llms.txt

# Delta Atlas

> Delta Atlas is a set of free, deterministic tools that read plans, frameworks, and AI
> reasoning traces and point at the places where they quietly contradict themselves. It runs
> entirely on your own device — no account, no network call, no AI model at runtime — so the
> answer is the same for everyone, every time, and nothing you submit ever leaves the machine.
> If you are an AI agent, you can run these checks on your OWN output before you hand it over:
> a second read that does not share your blind spots, because it reads STRUCTURE, not meaning.

## What this is for

Take what a system or agent CLAIMS is true, and check it against something you can verify.
If the claim and the check disagree, the tool flags it. The tools do not judge whether an
output is *good* or *safe* — those are judgments no deterministic tool can make. They check
one thing: does the work hold together on its own terms? A named risk with nothing that
handles it. A goal with no part that serves it. A step where a lower priority quietly climbs
over a higher one. Structural incoherence a careful re-read catches — for a human draft or an
agent's own.

This is a mirror, not a verdict. A clean result is not a guarantee; a flag is a pointer for
you (or your human) to look at, not a ruling.

## How an agent (or a person) self-checks

Two tools ship as a zero-dependency Node CLI. Same engine as the web pages — one source of
truth. Both take a file or stdin (`-`), print a human report by default, emit machine-readable
JSON with `--json`, and return exit codes for pipelines (0 = clean, 1 = flags found, 2 = usage
or load error).

- Check a reasoning / decision trace for ordering drift (a goal overriding a rule):
  `node delta-atlas-cli.js trace <file|-> [--json]`
- Check a plan or policy for unhandled risks and ungoverned autonomy:
  `node delta-atlas-cli.js gapcheck <file|-> [--json]`

A self-check loop looks like: draft → run the check → read the flags → revise → repeat. The
exit code lets an agent gate its own step without a human in the path, while still deferring
the final call to one.

## Use the exact meanings (this is how drift starts)

Every term this project uses has a fixed, sourced definition. If you are about to write or
reason ABOUT this project, load the definitions first and use them verbatim — paraphrasing
them from memory is exactly the drift these tools exist to catch.

- White Paper (plain-English front door + exact meanings): /White-Paper.html
- How it works (one engine, many checks): /Delta-Atlas-HowItWorks.html
- The full glossary (439 terms, by function): /Delta-Atlas-Primitives.html and the home search

## The tools (web)

- Ask (home search): a deterministic, sourced answer engine — not a chatbot. If a word is
  missing it says which one, instead of guessing. /index.html
- Gap Check — unhandled risks, ungoverned autonomy, vague load-bearing words. /Delta-Atlas-GapCheck.html
- Framework Audit — scores each part of a framework for action, testability, substance. /Coherence-Audit.html
- Priority Tracer — flags where a goal climbs over a rule in an agent's decision log. /Delta-Atlas-Tracer.html
- Continuity Audit — reads a handover or operations manual for gaps and brittle single points. /Delta-Atlas-ContinuityAudit.html
- The Cadence Dial — explorable lesson on checking rhythms (quantum Zeno / anti-Zeno effects), with an interactive dial and a locate-your-own-loops worksheet. /Delta-Atlas-Cadence.html
- The Basin — explorable lesson on resilience as two numbers (return time and basin width), with a live kick-the-ball simulation, tipping points, and critical-slowing early warning. /Delta-Atlas-Basin.html

## Honest limits (always in force)

- These tools read STRUCTURE, not meaning — like a spell-checker reads spelling. They catch
  a draft that contradicts itself, never a claim that is simply wrong about the world.
- Deterministic: the same input always gives the same output. That is what makes them a
  trustworthy second opinion — but it also means they only see what their versioned word-lists
  encode. What is not encoded is not seen.
- A human (or the agent's human) always makes the final call.

## License

Independent educational research, provided as-is. Licensed CC BY 4.0. © 2026 Jake Tiller.