Sovereign-Lens

A long-horizon evaluator that preserves geopolitical AI commitments, revisits them as the world changes, and learns without rewriting history.

Web: sovereignlens.ai

Current recommendations: Next steps audit — 2026-08-23

Motto: The Horizon already exists. Sovereign Lens makes its programming observable.

I approached Long Horizon Agents Build Day from a different direction. Most teams are extending an agent's memory - giving it a longer past. My premise is that the Horizon already exists, and it is already being programmed.

The Horizon is a shared, forward-facing object shaped by capital, talent, compute, energy, infrastructure, culture, data, standards, institutions, human networks, and AI agents. States are already writing to it through investments, treaties, export permissions, infrastructure, education, regulation, and technical standards.

My contribution today is Programmable Horizons: an approach for representing these changes as observable, versioned write events.

The instrument is Sovereign Lens—an open-source, evidence-first system for observing how those writes change sovereign capability, dependency, control, and optionality.

To reduce dependence on any single worldview, the same evidence is examined through multiple model perspectives. Agreement is useful; disagreement is preserved as analytical parallax. The system can later compare each interpretation with what actually happened.

SovereignLens.ai makes the programmable future observable.

Demo: sovereignlens.ai
Code: github.com/Dim25/Sovereign-Lens

Thanks!

-Dima (Let's Connect: https://www.linkedin.com/in/dim25

2 min tour: https://sovereignlens.ai/demo2min

Prototype status and implementation boundaries

This README was produced rapidly with AI assistance during a hackathon. It intentionally mixes the working prototype with the intended architecture and has not yet received full human editing or methodological review. See What actually works today for the honest boundary between software and specification.

Sovereign Lens

A long-horizon evaluator that preserves geopolitical AI commitments, revisits them as the world changes, and learns without rewriting history.

Sovereign Lens is an open-source, evidence-first system for observing how artificial intelligence changes geopolitical influence, technological dependency, national capability, and state sovereignty.

The contribution: the programmable Horizon

Most long-horizon agent projects ask how an agent can remember more of its past. This project starts from a different premise: the Horizon already exists, and it is already being programmed. The Horizon is not an agent's long-term memory. It is a shared, forward-facing object outside any one agent—the accumulated option space produced by capital, talent, compute, energy, infrastructure, data, culture, institutions, human networks, standards, and AI systems.

States and other institutions already program this Horizon. Sovereign Lens observes their public actions as writes to eight registers: timers, commitments, parameters, defaults, permissions, priors, capacities, and categories. Each write records who acted, what changed, when it activates, how binding and reversible it is, which sovereign capabilities it affects, and the evidence supporting that observation.

Multiple models analyze the same write independently. Consensus becomes an observation; divergence becomes parallax—a measurement of how different model cultures parse the same change. The observer is therefore also observable.

The hackathon contribution has two parts:

The approach — Programmable Horizons: represent long-term trajectories as inspectable, versioned objects composed from classical assets, human agency, institutions, and AI agents—not as agent memory alone.
The instrument — Sovereign Lens: observe one of the largest active Horizon- programming processes in real time: states and institutions reshaping sovereign capability, dependency, control, and optionality around AI.

In one line: everyone is extending the agent's past; Sovereign Lens makes the programmable future observable.

See the concept memo, the hackathon contribution memo, the build-day toolchain and provenance record, the trajectory_hypothesis schema, and the open-source promo production package.

It is not intended to be another geopolitical news summarizer. Sovereign Lens converts source-linked evidence into a living graph of AI-related actors, assets, relationships, claims, and events. Multiple AI models then evaluate the same evidence independently, while a separate evaluation layer exposes where they agree, where they disagree, which assumptions drive their conclusions, and what evidence could change their views.

The objective is not to reproduce one country's narrative or one model provider's worldview. The objective is to create an independent, plural, and auditable evaluation layer for AI geopolitics.

Current status

Hackathon prototype / experimental research project

The initial prototype was started during Long Horizon Agents Build Day at AGI House on August 22, 2026.

The immediate objective is to demonstrate one complete loop:

Evidence
   ↓
Structured claims and events
   ↓
Temporal geopolitical AI graph
   ↓
Independent model assessments
   ↓
Agreement and disagreement analysis
   ↓
Human-readable provenance

What actually works today

Three states, because two were hiding things:

[x] verified end to end by a committed test or a reproducible run
[~] partly real — the code path exists and is exercised, but the qualifier beside it is doing real work
[ ] specification, not software

Ingestion and extraction

[~] Manual evidence ingestion — 11 source-linked records (publisher, date, type, language, URL) are committed and inspectable. They are hand-authored; no path takes a document in.
Claim and event extraction — nothing reads a document and emits structured claims.

Graph

[~] Actor and relationship graph construction — 12 actors and assets and 15 typed, dated relationships are projected into a graph at any as_of by tested code. The records are hand-authored, not derived from evidence.
Bitemporal snapshot with supersession preserved — facts are filtered on recorded_at <= as_of before validity, so the view reproduces what was knowable then. Superseded facts stay visible rather than being deleted, and each snapshot carries a content digest.
Interactive graph visualization — web/: a 2D relationship graph with a time scrubber, current-versus-superseded fact treatment, and a one-click thirteen-month advance.

Assessment

[~] Multiple perspective-model assessments — one same-prompt option-space panel has provider-API runs from Qwen 3.8 Max, GLM 5.2 and DeepSeek V4 Pro, each with a chatcmpl response id, endpoint profile and region. The Claude Fable 5 entry is not the same kind of artifact: it was captured through a Claude Code session rather than an independent Messages API call, has no provider response id, and is labelled provider_route: claude_code_session with a route_note saying so. A disclosed local Codex analysis is also included. Run metadata and one rejected truncated response are committed under data/model_runs/. The main temporal dossier still uses deterministic fixture text, and there is not yet a general-purpose provider adapter feeding arbitrary cases into the graph.
[~] Consensus and disagreement evaluation — divergence is preserved and shown rather than averaged, and material disagreement is distinguished from consensus. The option-space synthesis compares real provider outputs, but the synthesis remains human-authored; no tested evaluator generates axes automatically.
Evidence and provenance drill-down — every visible claim resolves to its source, publisher, and the account that source speaks for (participant, regulator, oversight). A test asserts every reference in the fixture resolves.

Memory and calibration

[~] Persistent assessment memory — assessments, predictions and methodology lessons persist to SQLite (python3 demo.py --db sovereign.sqlite). Nothing yet retrieves prior beliefs across runs.
Falsifiable prediction ledger — a prediction stores its probability, horizon, and the exact evidence-snapshot hash it was registered against.
Retrospective calibration against later outcomes — one prediction resolves and is scored (Brier 0.1225 against a 0.25 coin-flip baseline) and the accepted lesson moves methodology v1 → v2 with history preserved. Offline demo fixture, and n = 1.

Verify the ticks yourself:

python3 -m unittest discover -s tests    # projection, ledger, trajectories and eight registers
cd web && npm install && npm test        # 43 tests, including cross-language digest conformance

Provider credentials are not required for the offline demo. To reproduce or extend the multi-model panel, copy .env.example to .env, then follow docs/API_ACCESS.md and run python3 scripts/check_api_config.py.

Anything unchecked is specification, not software. Model plurality has now been demonstrated on one bounded analytical question; automated extraction and a general evidence-to-provider-to-evaluator pipeline remain future work.

This project is not production-ready and should not be treated as an authoritative geopolitical intelligence source.

Neutral example convention

All examples in this README are synthetic and use deliberately generic placeholders:

Country 1
Country 2
Country 3
Institution 1
Company 1
Model 1
Infrastructure Project 1

They illustrate the data model and analytical workflow without making claims about real governments, companies, or geopolitical relationships. Production records must use real entity identifiers and verifiable evidence.

Why this exists

AI is increasingly becoming a geopolitical instrument.

Access to foundation models, compute, cloud infrastructure, semiconductors, data, technical talent, deployment expertise, and evaluation capability can materially change the options available to a state. But access is not the same as ownership, control, or sovereignty.

A country may gain substantial AI capability while simultaneously becoming more dependent on another state or foreign provider for:

compute
model updates
API availability
cloud hosting
cybersecurity
technical support
language localization
evaluation tooling
operational continuity

If an external actor withdraws access, changes prices, restricts use, or modifies the system, the recipient state may discover that it possessed access without possessing control.

Sovereign Lens treats AI capability itself as a changing structure of power.

It is designed to help investigate questions such as:

Which states are gaining meaningful AI capability?
Who owns or controls the infrastructure behind that capability?
Which countries are becoming dependent on foreign providers?
Which partnerships increase optionality, and which create lock-in?
Can a state replace its current model, cloud, or compute provider?
Does announced access translate into operational capability?
Can local institutions independently evaluate imported AI systems?
Where do different AI models interpret the same evidence differently?
Which conclusions remain robust across models?
What evidence would change the current assessment?

Core thesis

In the AI era, sovereignty is not only the ability to build or purchase AI. It is the ability to independently evaluate, choose, operate, modify, audit, constrain, replace, and govern it.

A state may possess significant AI capability without possessing full AI sovereignty.

Sovereign Lens distinguishes between several concepts that are often collapsed into one score:

Concept	Question
Capability	What can the state actually do?
Access	What technology can it currently use?
Ownership	What does it legally and operationally control?
Dependency	What capability disappears if an external actor withdraws support?
Optionality	How easily can the state switch providers or operating models?
Influence	What leverage does another actor gain through the relationship?

A country can gain capability while losing optionality. That tension is one of the project's central analytical questions.

Analytical lineage and independence

Sovereign Lens is informed by the Evaluation State thesis: practical AI sovereignty depends not only on frontier-model ownership, but also on a state's capacity to independently evaluate, orchestrate, audit, replace, and govern AI systems.

This is a disclosed analytical perspective, not a claim of value-free neutrality.

The system is designed to make its own framework contestable. Claims remain linked to evidence, assumptions are exposed, model disagreements are preserved, and alternative interpretations can challenge both individual assessments and the project's underlying methodology.

Independence therefore comes from an open and auditable process—not from pretending that the project has no authorship or perspective.

How it works

flowchart LR
    A[Source-linked evidence] --> B[Claims and events]
    B --> C[Temporal evidence graph]
    C --> D1[Perspective model 1]
    C --> D2[Perspective model 2]
    C --> D3[Perspective model 3]
    D1 --> E[Evaluation layer]
    D2 --> E
    D3 --> E
    E --> F[Consensus, disagreement, assumptions, missing evidence]
    F --> G[Human interface]
    G --> H[Assessment history and calibration]
    H --> C

The intended pipeline has six stages:

Ingest evidence from primary and independently verifiable sources.
Extract claims and events without collapsing them into conclusions.
Resolve actors, assets, and relationships while preserving uncertainty.
Update a temporal graph of AI-related geopolitical change.
Ask multiple models to evaluate the same evidence package and graph state.
Compare their assessments and expose agreement, disagreement, assumptions, confidence, and provenance.

Core data model

Object	Purpose	Synthetic examples
Actor	An entity that creates, controls, funds, receives, restricts, or evaluates AI capability	Country 1, Ministry 1, Company 1
Asset / capability	An AI-related resource or operational capacity	Model 1, GPU Cluster 1, Training Program 1
Relationship	A persistent connection between actors or assets	provides, depends_on, hosts, funds, restricts
Event	A time-bounded change in the graph	agreement signed, access granted, restriction imposed
Claim	A structured assertion extracted from evidence	Country 2 provides Model 1 access to Country 1
Evidence	A source that supports, challenges, or contextualizes a claim	agreement, procurement record, technical document
Assessment	An interpretation produced by a model or human analyst	capability increased; dependency may also have increased

Claims are not assessments

Claim:
Country 2 granted Country 1 access to Model 1.

Assessment:
The agreement may increase Country 1's AI capability while also increasing
its dependency on infrastructure controlled by Country 2.

The first statement can potentially be verified from evidence. The second requires interpretation.

Synthetic worked example

Suppose the following events occur:

Country 2 grants Country 1 access to Model 1.
Company 1 begins constructing Infrastructure Project 1 in Country 1.
Country 2 launches a technical training program for Country 1's public sector.
Country 3 offers Country 1 access to an alternative cloud platform.
Country 1 adopts a national AI policy.
Institution 1 creates an independent model-evaluation laboratory in Country 1.

The graph might contain relationships such as:

Country 2
   ├── provides ──> Model 1 ──> accessible_by ──> Country 1
   ├── finances ──> Infrastructure Project 1 ──> located_in ──> Country 1
   └── trains ──> Technical Workforce 1 ──> employed_in ──> Country 1

Country 3
   └── provides ──> Alternative Cloud Service ──> accessible_by ──> Country 1

Institution 1
   └── operates ──> Evaluation Laboratory 1 ──> located_in ──> Country 1

Sovereign Lens would not merely summarize these events. It would ask:

How has Country 1's practical AI capability changed?
Which dependencies increased?
Did vendor optionality increase or decrease?
Who owns and operates the relevant infrastructure?
Can Country 1 continue operating if one provider withdraws?
Which external actor gained influence?
Did Country 1 gain independent evaluation capability?
Do different models interpret the same change differently?

Evidence and provenance

Documents are evidence, not the final analytical unit.

Each evidence record should preserve enough provenance for another contributor or analyst to inspect it:

{
  "id": "evidence_001",
  "source_type": "government_document",
  "publisher": "Institution 1",
  "published_at": "2026-08-01",
  "retrieved_at": "2026-08-22T18:00:00Z",
  "language": "en",
  "url": "https://example.org/document",
  "content_hash": "sha256-placeholder",
  "supports": ["claim_001"],
  "challenges": [],
  "reliability_notes": "Primary source; implementation not yet independently verified."
}

Important provenance fields include:

original source and publisher
publication and retrieval dates
source type and jurisdiction
language and translation method
content hash or archived copy
extraction method and model version
original excerpt
supported and challenged claims
corroborating and contradictory sources
reliability notes

News may be useful for discovery and context, but it should not automatically become the system's primary epistemic layer. Whenever possible, Sovereign Lens should prefer primary, corroborated, or independently verifiable evidence.

Perspective models

Sovereign Lens uses several independently developed models as analytical instruments.

Each perspective model should receive:

the same analytical question
the same evidence package
the same graph snapshot
the same requested time horizon
the same structured output schema

This reduces avoidable context differences while preserving meaningful differences in reasoning.

Models are not country delegates

A model developed in Country 2 should not automatically be interpreted as "Country 2's position," just as a model developed in Country 3 should not automatically represent that state's worldview.

Models may differ because of:

training distributions
architecture
post-training
safety policies
system prompts
language context
access to information
institutional incentives
reasoning behavior

Those differences can be analytically useful, but they should not be overinterpreted.

Example assessment schema

{
  "agent": "model_1",
  "question": "How does Event 1 affect Country 1's AI sovereignty?",
  "assessment": "Country 1 gains capability but may become more dependent on externally controlled infrastructure.",
  "confidence": 0.68,
  "drivers": [
    "model access",
    "technical training",
    "external cloud dependency"
  ],
  "counterarguments": [
    "Country 1 retains access to an alternative provider",
    "the agreement may have limited operational scope"
  ],
  "assumptions": [
    "the announced program becomes operational",
    "Country 1 cannot host the model independently"
  ],
  "supporting_evidence": ["evidence_001"],
  "contradictory_evidence": [],
  "missing_evidence": [
    "contract termination rights",
    "local hosting rights",
    "availability of alternative providers"
  ]
}

Evaluation layer

The evaluator compares model outputs rather than simply averaging them.

It should identify:

Consensus: conclusions several models reach independently
Divergence: conclusions on which models materially disagree
Assumptions: explicit or hidden premises behind each conclusion
Evidence selection: which sources each model relied upon or ignored
Confidence: how certain each model claims to be
Missing evidence: what information could resolve uncertainty
Prompt sensitivity: whether small framing changes alter the outcome

Synthetic example:

Question:
Did Event 1 increase Country 1's dependency on Country 2?

Model 1: Likely yes — confidence 71%
Model 2: Possibly — confidence 55%
Model 3: Insufficient evidence — confidence 43%
Model 4: Primarily capability-building — confidence 64%

The most useful output may not be the average. The useful output is often:

Why do they disagree?

Sovereignty dimensions

Sovereign Lens should not reduce sovereignty to one unexplained number.

A more defensible representation is a multidimensional profile:

Dimension	Core question
Model sovereignty	Can the state access, inspect, modify, host, and replace its models?
Compute sovereignty	Who owns and controls the compute?
Infrastructure sovereignty	Where are cloud services and data centers located, operated, and governed?
Data sovereignty	Who controls training, operational, and citizen data?
Talent sovereignty	Does the state possess domestic technical and institutional capability?
Vendor optionality	Can the state realistically change providers?
Evaluation sovereignty	Can the state independently test whether systems work and remain acceptable?
Governance sovereignty	Can institutions audit, authorize, constrain, suspend, and replace AI systems?

Illustrative qualitative profile:

Country 1

Model sovereignty:          moderate
Compute sovereignty:        low
Infrastructure sovereignty: low–moderate
Data sovereignty:           moderate–high
Talent sovereignty:         moderate
Vendor optionality:         low
Evaluation sovereignty:     emerging
Governance sovereignty:     moderate–high

These bands are illustrative only. Sovereign Lens should not publish numerical sovereignty scores until the definitions, evidence rules, weighting choices, uncertainty treatment, and validation methodology are documented and reviewable.

Potential data sources

Sovereign Lens should prioritize evidence diversity and provenance over raw document volume.

Government and policy

national AI strategies
legislation and executive orders
ministry announcements
bilateral and digital-cooperation agreements
procurement records and tenders
export controls and sanctions
budgets and parliamentary records
development programs

Models and software capability

API and model availability
open-weight releases
model localization and language support
licensing, hosting, and fine-tuning rights
government model deployments
usage restrictions and withdrawal conditions

Compute and infrastructure

GPU procurement
national supercomputers
data-center construction
cloud-region expansion
sovereign-cloud agreements
semiconductor imports and exports
power availability
subsea cable and network infrastructure

Corporate, talent, and research activity

investments and joint ventures
public-sector contracts
research collaborations
acquisitions and infrastructure deployments
technical training programs
university partnerships
researcher migration
workforce and laboratory formation

Economic and multimodal evidence

trade and foreign-direct-investment data
infrastructure financing
geospatial observations
shipping and aviation data
network measurements
power-consumption signals
data-center and semiconductor supply-chain indicators

Multimodal sources are future-facing unless explicitly implemented and documented in the repository.

Hackathon MVP

The hackathon version should remain narrow and verifiable.

A credible same-day MVP is one complete evidence-to-evaluation loop with approximately:

5–15 source-linked evidence records
3–5 actors
3–10 AI assets or capabilities
10–30 structured relationships
3 perspective models
1 evaluator
1 graph view
1 disagreement panel
source-level provenance

The demo should prove that the system can:

ingest a small evidence package
extract actors, assets, relationships, claims, and events
update a temporal graph
ask several models to assess the same change independently
identify consensus and disagreement
expose assumptions and missing evidence
trace every conclusion back to its sources

Long-horizon memory and calibration

See docs/LONG_HORIZON_CONCEPT.md for the conceptual framing: the horizon already exists, and the engineering task is to make experience, succession, and learning auditable across agent generations.

Geopolitical analysis becomes more useful when the system can compare previous assessments with later outcomes.

Previous assessment
        ↓
New evidence
        ↓
Observed outcome
        ↓
Calibration review
        ↓
Updated analytical method

Synthetic example:

Initial assessment:
70% probability that an announced agreement produces operational deployment
within six months.

Six-month observation:
No operational deployment detected.

Potential learning:
The system may have overweighted official announcements and underweighted
implementation capacity.

A persistent memory layer may store:

prior assessments
model predictions
reasoning summaries
discovered evidence
past disagreements
calibration errors
analyst corrections
unresolved questions
recurring assumptions

Design principles

Evidence before narrative

Narratives should be constructed from observable, traceable evidence.

Claims before conclusions

The system should separate what happened from what it may mean.

Events before documents

Documents are sources. Events and graph changes are the analytical objects.

Graph before feed

Geopolitical relationships evolve over time and should not be reduced to a content stream.

Multiple models before one oracle

No single model should define the system's worldview.

Disagreement is information

Differences should be exposed and explained rather than averaged away automatically.

Provenance everywhere

Users should always be able to ask: Why does the system believe this?

Sovereignty means optionality

Access to advanced AI is not equivalent to sovereign control.

Independence is architectural

Models, providers, prompts, and data sources should be replaceable.

Prototype target stack

The target stack may include:

Backend: Python, FastAPI, Pydantic
Frontend: Next.js, React, TypeScript
Agent orchestration: LangGraph or a lightweight asynchronous runner
Graph: NetworkX for the prototype; Neo4j or PostgreSQL-based graph storage later
Storage: JSON or SQLite for the prototype; PostgreSQL and pgvector later
Models: provider adapters for OpenAI, Anthropic, Google, GLM, Qwen, DeepSeek, and open-weight models
Memory: Stash or an append-only assessment history
Visualization: React Flow, Cytoscape.js, D3.js, or a map layer

This list describes the intended prototype architecture. The actual repository dependencies and committed code are the source of truth.

Local development

Micro partner integrations

The optional integrations/ slice demonstrates the event-partner architecture without putting a hosted service in the analytical critical path:

python3 -m integrations.micro --out /tmp/sovereign-lens-integrations

It emits a Temporal-ready reassessment task, a Stash-ready research-memory record, and a Coframe presentation-safety contract. The optional Temporal workflow is in integrations/temporal_workflow.py; the web interface emits the corresponding safe UI events. See the live build integration notes.

Official references: Temporal documentation, Stash documentation, and Coframe documentation.

Three-minute offline demo

The initial demo exercises the long-horizon memory loop with Python and SQLite, without API keys or network access:

python3 demo.py --paced

It demonstrates a dated graph snapshot, three transparent perspective fixtures, a falsifiable prediction, later evidence, bitemporal supersession, calibration, and a human-reviewed methodology update. The perspective text is deterministic demo data, not output from live model providers.

Run the tests with:

python3 -m unittest discover -s tests -v

See docs/PITCH.md for the three-minute presentation and docs/DEVELOPMENT_PLAN.md for the database and build plan.

Presentation interface

The web/ dossier renders the same loop as one screen, built for a projector at 1440x900:

cd web
npm install
npm run dev      # http://localhost:5173
npm test         # 43 unit and component tests

It reads a frozen local fixture through a DataSource seam, so it can be repointed at a demo.py --json export or an HTTP endpoint without changing a component. See web/README.md for the design rules, the data seam, and what the screen explicitly does not claim.

Planned full application

git clone https://github.com/Dim25/Sovereign-Lens.git
cd Sovereign-Lens

The two run paths above are verified against the committed repository. The ingestion, extraction and provider-adapter layers are not built, so there is no launch command for them yet.

Do not commit API keys, private evidence, restricted data, or credentials.

Seed evidence

The prototype is intended to use a small set of real, source-linked records rather than documentation-only fixtures.

One case is committed and inspectable: case_uae_us_ai_infrastructure, in web/src/data/uae-us-ai-infrastructure.fixture.json.

	Count
Sources	11
Actors and assets	12
Typed relationships	15
Facts	11
Perspective assessments	6 (two snapshots x three perspectives)
Disagreement axes	5
Predictions	1 registered, 1 resolved
Methodology lessons	1 accepted

Every source is a real, public, first-party or oversight document with a live URL: UAE Embassy statements, Microsoft and G42 announcements, the Congressional Record, a Commerce export statement, and a Congressional Research Service overview. Two facts are lifted verbatim from demo.py so the screen and the terminal cannot disagree.

The assessment text over that evidence is deterministic fixture data, not model output. The 2026 capacity figure is reported by a participant government and is not independently verified.

Synthetic Country 1 / Country 2 examples are used only for explaining the methodology.

Limitations and responsible use

Known or expected limitations include:

incomplete or delayed public evidence
strategic misinformation
translation errors
extraction mistakes and hallucinations
duplicated or circular sourcing
hidden contractual terms
ambiguous entity resolution
shared biases across models
model-version drift
prompt sensitivity
false confidence and false precision
difficulty distinguishing announced access from operational capability

Sovereign Lens should not be used as the sole basis for:

military action
sanctions
diplomatic retaliation
intelligence attribution
public accusations
critical-infrastructure intervention
automated policy decisions

High-impact conclusions require independent human review and additional verification.

Sovereign Lens is not a geopolitical prediction oracle, propaganda engine, military targeting system, automated foreign-policy decision-maker, or substitute for expert analysis.

Contributing

Contributions are welcome, especially across countries, disciplines, languages, and model ecosystems.

Useful contribution areas include:

agent orchestration
knowledge graphs
geopolitical datasets
source verification
structured extraction
model evaluation
multimodal ingestion
international relations
sovereign AI research
provenance and uncertainty
long-horizon memory
calibration
visualization
multilingual analysis

A typical contribution flow:

Open an issue describing the proposed change.
Create a focused branch.
Add or update tests where possible.
Document methodological assumptions.
Submit a pull request.
Explain how the change affects provenance, reproducibility, or independence.

Evidence contributions should include enough provenance for another contributor to verify the source.

Upstream contributions are strongly encouraged, but they are not required by the Apache 2.0 license.

Open-source philosophy

Sovereign Lens is intended as open infrastructure.

The project should remain:

inspectable
reproducible
evidence-first
model-plural
internationally usable
vendor-independent
open to methodological criticism

No single model provider, state actor, institution, or source should automatically become the project's source of truth.

License

Licensed under the Apache License 2.0.

See LICENSE for details.

Apache 2.0 permits commercial and non-commercial use, modification, and distribution while preserving notices and providing an explicit patent grant.

Maintainer

Dima Filatov

Building Sovereign Lens at the intersection of sovereign AI, multi-agent evaluation, geopolitical intelligence, and deployed AI systems.

LinkedIn: linkedin.com/in/dim25

Hackathon origin

Initial prototype created during:

Long Horizon Agents Build Day AGI House August 22, 2026

We are not building another geopolitical news summarizer. We are building a living graph of how AI capability moves power between states—and using independent AI models to evaluate what those changes mean.

Name		Name	Last commit message	Last commit date
config		config
data		data
docs		docs
integrations		integrations
promo		promo
schemas		schemas
scripts		scripts
site		site
tests		tests
web		web
.env.example		.env.example
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
demo.py		demo.py
wrangler.jsonc		wrangler.jsonc

Latest commit

History

Folders and files

Repository files navigation