A source-grounded knowledge engine for understanding people, places, events, and ideas over time.
Information is abundant. Understanding is rare.
World turns documents, events, records, research, and conversations into structured knowledge that keeps its evidence attached. Every conclusion stays connected to where it came from, when it was learned, what supports it, and how confident the system should be.
World is not another chatbot. It is an attempt to build an evidence-backed understanding that can grow, evolve, and explain itself — while preserving provenance, uncertainty, temporal context, freshness, corrections, and human judgment.
World doesn't try to answer everything. It builds an evidence-backed understanding that can grow, evolve, and explain itself.
- Status — a portfolio showcase
- Why World Exists
- The Problem
- The World Approach
- Inside the Engine
- Architecture
- Core Capabilities
- Provenance and Trust
- Time-Aware Knowledge
- Honest Uncertainty
- Concept Diagrams
- Engineering Highlights
- Privacy and Local-First Philosophy
- Repository Guide
- Roadmap
- Part of the NemoClaw Ecosystem
- Author and Portfolio Context
- License
This is a public, recruiter-facing showcase repository. It presents the concept, architecture, knowledge model, and engineering philosophy of World. It is not an open-source production release.
To keep the presentation honest, everything here is framed against four explicit tiers, and the same labels are used throughout the documentation:
| Tier | Meaning |
|---|---|
| 🟢 Implemented foundation | Established design decisions and working conceptual foundations that the model is built on. |
| 🔵 Architectural direction | The intended structure of the system — described, not claimed as fully built. |
| 🟣 Product concept | Illustrative experiences that show where the engine leads. |
| ⚪ Future roadmap | Planned and research directions that are not yet built. |
There are no invented metrics in this repository — no test counts, benchmarks, customers, deployments, or production usage. Diagrams and any names shown are illustrative. The production implementation, its private history, and any credentials are intentionally kept in separate private repositories and are not included here. See the License and SECURITY.md.
We are surrounded by information and starved of understanding.
Search returns links. Chatbots return fluent text. Neither preserves why an answer should be believed, when it was true, how fresh it is, or what remains unknown. When the answer is wrong — or simply out of date — there is often no way to trace it back and correct it.
World starts from a different premise: knowledge is only trustworthy when it can explain itself. Understanding should accumulate like a well-kept record, not evaporate like a chat session.
A useful understanding of the world has to hold several hard things at once:
- Provenance — where each fact came from, and who or what asserted it.
- Time — when a fact was true versus when it was learned.
- Freshness — whether the supporting evidence is current or stale.
- Uncertainty — how confident to be, without hiding doubt behind fluent prose.
- Correction — the ability to revise, supersede, and keep a history.
- Human judgment — a place for people to confirm, override, and be accountable.
Most systems optimize for a confident-sounding answer and quietly discard all six. World is an argument that these are not optional extras — they are the knowledge.
World treats every piece of information as evidence first. Raw sources become immutable, source-linked evidence records. Evidence supports claims that carry confidence. Claims connect entities — people, places, organizations, events, concepts — into a knowledge graph layered with time and provenance. Answers are then assembled from that graph, so they can always be traced back.
Language models help read and interpret along the way, but they are not the authority. The system of record is the deterministic, inspectable evidence-and-claim layer that a human can check. See KNOWLEDGE_ENGINE.md.
One observation is carried all the way through — from an untrusted paragraph, to a grounded evidence record, to a claim with explicit confidence, to a durable and connected node in the graph. Nothing loses its lineage on the way.
The core flow is deliberately linear and inspectable:
Sources
↓
Ingestion
↓
Parsing and Understanding
↓
Evidence Records
↓
Claims
↓
Knowledge Graph
↓
Temporal and Provenance Layers
↓
Query and Reasoning
↓
Applications
Governance, privacy, human review, and stewardship are cross-cutting — they apply across every stage rather than sitting at the end as decoration. A full walk-through of the layers, trust boundaries, and data flow is in ARCHITECTURE.md.
World organizes knowledge around a small, durable set of entity types:
people · places · organizations · events · documents · claims · evidence · concepts · relationships · timelines · sources · corrections
Around that model, the engine is designed to:
- 🔵 Convert heterogeneous sources into immutable, source-linked evidence records.
- 🔵 Derive claims that carry confidence and cite their evidence.
- 🔵 Maintain a knowledge graph of entities and relationships.
- 🔵 Track four kinds of time and a freshness signal on every claim.
- 🔵 Record version history, corrections, and supersessions.
- 🟢 Prefer deterministic, inspectable behavior over opaque generation.
- 🟢 Keep data local by default, with human review as a first-class step.
Capability tiers use the labels from Status; see ROADMAP.md for what is foundation versus planned.
A meaningful piece of knowledge should be able to answer:
- Where did this come from?
- Who or what asserted it?
- When was it learned?
- What time period does it describe?
- What evidence supports it?
- How confident is the system?
- How fresh is the evidence?
- Has it been corrected or superseded?
- What remains unknown?
Provenance in World is not a citation footnote — it is the spine of the data model. Details are in PROVENANCE.md.
Facts have a lifespan. A role, a location, a relationship — each is true for a period, and World was learned about it at a moment that is usually later. World keeps valid time (when it was true in the world), observed time (when World learned it), ingestion time, and freshness (when it was last confirmed) distinct, so answers can respect when they are being asked about.
World is designed to say "I'm not sure," "this may be stale," and "this is unknown" — plainly. Confidence is an explicit property of a claim, not a tone of voice. When evidence conflicts, the conflict is preserved rather than averaged away. A quiet, well-marked unknown is treated as more valuable than a confident guess.
Product concept. Entity names shown are illustrative, not real records.
The graph is the place where evidence and claims become a connected understanding. Each node links back to the claims and evidence that justify it, so the picture can always be unfolded into its reasons.
- Evidence-first data model — immutable, source-linked records as the system of record.
- Provenance as a first-class citizen — source identity, origin, and locator on every record.
- Bitemporal thinking — valid time and observed time modeled separately.
- Freshness and staleness — knowledge knows how old its evidence is.
- Explicit confidence — uncertainty is data, not phrasing.
- Correction and version history — nothing is silently overwritten; supersession is tracked.
- Deterministic, inspectable pipeline — models assist interpretation; they are not the authority.
- Local-first storage philosophy — designed around durable, embeddable local storage.
- Human-in-the-loop by design — review and override are architectural, not afterthoughts.
- Long-horizon stewardship — knowledge is meant to be maintained, refreshed, and corrected over time.
See ENGINEERING_PHILOSOPHY.md for the principles behind these choices.
World is designed to run on your own machine, with your own data, under your own control. Local-first is not just a deployment choice here — it is a trust choice:
- Data stays local by default; there is no silent exfiltration to remote services.
- Understanding is portable, inspectable, and yours to export or erase.
- Human review sits between automated interpretation and durable knowledge.
No production secrets, personal datasets, credentials, or private source belong in this public repository. See SECURITY.md.
| Document | What it covers |
|---|---|
| README.md | This overview. |
| PROJECT_SUMMARY.md | A recruiter-friendly, few-minute walk-through. |
| ENGINEERING_PHILOSOPHY.md | The principles behind the build. |
| ARCHITECTURE.md | Layers, data flow, trust boundaries, and models. |
| KNOWLEDGE_ENGINE.md | How information becomes durable knowledge. |
| PROVENANCE.md | The provenance and trust model in depth. |
| ROADMAP.md | Foundation, in development, planned, and research. |
| RELEASE_NOTES.md | The v1.0.0 showcase release. |
| CONTRIBUTING.md | How this showcase handles contributions. |
| SECURITY.md | Responsible disclosure and what never belongs here. |
| LICENSE | Viewable showcase terms. |
| assets/ | All diagrams, authored as self-contained SVG. |
World is presented in honest phases rather than as a finished product:
- 🟢 Foundation — the evidence-first model, provenance and temporal thinking, and local-first, deterministic design.
- 🔵 In development — the claim and knowledge-graph layers and their query surface.
- ⚪ Planned — richer reasoning, correction workflows, and application interfaces.
- 🔬 Research direction — long-horizon stewardship and trustworthy assisted interpretation.
The full breakdown is in ROADMAP.md.
NemoClaw is a portfolio of trustworthy software systems built around evidence, provenance, consent, and human authority.
- Skippy — human-centered personal AI
- Polis — governed civic intelligence
- World — source-grounded knowledge infrastructure · you are here
- Crate — local-first digital stewardship
Each project keeps its own identity and stack. What they share is the engineering posture, not a runtime.
Within that ecosystem, World is intended to serve as the knowledge and context layer — the place where trustworthy, source-grounded understanding is built and maintained. World providing that layer to its sibling projects is an intended ecosystem role, not a completed integration.
I conceived World, defined its evidence-first and provenance-first principles, and directed its architecture and knowledge model. This repository is a portfolio showcase of that systems thinking: how to build knowledge infrastructure that is traceable, time-aware, uncertainty-honest, local-first, and stewarded by people over the long term.
It demonstrates systems architecture, data modeling, provenance and temporal design, and technical communication — with a deliberate emphasis on trustworthy AI infrastructure over chatbot novelty.
Released under the World Public Showcase License — free to view and reference for evaluation, but not to copy, redistribute, use commercially, or present as your own. Production source code is not included. See LICENSE.
World — build an understanding that can explain itself.