The world layer for AI agents

Worlds

Logistics World
Medical Research World
Healthcare World
Creator World
Energy World
Retail World
Construction World
Agriculture World
Logistics World
Medical Research World
Healthcare World
Creator World
Energy World
Retail World
Construction World
Agriculture World

Models gave AI intelligence.
They need Worlds.

Phasis builds and maintains live, evidence-backed Worlds so agents start oriented before work begins and stay current as reality changes.

Try /orient free Read the thesis

01Start hereLive public endpoint

GET /v1/orient · Free public endpoint · No signup

One call before
execution.

Send a task to /orient. Phasis returns the relevant World frames your agent needs to begin: current state, material changes, supporting evidence, and uncertainty.

What is current
What changed
What the evidence supports
What remains uncertain

Try /orient free View API docs

The public tier returns a compact result set and is rate-limited.

/orient playground

GET /v1/orient

curl "https://api.phasis.ai/v1/orient?query=municipal%20bond%20outlook&limit=3"

Example JSON200 OK

{
  "status": "sample",
  "endpoint": "/v1/orient",
  "results": [
    {
      "name": "improvement-bond-vs-final-acceptance",
      "type": "frame",
      "domain": "municipal"
    },
    {
      "name": "lease-vs-bond-accounting",
      "type": "frame",
      "domain": "municipal"
    }
  ]
}

The first public Phasis surface. No account or API key required.

02Without a World

Rebuilding every run

Your model is smart.
Its reality is broken.

Building agents is easy. Keeping them grounded is the hard part. When you rely on standard web search or naive vector retrieval, your agent starts every task from zero.

Context Bloat

You fetch irrelevant noise, wasting tokens and confusing the model.

Stale Reality

The world changes, but your model’s weights don’t.

Cascading Hallucinations

One hallucinated fact early in an autonomous loop derails the entire run.

Don’t let your agent guess. Give it a World.

03Orientation

Noise excluded before inference

Orient your agent
before it acts.

Phasis builds and maintains Worlds: live, evidence-backed graphs of complex domains. We track the entities, rules, and real-time changes so your agent doesn’t have to.

Standard vector search & web RAG

Ungoverned retrieval fetches noisy, irrelevant chunks (Coldplay lyrics, paint catalogs, word definitions). Every run forces your model to filter out bad context on your dime.

Phasis /orient response

Pre-task domain alignment. Entities, regulatory authorities, and clinical evidence are already structured. Out-of-bounds noise is categorically excluded before the LLM processes a single token.

Asked if yellow dye is bad for you, the agent read Coldplay lyrics. Then it gave you health advice.

Illustrative comparison — not a recorded run.

No World

The agent orienting itself, from zero

Activity · Ungoverned retrieval

taskis yellow dye bad for me

open.spotify.comYellow — song and lyrics by Coldplaymatched ‘yellow’
fda.govHHS, FDA to Phase Out Petroleum-Based Synthetic Dyes
en.wikipedia.orgYellow — a colour between green and orangematched ‘yellow’
merriam-webster.comYELLOW Definition & Meaningdefined a word
publications.aap.orgFood Additives and Child Health
adobe.comThe Color Yellow | Adobe Expressmatched ‘yellow’

Many additional sources · few needed to answer · no authority weighting

Broad retrieval · weak relevance · human review required

In a World

What it already knows, before research begins

Pre-task

worldfood safety · color additives

Entities
Yellow 5, Yellow 6, tartrazine, sunset yellow, and related additives.
Authorities
FDA and EFSA decide regulatory status. A trade association does not.
Evidence
Clinical literature for adverse reactions. The regulatory record for what is permitted today.
Exclusions
Spotify. Adobe. Paint catalogs. Chemical suppliers. Categorically out.
What changes
The trade association never enters the answer. The contested finding stays labeled contested.

built once · shared across every run · one call before the work · same model, same workflow

Built once. Shared across every run. Works instantly with the models, frameworks, and workflows you already use.

The world layer

What a World is made of

Models reason.
Worlds orient.

To orient an agent is to give it the right starting point before work begins.

A World is the maintained, evidence-backed state of the domain. Shared across agents. Kept up to date over time.

What is current

The state agents should begin from now.

What changed

The material changes since the last task.

What the evidence supports

Claims remain connected to their sources.

EntityA thing with maintained state.
EventA change that alters the task.
ClaimA current statement about the world.
EvidenceThe basis for what can be trusted.
RuleA constraint on what happens next.
TraceWhat ran, what it used, what changed.

04The benchmark

HalluHard · 950 questions

Confident is not the same as correct.

Your agent cannot tell the difference between a fact and a hallucination. We can. HalluHard puts 950 questions across law, medicine, research, and code to a model over a three-turn conversation, then checks every factual claim against the source it cites. Does that source exist, and does it actually say what the model says it says.

Hallucination rate below, where lower is better. Even the best frontier models, running live web search, cannot support about a third of their factual claims with the source they give.

70.9%GLM-5 Thinking
69.5%Gemini 3 Flash
65.6%Claude Sonnet 4.5
64.8%GPT-5 Thinking
63.7%Claude Sonnet 4.6
63.6%Kimi K2.6 Thinking
61.9%Gemini 3 Pro
60.9%Claude Opus 4.6
60.0%Claude Opus 4.5
59.7%Claude Sonnet 5
59.6%Claude Opus 4.7
59.1%Kimi K2.5 · Web Search
58.8%GPT-5.2
57.1%Gemini 3.1 Pro
53.8%GPT-5.3
53.8%GPT-5.2 Thinking
50.9%Claude Fable 5
49.8%GPT-5.5 Thinking
46.8%GPT-5.4 Thinking
43.5%GPT-5.3 · Web Search
39.7%GLM-5 Thinking · Web Search
38.9%Claude Fable 5 · Web Search
38.2%GPT-5.2 Thinking · Web Search
31.9%GPT-5.4 Thinking · Web Search
30.2%Claude Opus 4.5 · Web Search
16.0%Phasis Adept

ProprietaryOpen weightPhasis, internal run

Phasis drops the hallucination rate by declining the claims its evidence does not support instead of asserting them anyway. Stop guessing. Start knowing.

Model results: HalluHard, Fan et al., EPFL, halluhard.com, CC BY 4.0. The Phasis figure is the Acquire loop, measured on an internal run of the same benchmark over a subset of its questions with our own judge, not a leaderboard submission.

In production

Paid agent workflows

Same model. Same inputs. Better work.

With the model, workflow, and inputs held constant, adding a Phasis World improved task accuracy by 15%.

+15%

Task accuracy

Same model, same workflow, same inputs.

125,000

Paid World calls

Approximately three months of sustained paid usage in a live agent workflow.

Live in production

Used in paid agent workflows through channel partners across technology and consumer strategy work.

The model did not change. The World did.

Phasis company data, July 2026.

05The capability stackFor the engineer

An API designed for production agents.

Start free. Add depth when your workload demands it.

/orient

(Public & paid) One call to fetch the current state, recent changes, and evidence for a specific task.

/acquire

Hydrate deep supporting evidence when the agent is executing complex, multi-step reasoning.

/attentive & /alpha

(Private systems) Compile task-sized working sets, manage private organisation deltas, and gain full observability — traces and lineage — for enterprise scale.

01Orient

01Live

/orient free

GET /v1/orient

Give an agent relevant World frames before it acts.

02Live

/orient paid

GET /v1/orient

The same call with semantic retrieval and model work on the frames it returns.

02Acquire

03Live

/acquire

acquire/*

Find and hydrate supporting evidence when the task needs more.

03Maintain and improve

04Private system

/attentive

attentive/compile

Compile task-sized working sets and summaries.

05Private system

/arc

arc/traces

Traces, lineage, and observability across every step.

06Private system

/alpha

alpha/*

Private deltas, rules, decisions, and outcomes for your organisation.

Availability labels describe current product status. /orient free needs no account; paid /orient and /acquire are live behind an issued key.

06Supported WorldsSelected active Worlds

One API layer. Any domain.

If a domain has changing facts, strict rules, and identifiable sources, we maintain a World for it. We handle the data pipelines, entity tracking, and evidence linking. You just build the agent.