← Delta Atlas

How it works

One engine, many checks. What every tool here is actually doing — and how much each check is worth.

Every tool here does the same thing. Take what a system or agent claims is true, and check it against something you can actually verify — something it can't just make up. If the claim and the check don't match, the tool flags it. We don't read the agent's mind and we don't police its wording; we check whether the story it tells matches something real. The subject changes — a software dependency, a reasoning trace, a handover manual, the electricity a chip draws — but the move never does.

One thing we say out loud that most of this field hides: not all checks are equal. A system that grades its own homework can look perfect and still be lying. So the tools sit on a ladder — the higher the rung, the harder the check is to fake, and the more a “pass” is actually worth.

Which check fits your job?

Auditing an agent's decision or reasoning logPriority Tracer → Reviewing an AI plan, policy, or system promptGap Check → Checking whether a framework's parts hold upFramework Audit → Reviewing a handover or operations manualContinuity Audit → Just want a word explainedAsk / glossary →

Or, if you're an AI agent (or telling one what to do): run these checks from a pipeline on your own output — see llms.txt.

The ladder — how much a check is worth

The document against itself Rung 1 · weakest The check stays inside one text: does the part that names a risk also name something that handles it? These catch the everyday failure — a plan with holes — but they share one blind spot: a careful lie passes. A document that names controls it doesn't really have will still read clean. Cheap, useful, and honestly the weakest rung. Gap Check · Framework Audit · Priority Tracer · Continuity Audit
One thing against a separate thing Rung 2 · stronger The check is a second item, made at a different time. Today's server contract against a fingerprint you approved last week. The day you planned against the day you closed — locked in a chain so neither can be quietly edited after the fact. Now a lie has to fake both, which is much harder — and if someone tampers, the chain breaks right at the spot they touched. day-ledger · mcp-manifest-pin
Against the physical world Rung 3 · the frontier The check comes from outside the software entirely: the power a chip actually drew, or a hardware seal proving what actually ran. This is the rung a lying agent can't beat by writing a tidier log — as long as the check sits on something the agent can't reach. We've tested the power end and we'll be straight about it: it reliably catches “is it working, or just claiming to be,” but a single sensor is only as trustworthy as the wire it sits on. The real top of this rung is a check rooted in trusted hardware. We're building toward it — it isn't done, and we won't pretend it is. Physical footprint experimental

Where each tool stands

ToolRungWhat it claimsWhat it's checked against
Gap Checkitselfthe risks a document nameswhether that same document names a fix for each
Framework Audititselfa stated goalwhether the document's own parts have real substance
Priority Traceritselfthe priority a trace says it holdsthe order that same trace actually follows
Continuity Audititselfwhat a handover manual promiseswhether the same manual holds the backups it needs
day-ledgerseparatethe day you plannedthe locked, signed day you closed
mcp-manifest-pinseparatethe tool contract a server serves todaythe fingerprint you approved earlier
Physical footprint expphysicala log claiming the machine is idlethe real power the chip drew
The rule underneath all of it. We don't check for good; we check for true. “Safe” is a judgment no meter can read. “This machine ran exactly what it said it ran” is something you can hold up and check. We only measure the second kind — and for every tool, we tell you how strong that check really is. That last part is the honest half most of the field leaves out. One more thing, plainly: a check is only as good as how independent it is. Three numbers from the same tool aren't three checks — they're one check wearing three hats. A real second check comes from a different place that could fail a different way.
This page's own little ledger (it practices what it preaches):
DateChangeReason
2026-07-05v1: the “one engine, many checks” explainer — the ladder, the tool map, the underlying rule.To give the whole toolset one plain-English thesis a first-time reader gets on sight, without jargon.
(next)(update whenever a tool moves rungs or a new check is added)