Check an AI plan, trace, or term against fixed, plain-language rules. The tools are deterministic and run in your browser.
Your plan and search text are analyzed locally. Atlas does not place submitted text in request URLs or send it to a model or API. Cloudflare may collect ordinary request metadata plus page-performance and visit measurements; read Cloudflare's RUM description (opens in a new tab). This browser may retain a one-use Quick handoff, missed-search history, or verification checklist.
See a sample Gap Check before using your own plan
Live · the real tool, running on your device
This fixed example shows how the tool separates named gaps from controls it can find. It is a demonstration, not a verdict.
I HAVE A
Plan or policy
Gap Check finds what a plan leaves unhandled: a risk with no control, autonomy with no oversight. Framework Audit scores what's there.
··
I HAVE AN
Agent's trace
The Priority Tracer catches ordering drift: the moment a goal quietly climbs over a rule it was supposed to obey — and credits where the rule held.
·CLI for pipelines in the repo
I WANT THE
Words explained
The candidate source inventory records 439 terms. The current Ask and Explore snapshot contains 435; the public data-sync baseline records the four-term gap.
Candidate source inventory: 439 vocabulary records · 6 areas · 193 source records · Ask and Explore snapshot: 435 records · free to read
Get the app:
Android — tap Install app above (or Chrome menu › Add to Home screen) ·
iPhone/iPad — in Safari tap Share › Add to Home Screen ·
Computer — in Chrome/Edge click the install icon at the right end of the address bar.
Free. No app store or account. A successful install pre-caches a defined subset of core routes and engines, not every page; optional assets and an existing browser cache can vary. Inspect the exact cache list (opens in a new tab). Cloudflare Web Analytics measures page visits and performance separately from Atlas analysis input.
Browse by area
These are the six topic filters available inside Explore.
FoundationsWhat it's built from
Data & ProvenanceWhat it's anchored to
Agentic MechanicsHow it acts
Failure Modes & RisksWhat goes wrong
Control & SafetyHow it's constrained
Oversight & GovernanceWho's accountable
What is this?
Delta Atlas is a function-first glossary of language used in agentic AI and its governance. Each recorded term is described by what it does, then linked to related functions and controls. A candidate method called State-Delta organizes claims, observations, proposed changes, and accepted state as separate things; the public evidence status below shows where that method is and is not yet fully applied.
Why not just ask a chatbot?
You can, and for many things a chatbot is useful. But a generated answer can vary with model, context, and prompt history; long or conflicting context can also blur old and new claims.
Delta Atlas uses fixed browser code over recorded data snapshots. With the same pinned engine, snapshot, and input, its tested scoring behavior is reproducible. That does not make the underlying definitions true, and several older pages still carry different snapshots. The method is to keep claims small, expose those mismatches, and require evidence before promoting them.
Why?
The field invents words faster than most people can learn them, and that gap quietly decides who gets to take part. Someone entering the AI workforce should be able to find out what an Atlas-specific term such as "ontology gate," or a broader term such as "guardrail" or "excessive agency," means without a course and without spin.
This is open research: free to read, check, and build on. Helping people understand the work is part of doing the work.
What this doesn't do
Where it sits: these tools compare stated text with explicit browser-side rules. They can point to missing or conflicting structure; they do not enforce policy, authorize an action, observe the outside world, or bind a system.
It does not certify anything. The groundedness score is a source-support heuristic, not a truth, confidence, acceptance, or safety result.
It is not legal, compliance, or professional advice, and it is not trying to out-dictionary the dictionary. It focuses on a deliberately selected subset of terms for understanding agentic AI.
Evidence status
All 439 records currently carry the repository label reviewed. That label is not yet a per-term receipt: reviewer identity, review date, exact source locator, and source digest are not recorded consistently. Some public tools also still contain older generated vocabulary snapshots. The public data-sync baseline (opens in a new tab) records those differences instead of hiding them. The next safe step is an owner-approved materialization policy and receipt-grade per-term evidence.
Research note
Delta Atlas is independent open research, not a product certification. The tools are shared so people can inspect the rules, reproduce the checks, and report where they fail. A falsified claim or broken fixture is useful evidence: it should lead to a recorded correction, not a hidden rewrite.
Everyday tools
Look up a word in the search above. Then these two.
General tool
Not about AI vocabulary. It checks any framework or plan for coherence, on any topic.