Faros
Faros. Ethen's evolving intelligence program.
Faros is where Ethen builds the capability to choose, evaluate, adapt and, over time, develop its own intelligence. Today, Faros in Ethen Chat runs on Meta's Muse model.
ConceptualConceptual visualization. A beacon sweeping a field of models: Faros chooses, evaluates and improves the intelligence it uses.
Faros today
An intelligence program, not a model launch.
Faros is Ethen's model and intelligence-development program. In Ethen Chat today it runs on Meta's Muse model, reached through Meta's API. The Ethen workspace remains a separate product, and proprietary models open only behind published gates.
- Current
In Ethen Chat
Faros in Ethen Chat runs on Meta's Muse model, reached through Meta's API. You can use it today.
- Building
Routing policy
Faros chooses execution configurations with rules today. Logging of candidate sets and selection probabilities is being added so learned policies can be tested fairly later.
- Gated
Proprietary models
Ethen does not train its own models today. First-party verifiers, a learned router and specialist models each open only when a published gate is met.
Model development system
How intelligence gets built here.
Eleven stages, one loop. Select a stage to see what it means, what Faros does at that stage today, and the research behind it.
Stage 01 / 11
ResearchingResearch question
- What the stage means
- Every cycle starts from a question that can be falsified: which configuration produces more verified outcomes per unit of cost, and where do models fail?
- What Faros does here today
- Faros questions are written up as position papers, surveys and protocols in Ethen Research Lab.
ConceptualStages describe the development loop; each stage's status comes from the record.
Development status
Where each workstream actually stands.
Each row has one status, assigned only from the implementation and research record. Most of the work is early, and the board says so.
| Workstream | Researching | Designing | Building | Evaluating | Validated | Future |
|---|---|---|---|---|---|---|
| Rules-based routing policyRuns in Chat; candidate-set logging being added. | Building | |||||
| Model certification harnessExists for Chat, text scope. | Evaluating | |||||
| Sealed certification and safety setsAt least 200 prompts plus a safety set, planned. | Designing | |||||
| Propensity loggingResearch note published; logging in design. | Designing | |||||
| Verifier calibration (Ethen-Verify)Protocol published; not run. | Researching | |||||
| Model-change assuranceProtocol published; not run. | Researching | |||||
| Capability transferMethods paper and protocol; not run. | Researching | |||||
| Learned routing (shadow, then learned)Deferred until enough logged decisions exist per task class. | Researching | |||||
| Specialist adaptation, fine-tuning, distillationGated on volume, sealed evals and licences. | Future | |||||
| Configuration synthesisLong-term bet: choose harness, tools and verifier together. | Future | |||||
| Foundation-model trainingNot pursued now. Requires rights-cleared data at scale, passed research gates and outside capital. | Future |
No Faros workstream has reached Validated yet.
Evaluation
Evaluation is how intelligence is built, not a scoreboard added afterwards.
Faros treats measurement as part of model development. Each method below has a question, a method and a decision gate, and each one says how far it has got.
Verified outcomes
- Question
- Did the work actually succeed, by a check the producer did not control?
- Method
- Grade with deterministic checks first, then programmatic verifiers, then expert rubrics, then calibrated judges. A producer is never the sole acceptor of its own output.
- Decision gate
- Verified success, not plausible output, is the unit every other gate is measured in.
- Evidence status
- Proposal / Hypothesis
Learned routing must beat strong rules
Research protocolThis is the decision gate from the routing protocol. The axes show the comparison the protocol will make. No results are plotted, because the experiment has not been run.
Evaluation checkpoints (logged decisions per task class)
- 1K
- 5K
- 10K
- 25K
- 50K
| Cost | At least 15% lower full cost per verified outcome than rules |
|---|---|
| Quality | One-sided 95% lower confidence bound above −2 percentage points |
| Safety and latency | No regression |
| Results | None; the experiment has not been run |
Architecture
Faros is wider than routing, and Gateway is a separate product.
Research feeds the program, the program produces intelligence and policy, and products reach that intelligence through runtime access. Select a layer to trace its connections.
Faros
Model development, evaluation, data systems and routing policy. The program that turns research into intelligence.
Connects to: Ethen Research Lab, Gateway & runtime access
ConceptualA structure diagram of the program, not a deployment map.
Model Choice
Choose models directly.
Explore models through Open Intelligence and use supported access paths. Faros is optional.
Explore open modelsGateway
Access through Gateway.
Gateway provides access and routing for supported providers. It works without Faros.
Explore Gateway
Capability evolution
From integrating intelligence to developing it.
External intelligence integration
Faros runs on rented frontier models, chosen and configured by Ethen.
RoadmapThese are capability eras, not version releases. None has a date. Each one opens only when its gate is met.
Faros research
Explore the research behind Faros.
Faros research is a program of Ethen Research Lab. Research direction remains separate from available capabilities.
- P08Faros: Researching How Intelligence Should Choose IntelligencePosition PaperResearch Synthesis
- P09Why Learned AI Model Routing Must Beat Good RulesSurveyResearch Synthesis
- P11Model Change Assurance: Testing AI Upgrades Before They Reach Real WorkResearch NoteResearch Synthesis
- P20Why Better Foundation Models May Make Evaluation More Valuable, Not LessPosition PaperResearch Synthesis
- P26How to Test Whether Learned AI Routing Beats Strong RulesResearch ProtocolProtocol / Planned Experiment
- P33A Research Protocol for Model Change AssuranceResearch ProtocolProtocol / Planned Experiment
- P34Why AI Routers Should Log Propensities From Day OneResearch NoteResearch Synthesis
One system
Research, intelligence and product compound together.
The loop runs in every direction. Product use raises research questions, evaluations reopen old answers, and some research ends in knowledge that never becomes a product.
Discovers and validates
Ethen Research Lab
The research organization studying intelligence, computing and science, with an explicit evidence status on everything it publishes.
Explore Ethen Research LabTurns research into intelligenceYou are here
Faros
Ethen's model and intelligence-development program: evaluation, routing, data and adaptation, building toward increasingly proprietary capability.
Puts intelligence to work
Ethen
The product people use to think, research, write and build with AI.
Explore EthenCloses the loop
Evidence
Outcomes, failures and evaluations become the questions the next round of research starts from.
FAQ
What is Faros?
Faros is Ethen's model and intelligence-development program. In Ethen Chat, Faros is also the name of the default conversational experience, which today runs on Meta's Muse model, reached through Meta's API. The Privacy Policy and Subprocessors list name every model provider.
Has Ethen trained its own Faros model?
No. Ethen does not train its own models today. First-party verifiers, a learned router and specialist models are each gated on published evidence, and foundation-model training is not pursued now.
Does Faros replace the Ethen workspace?
The Ethen workspace remains a separate product. Faros is the intelligence program behind it.
Do I need Faros to use models or Gateway?
You can select models directly and use Gateway without Faros.
Follow the evidence as it arrives.
Read the Faros research record, or use Faros in Ethen Chat today.