The gates, and how they map to the four layers

A gate is a hand-built, falsifiable rulebook. Not a model's opinion, not a
classifier, not a score. It decides one narrow question exactly, or it declines
to decide and says so.

There are now twenty-one of them, most published as public repositories anyone
can clone and run. This post sets out how they map onto the four architectural
layers, because that mapping is the part that was previously implicit.

The mapping

Inside DTL, the layer that asks whether an output is allowed, sit the
checkers that settle a specific kind of claim: dimensional analysis for physics
equations, chemistry formula and reaction balancing, statistical consistency
checks that catch impossible summary figures, structural sanity checking for
hardware description code, exact arithmetic, security weakness families, claim
extraction and routing, and documentation linting for unsupported claims.

Inside DAX, the layer that asks whether a physical action may execute, sits
the fail-closed permission boundary itself, along with early work on orbital,
flight-envelope and grid operational constraints. Two of those are research
scope and labelled as such.

Inside DCLA, the layer that tracks control changes, sit the open claim
taxonomy and verdict schema, and an application profile mapping lanes and
evidence onto specific regulatory articles.

Inside DELA, the layer that tracks evidence changes, sit scientific state
diffing with dependency impact and selective replay, pre-flight validation of
structured job metadata, confidence-contract routing for predicted structures,
provenance continuity checking, and replay of sealed evidence packs to detect
drift.

Why narrow is the point

Every gate on the list has a section stating what it does not check, and
that section is not a disclaimer. It is the specification.

Dimensional correctness is necessary but not sufficient: an equation can be
dimensionally valid and still wrong. A structural RTL check is not formal
verification and not timing sign-off. Statistical consistency finds arithmetic
impossibility, not fraud and not poor study design. Provenance continuity
verifies the chain, not the quality of what is at the end of it.

A checker that claims a broad competence is a checker whose failures are
unpredictable. A checker with a stated boundary can be trusted right up to that
boundary and no further, which is what makes it usable in a workflow where
somebody has to sign off.

Status, honestly labelled

Three labels, no fourth:

• Open source. A public repository with working code you can run today.
• Interactive lab. A runnable console, plus a repository where one exists.
• Research. Early scope, not for operational use, and it says so on its own
  page.

There is no "coming soon" on the gates index. A capability with nothing behind
it does not get an entry, because a page full of placeholders tells a reader
more about the company than a shorter page of real things would.

Test them

You do not need our permission, our source or our involvement. The audit
protocol is public: clone a gate, run it against your own cases, and if you
find where it is wrong, that is the most useful thing anyone can do for this
work.