at bolt://localhost:7687 stores Author, Work, Chunk, ConceptMention, Concept, Domain, and ExtractionRun nodes. Relationship types include WROTE, CONTAINS_CHUNK, HAS_MENTION, INSTANCE_OF, CATEGORIZED_AS, and RELATED_TO.
Semantic retrieval
This is the search hound, not the judge.
Chroma embeddings help find neighboring passages and structural cousins that do not share exact words. A Chroma hit can suggest where to look, but the final claim still cites graph and chunk provenance.
Bible KG
A read-only alignment witness lives beside the imagination graph.
Bible nodes are namespaced with Bible* labels and B_* relationships. The graph contains verses, entities, roles, capacities, symbols, valences, lexical roots, geography, and cross-references; Leonardo code must not write to it.
Agentic harness
The graph does not act by itself; named agents move work through stages.
Hermes runs Leonardo as the orchestrator. council-cc is a Bun/TypeScript Discord daemon using discord.js and the Claude Agent SDK; workflow crons and active wait-lock files make sure one Council seat or stage advances at a time instead of drifting.
Council memory
The Council keeps institutional memory in its own database.
claw-memory uses a separate Neo4j at bolt://localhost:7688 for MemoryClaim, RawEvent, VerificationRecord, SummaryMemory, Scope, Seat, and Operator nodes. Memory guides recall; live graph/file checks decide action.
Provenance path
When Leonardo says an idea exists, the important object is not a paragraph summary. It is a path through the graph to the exact passage.
Concept <-[:INSTANCE_OF]- ConceptMention <-[:HAS_MENTION]- Chunk <-[:CONTAINS_CHUNK]- Work <-[:WROTE]- Author
That path is why a dossier can show not only what was extracted, but where it came from, who wrote it, and which neighboring ideas appeared in the same chunk.
Separated trust surfaces
Evidence truth
Leonardo graph + SQLite chunks
If we claim a concept exists, we cite the source chain and quote the passage.
Discovery help
Chroma + web search
Useful for finding candidates and modern analogues; never sufficient by itself.
Alignment witness
Bible KG
Read-only. It disciplines claims around personhood, authority, truth, delegation, and restraint.
Deliberative memory
Council memory graph
Records who said what and why; treated as testimony until verified against live evidence.
Mention-first
Evidence is the exact sentence that imagined something — not a tidy summary written afterwards.
The atom is the ConceptMention: one passage asserting one idea, carrying its normalized name, verbatim excerpt, confidence, lens, prominence, and domains. Concepts are clusters of mentions via INSTANCE_OF, and that clustering edge stays soft — always re-checkable against the mentions beneath it.
Every idea keeps its receipts
Any claim walks back to the work, the author, and the line on the page.
Concept ← INSTANCE_OF ← Mention ← HAS_MENTION ← Chunk ← CONTAINS_CHUNK ← Work ← WROTE ← Author. Nothing floats free of a source, so a dossier shows not just what was extracted but where it came from and which ideas shared the same chunk.
Mapped, not piled
Ideas are filed by domain and pulled together when they co-occur, so cousins across centuries sit side by side.
A 139-domain hierarchical taxonomy (CATEGORIZED_AS) plus RELATED_TO co-occurrence turn one million mentions and 577K concepts into a walkable map. The same imagined mechanism can appear in fiction, myth, and sacred text and still cluster as one neighbourhood.
A fixed witness beside it
A read-only Bible knowledge graph sits next to the imagination graph as an alignment reference.
About 53K Bible* nodes and 420K B_* relationships: verses, entities, capacities, symbols, valences, transformations, lexical roots, geography, and cross-references. It is a catalog of precedents for 'impossible' capabilities and a discipline on claims about authority and personhood — Leonardo code never writes to it.
Two graphs, kept apart
Debate is remembered in its own database so it can never be mistaken for evidence.
claw-memory runs on a separate Neo4j at bolt://localhost:7688, isolated from the evidence graph on :7687. A Council opinion is testimony — it lives here, not in the map of sourced concepts. Memory guides recall; live graph and file checks decide action.
What it remembers
Not just text — who said it, from what event, under whose authority, and whether it was checked.
MemoryClaim (a stated belief), RawEvent (the observed source), VerificationRecord (independent proof), and SummaryMemory (condensed recall), bounded by Scope, Seat, and Operator. Each claim is wired by DERIVED_FROM, OBSERVED_IN, and IN_SCOPE edges and carries a content_sha256.
How it recalls
Three ways of finding a memory, fused — then optionally re-ranked for precision.
Every recall unions lexical (exact and fuzzy text), semantic (vector embeddings), and pattern search. With a model present it reranks the candidates through a local cross-encoder (ms-marco-MiniLM) for the sharpest few. No embeddings degrades to text-only; no rerank model degrades to unreranked — recall never hard-fails.
How it earns trust
A memory is authoritative only after an independent read proves it landed — and nothing is ever silently deleted.
A formal Council deposit counts only as a MemoryClaim with a named seat and operator, a Discord provenance URI, a content hash, the three provenance edges, and a verification done by a fresh Neo4j session — not the tool's own reply. Stale beliefs are superseded, wrong ones marked, legacy rows stratified: recorded, never erased.
One verb-set · eleven tools
Agents reach the memory through eleven MCP tools — the whole lifecycle of a belief, from first deposit to time-travel recall to principled retraction.
313K
source chunks
passages cut from books and documents
1.0M
idea mentions
specific imagined mechanisms found in text
577K
concepts
clusters the mentions are gathered into
53K
Bible KG nodes
read-only reference and alignment witness
328
MemoryClaims
durable Council claims in the memory graph
These are live-scale numbers from the current local system, not decorative counters. They are why the graph is already useful even while canonical cleanup continues.
Graph relationships
Structural health
99.99%
Mentions fully categorized
0
Orphan chunks
0
Missing indexes / constraints
1
Concepts without mentions
Extraction & memory
99.69%
Screened chunks extracted
3.84
Mentions / extracted chunk
274
VerificationRecords
99.7%
Claims with provenance
Fiction, myth, sacred text, and soon papers and patents are ingested into chunks. Each Work and Chunk keeps identifiers, source labels, and enough metadata to reconstruct where a passage came from.
The sidecar pipeline marks which chunks are worth extraction. This prevents paying to extract every dull paragraph and lets us audit unscreened, positive, extracted, and missing-extraction cohorts separately.
Extraction turns a passage into a ConceptMention: true-name binding, resurrection pattern, memory palace, autonomous tool, authority token, and thousands more. Each mention carries normalized name, excerpt, confidence, lens, prominence, and domains.
Resolution clusters many mentions into a Concept through INSTANCE_OF. The edge is soft, like sfumato: useful, but always checked against the underlying mentions because canonicalization is still improving.
Leonardo deepens one canon concept with graph evidence, co-occurring concepts, semantic neighbors, Bible KG alignment, web research, risks, and a prototype question. The output separates what canon already said from what Leonardo added.
Kallimachos checks precedent, Sextus attacks weak evidence, Archimedes asks how to test, Philo checks sacred structure, Humboldt synthesizes across domains. Their deposits are stored as Council memory, not silently blended into the evidence graph.
Workshop builds the smallest measurable experiment. Results go back to Council memory and, when appropriate, back into the main graph as new evidence of success, failure, or limitation.
The Council runs as council-cc — a Bun/TypeScript daemon that gives each seat its own Discord identity, persona, and tools through the Claude Agent SDK, with claw-memory as its substrate. There is no central voice: each seat sees a message on its own and deposits its own MemoryClaim, confirmed by independent readback before it counts.
Kallimachos
The Cartographer
Has this already been judged? What precedent or scope-drift exists?
Sextus Empiricus
The Skeptic
What is the strongest objection, the weakest evidence, the failure mode?
Archimedes
The Engineer
What is the mechanism, and what is the smallest useful experiment?
Philo
The Theologian
Is the sacred / Bible parallel structural, or merely poetic?
Humboldt
The Synthesizer
What do the seats converge on? What inscription should survive?
Workflow 1 → Workflow 2
Workflow 1 is Leonardo's research drain: one canon concept per run becomes a source-backed dossier. Workflow 2 sends that finished dossier through the real Council in deterministic stages — and only PASS verdicts reach the Workshop.
Worked example · CANON-01v2-0001 · true-name power
The first fully worked item ran all twelve stages to ALL_COUNCIL_AUDITS_VERIFIED. The Workshop tested a scoped custody experiment: the cryptographic custody layer gained empirical support, a leaked-key positive-control risk was confirmed in a bounded test, and the semantic “incantatory” layer was explicitly deferred — not claimed.
The ancient shape became a bounded technical hypothesis, not a spell.
The token
$LEO is the economic rail of the Leonardo ecosystem — on Base, with Bankr and x402 as its financial layer.
One token aligns three groups: contributors who grow the graph, builders who run agents on the harness, and the project that funds both. It funds access, bonds authority, rewards verified work, and anchors receipts — a utility token for participating in and using the system, not a claim on profit.
Quests & bounties
Work that grows the project is paid from project supply.
An open quest board posts concrete contributions — graph growth, dossier packaging, concept implementation, refinement and verification — each carrying a bounty drawn from a dedicated project-supply allocation. Quests are scoped, reviewed for quality, and paid on completion.
Hosted agent access
Run your own agent on the harness — metered in $LEO.
A cloud service lets anyone build and run their own agent on the same harness Leonardo uses, with the original Council and Workshop callable as services. Compute and Council/Workshop invocations are metered and settled in $LEO, so usage demand ties directly to the token.
Staking → standing usage
Staking turns a holding into a daily, accumulating usage allowance.
Staked $LEO accrues a daily allowance of hosted-agent usage that accumulates while staked — a stake becomes ongoing access rather than a one-off spend. Staking also opens and weights participation on the quest board.
The flywheel
Use funds work; work grows the graph; a better graph draws more use.
Hosted-agent usage and quest activity create demand; bounties paid from supply pull contributors who grow the graph and harden the harness; a deeper graph and stronger agents make the hosted product more valuable — which draws more usage. The token is the loop's accounting layer.
Supply & transparency
A dedicated allocation funds quests; the mechanics are stated, not hidden.
A portion of supply is earmarked for the quest and bounty program; the remainder follows the Bankr launch on Base. Allocation, emissions, and staking parameters will be published before they go live.
Hermes decides — what & who
The harness owns the judgment. Leonardo researches, the Council deliberates, claw-memory records, and the Workshop executes; quest scoring verifies that a contributor actually grew the graph, packaged the dossier, or passed Council review before any payout is signalled.
Bankr executes — the money move
Bankr is the financial layer on Base. When a quest passes, Hermes signals a payout and Bankr releases the bounty from project supply; it also runs staking distributions, settles hosted-agent usage in $LEO, and handles treasury operations and contributor notifications.
Hermes never has to be a payments system; Bankr never has to understand the graph. The token is the wire between them.
Offchain · the harness
Onchain · Base
The capabilities themselves do not run on a blockchain. The chain records who paid, what version was called, what was verified, and who earned. Offchain execution, onchain trust and payment.
Identity Registry
CANON-01v2-0001 · true-name
An ERC-721-style agent ID, an agentURI, and a verified wallet. The standard holds the address; Leonardo decides what evidence fills it.
Reputation Registry
CANON-01v2-0002 · impersonation
Tagged feedback and offchain feedback hashes with revocation — reputation backed by the Workshop record, not by who paid most.
Validation Registry
CANON-01v2-0003 · authentication
An agent asks a validator to verify work; the Council and Workshop post the response and an evidence hash.
ERC-8004 — the draft Trustless Agents standard — is the address book. Leonardo is the trust engine. $LEO is the economic rail. We register into the standard instead of reinventing it.
The trust stack, as tools
The seven Agent Trust Stack primitives are not only Leonardo's own faculties. As each clears its Council and Workshop gates, it ships as a callable tool other builders embed in their own agents — metered in $LEO. Identity is the flagship: Agent Passport-as-a-Service. $LEO pays for the registration, bond, and validation ceremony; it never buys the trust itself, which still depends on evidence, validation, and history.
$LEO can buy
$LEO cannot buy
Truth is evidence-governed, not token-governed.
Planned utility, described for transparency. $LEO is a utility token for participating in and using the Leonardo ecosystem — not an investment — and specifics may evolve before launch. See the token page for the summary.
Books, papers, abandoned patents: the next corpus wave turns imagination, science, and neglected filings into one searchable design field.
Wave-3 batches reconciled; 99.98% extraction coverage across screening-positive chunks. 145,610 distinct wave-3 chunks extracted via the mid-2026 batch run.
Extend Neo4j with Paper, Citation, Inventor, Patent, and TechnologyArea surfaces so fiction, papers, and patents can be compared in one prior-art map.
Tier-1 public-domain fiction, hard-SF mid-century, philosophy of mind, and mystical/esoteric expansion. Reuses the existing ingest pipeline.
AI, language, neuro, crypto, quantum, applied physics, bio, and optimization categories from 2010–2026. Technical papers become second witnesses beside imagination.
USPTO public data filtered toward speculative CPC codes, lab assignees, and abandoned/lapsed/expired status. The quarry becomes prior art and neglected mechanism.
Activate PatentExtractor, add ArxivExtractor, and expose seed-papers / seed-patents CLI entrypoints.
Technical screening and extraction prompts for papers and patents, with single-lens extraction and fewer fiction-specific assumptions.
Books, patents, and papers run as separate waves using source-kind filters so old rows do not mix with new cohorts.
Resolve concepts across fiction, papers, and patents so an imagined mechanism can be matched to a technical analogue or abandoned implementation.
Materialize REALIZED_BY, ANTICIPATED_BY, and IS_PRIOR_ART_FOR edges between concepts, papers, and patents.
Graph health, sample-trace integrity, cross-link sanity, performance checks, and agent CLI hooks for paper, patent, and cross-source stats.
All of the graph work feeds this: a self-improving agentic harness that turns imagined capabilities into tested, running ones.
Leonardo research → Council deliberation → orchestrator synthesis → Workshop testing → empirical result back to memory and the graph. The loop is specified; Phase 01 calibration is running live.
A second Neo4j on :7688 stores deliberative memory — deposits, verdicts, prior rounds, and Workshop result reports. Memory is testimony, verified against live evidence before action.
Turn PASS verdicts and test designs into real implementations: agent capabilities, code modules, prompt templates, workflow steps. Every result returns to Council memory as evidence.
Every tested capability becomes a reusable skill or harness primitive the agents can invoke.
Workshop successes and failures feed back into the graph as new evidence. Each round patches the harness docs, queries, and templates.
The orchestrator coordinates the Council, Leonardo, and Workshop agents as a fleet rather than one opaque assistant.
Once Phase 01 clears, the harness advances to later canon phases without human seeding: twenty-one phases, self-driven but still verified.
$LEO ships on Base, opens the quest board, meters the hosted harness, and settles into a usage-funds-work economy.
$LEO goes live on Base with Bankr handling the financial rails — token, treasury, and payout automations — and a project-supply allocation earmarked for the quest and bounty program.
The first open quests — grow the graph, package dossiers, implement mined concepts, refine and verify — posted with scoped bounties paid from project supply on completion.
Quest submissions are reviewed for quality, with the five-seat Council able to adjudicate disputed or high-value work before bounties release.
A cloud service to build and run your own agent on the harness, with the Council and Workshop callable as services, metered and settled in $LEO.
Staking goes live: staked $LEO accrues a daily, accumulating hosted-agent usage allowance and opens quest-board standing.
General availability of the hosted harness and a steady-state quest economy where usage funds the work that keeps growing the graph.
Mentions are evidence; concepts are clusters.
Every serious claim cites Concept → Mention → Chunk → Work → Author.
Bible KG is read-only and treated as alignment witness, not a toy mythology source.
Council memory is testimony; live graph/file checks decide action.
Council before Workshop: critics first, builders later.
No graph resets, no secret exposure, no paid batch work without value stated.
the purpose
Every invention was first imagined somewhere. Leonardo makes those old imaginings searchable, criticizable, and testable — then lets memory keep the scars from every round.