m590.research.x1000.ai

LIGHT HOPE · Research Infrastructure Node

Compile research once.
Keep the knowledge fresh.

m590 is a research knowledge compiler. It takes raw sources — papers, repositories, articles, transcripts — and continuously turns them into structured, interlinked, auditable knowledge. The LLM is the librarian, not the search engine.

Request access See how it works
v0.1.0 · skeleton Open architecture, inspectable design — built in the open, claimed only when real.
01

Ingest

Papers, repos, web, transcripts enter as immutable sources.

02

Normalize

Every source type maps to one standard intermediate format.

03

Compile

An agent distills sources into linked knowledge — swappable runtime.

04

Structure

Topic → Claim → Evidence → Source, with typed relations.

05

Render

An Obsidian-compatible wiki with backlinks and a knowledge graph.

06

Export

Quality-gated knowledge objects, portable as m495 skills.

The idea

Traditional retrieval rediscovers everything on every query. Compilation accumulates.

Following the Karpathy LLM Wiki pattern: instead of re-reading raw fragments each time a question is asked, m590 compiles knowledge once into a persistent, linked structure and keeps it updated. Knowledge compounds instead of evaporating.

🗂️

Structured, not scattered

Sources become a graph of topics, claims, and evidence — each assertion traceable to where it came from.

🔗

Interlinked by design

Bidirectional backlinks and typed relations (is_a, extends, contradicts) turn isolated notes into navigable knowledge.

🧾

Auditable provenance

Every claim requires evidence. Conflicts are annotated, not silently overwritten. Every change writes a revision record.

Architecture

Six layers, every dependency swappable

Each external tool sits behind an abstract adapter. Swap the agent runtime, swap the wiki renderer, swap the model provider — without touching business logic. Anti-lock-in is the design, not an afterthought.

Layer
Name
Responsibility
Swappable via
L0
Policy
Compile rules, taxonomy, quality gates
config file
L1
Source
Immutable raw materials — never modified by agents
storage adapter
L2
Normalization
Standardized intermediate format for all source types
transform adapter
L3
Compile
Agent-driven knowledge compilation
agent adapter
L4
Knowledge
Structured objects: Topic, Claim, Evidence, Relation, Revision
store adapter
L5
Export
Obsidian vault, JSON, m495 skills, API
export adapter

Knowledge model · MCP v1

Objects that hold their evidence

The knowledge protocol is small and strict on purpose. Immutable sources at the bottom, revisions on every change, evidence behind every claim.

  • SourceImmutable raw material — paper, repo, web page. Agents cannot modify it.
  • TopicA knowledge node at taxonomy level L0–L3.
  • ClaimA specific assertion within a topic, backed by evidence.
  • EvidenceLinks a claim to its supporting source.
  • RelationTyped edge between topics — is_a, extends, contradicts.
  • RevisionVersion record written on every knowledge change.

Trust & taxonomy

Sources are ranked, not flattened

# taxonomy levels L0 Domain AI, Biology, Economics L1 Sub-domain LLM, Genomics, Macro L2 Method Transformer, CRISPR, MMT L3 Detail RoPE, Cas9, Taylor Rule # source trust levels A "peer-reviewed / official docs" B "reputable articles / known experts" C "community — forums, discussion" D "unverified provenance" # quality gate (compile) min_evidence_per_claim: 1 min_confidence: 0.6 max_orphan_ratio: 0.20

Who it is for

Built for teams that compile knowledge across domains

One compiler, many domains. Behavior is driven by a compile-policy file — not hardcoded prompts — so the same infrastructure serves very different subject matter.

Research & AI teams

General research compilation

Turn a stream of papers, repos, and articles into a maintained wiki that new questions can query instead of re-deriving.

Game & product research

Design-knowledge packs

A dedicated game-research compile policy supports design knowledge for projects like m595 COMPASS and future titles.

Specialized domains

Domain-tuned policies

Configurable policy packs exist for higher-rigor domains. Domain packs set stricter citation and trust rules per subject.

Note: m590 is knowledge-compilation infrastructure. Compiled material is research knowledge, not professional, medical, or legal advice — trust levels and evidence links are provided so downstream readers can verify sources themselves.

Where it stands · m8300 proof-before-claims

Honest about what is built and what is next

m590 is at v0.1.0 — project skeleton. We show the roadmap plainly rather than dressing up plans as shipped features. The list below mirrors the project's own development roadmap.

Built today

v0.1.0 · skeleton

  • ✓ FastAPI service, V5.3/V5.4 compliant, port 8590
  • ✓ Core data models: Source, Topic, Claim, Evidence, Relation
  • ✓ Abstract adapter interfaces (anti-lock-in)
  • ✓ API surface: research, vault, export endpoints
  • ✓ Docker Compose with resource limits
  • ✓ Configurable compile policy + domain policy packs
  • ✓ MCP v1 knowledge protocol defined

On the roadmap

v0.2.0 → v1.0.0

  • ○ End-to-end compile loop: ingest → compile → wiki
  • ○ PostgreSQL knowledge store + Obsidian backend
  • ○ Quality gates enforced, conflict annotation
  • ○ Real m495 skill export integration
  • ○ Multi-project scope isolation
  • ○ Nightly scheduled compile runs (m8005)
  • ○ m8101 knowledge-center integration, audit trail

Access

Want to compile a corpus with m590?

m590 is an internal LIGHT HOPE infrastructure node, opening to early collaborators. There is no public pricing yet — reach out to discuss a compile scope, or join the waitlist for the first compile-loop release.

Contact the team Join the waitlist