World layer

The world layer for teams building software and AI agents

Phasis gives agents a maintained view of the world before they act: current domain state, evidence, verification, and the private Alpha that makes it yours.

Request access Read the research

The problem

Models can reason. They still do not know your world.

Scale compresses how to think. It does not absorb what is current, private, or specific to the work in front of your team.

Shared World

One verified map, maintained once.

Public domain reality, sources, dates, standards, and failure modes are kept current outside the model and served to every agent that needs them.

Alpha

Your private edge stays private.

Team data, policies, traces, and working sets sit as a thin overlay on the shared world. They do not flow back or sideways.

World API

Orient before the agent acts.

One orient call returns what matters, what changed, what to trust, what to check, and which private differences apply.

Verification

Not every trace becomes world.

Production work passes through source, freshness, scope, and privacy gates before it becomes shared state, Alpha, an eval, or nothing.

Compounding

Work should make the next run smarter.

Today, agents rebuild the same context from cold. Phasis captures the verifiable parts so later agents start ahead.

For builders

Use any model. Give it a world.

Smaller models get missing domain state. Frontier models spend reasoning on the hard remainder instead of reconstructing reality.

Product surfaces

API, endpoint, workspace.

World API for direct orientation. Adept when inference should arrive with the world built in. Workspace for oversight, traces, evals, and approvals.

Start here

A maintained world before action.

Build agents that orient, act, trace, and re-orient as the world changes.

Request access Read the essays

Shared World · Alpha · orient · evidence · verification · traces

Same question, two answers

Same question, two answers.

Ask a raw model for a domain decision and you get plausible prose. Give the agent a maintained world first and you get work shaped by current state and a standard of judgment.

Raw model

Names broad trends without ranking what mattersplausible but not governed
Mixes weak sources with stronger evidencesource hierarchy unclear
Sounds confident where the data is thinunsupported claim risk
Misses the workflow format the buyer needsreview burden remains
Does not record why the answer should improveno useful trace
Requires expert cleanup before it can shipaccepted output uncertain

expert review still required

Phasis world layerVerified path

orient({ domain: "taste & flavor", task: "brand strategy" })

worldReturns the current shared map, sources, dates, and live caveats.

alphaAdds the team's private policies, data, traces, and working set.

orientClassifies the task, stakes, evidence standard, and likely failure mode.

verifyChecks groundedness, unsupported claims, freshness, and rubric fit.

traceRecords what can safely improve the world before the next run.

What it is

A world agents can call before they act.

Phasis is the maintained layer between general models and real work: Shared World, private Alpha, evidence, verification, and traces that make the next run start ahead.

01: Maintains the shared world outside the modelWorld API
02: Adds private Alpha for one team without leaking it sideways
03: Orients each task to current state, evidence, and risk
04: Selects the working set the agent actually needs
05: Verifies sources, freshness, groundedness, and rubric fit
06: Turns approved traces into compounding world state

Shared World · Alpha · orient · evidence · verification · traces

The clean line vs retrieval: retrieval fetches what already exists. The world layer maintains what is true, current, and usable before an agent starts work.

Coverage

Worlds for domains where current context matters.

Cultural arenasLiveBrand strategyLiveAI landscapeCybersecurityFinancial marketsLegalHealthcareRoboticsInfrastructureEnterprise buyingResearch workflows

Every world is maintained around the same operating questions: what evidence matters, what a generic model would miss, and what standard the output must satisfy.

Adept API

Use Adept like a model endpoint.

Send a task, a domain, and the work you need done. Adept handles orientation, evidence, verification, and traceability behind the endpoint.

REST API
Streaming
Evals
Traces

Request shape

{
  "domain": "cultural-arenas",
  "task": "analyze_brand_position",
  "input": "How is this brand showing up in wellness culture?",
  "mode": "verified"
}

Response shape

{
  "answer": "...",
  "evidence": [{ "source": "...", "why_it_matters": "..." }],
  "verification": { "status": "passed", "caveats": [] },
  "trace": { "task_class": "brand_strategy", "route": "verified" }
}

Proof

Proof should measure usable work.

Accepted Output: MeasuredAdept is evaluated against raw models on real domain workflow examples.
Review Burden: TrackedBenchmarks look at expert edit time, unsupported claims, and workflow fit.
Traceability: Built InEvidence, verifier results, and run traces make the output inspectable.

Manifesto

Fluency is abundant. Domain judgment is scarce.

Models get smarter every month. That does not make them competent inside your domain. Drop an agent into strategy, compliance, brand, or market intel and it has to figure out a world it was never trained on — so it fills the gaps, builds on weak assumptions, and goes deeper in the wrong direction. Confidently. It doesn’t know what it doesn’t know.

Phasis builds the world layer around the model: Shared World, private Alpha, evidence, verification, and traces before the work is trusted. The model stays general. The agent gets a world it can act in.

FAQ

Questions, answered head-on.

What is Phasis?: Phasis builds the world layer for software and AI agents: Shared Worlds, private Alpha, evidence, verification, and traces before work begins.
What is Adept?: Adept is inference with the Phasis world layer built in. It works like a model API, but behind the endpoint Phasis orients the task, selects evidence, routes the model path, verifies the work, and records traces.
How is it different from web search or RAG in my agent?: RAG retrieves context. The world layer starts earlier: it maintains what is true, current, and relevant, then decides what kind of judgment the task requires and what evidence should count.
What comes back from a call?: A domain-shaped work product: the answer, the evidence used where available, verification notes, caveats, and trace metadata on supported plans.
Which domains do you support?: Phasis-owned worlds come first, especially cultural and brand strategy work. Additional worlds are maintained with workflows, source material, and eval examples. We do not treat every domain as equally deep on day one.
How do you measure accuracy?: We measure by domain outcome, not generic accuracy. The core metrics are accepted-output rate, unsupported-claim rate, expert review time, traceability, and cost per accepted output.
Does Adept learn from use?: The world layer can improve from permitted feedback, traces, evaluations, and public or licensed research signals. Customer-specific learning stays customer-specific unless explicitly configured otherwise.
How do I access it?: Request access for World API or Adept. The first public story is one orient call before action; tool and workflow integrations can be layered around that surface as supported products mature.
Where does the data come from?: Public, licensed, customer-provided, and workflow-derived sources, depending on the domain and plan. Customer-provided data is handled as customer-specific unless a separate agreement says otherwise.

Phasis orients the work, supplies current domain state, checks the evidence, and keeps the world improving between runs.

Request access Read the docs