The instruments — one rule: prove it, or say nothing.

One platform, one reader for each kind of AI in your company: models you download, models someone trained further, code your AI writes, and agents that act for you. Each block below says what the reader looks at, what it catches, and — honestly — how finished it is today.

These four watch one river. Two more instruments complete the house: Proofora, the machine-checked proof under every reading, and Solora, the keystone the whole house is built and watched from.

In one breathOne reader per thing you run — each proves what it claims, or says nothing at all.

Mid-water twilight: the last light from the surface fades into warm darkness around a small ember glow. — **fig. 03** mid-water — the light thins, the reading begins

Reads · Modelsstation −250 m

The evaluator — what a model really is

early · works on popular families

Know what a downloaded model really is — before you trust it.

It reads the model straight from the file and reports, in plain grades, what it tends to do and exactly what a stranger's changes did to it. It answers "what is this?" — never "is it safe?" (that's the curator's job, below). And when it can't see something clearly, it says so plainly instead of guessing.

browse the public model atlas — ardora.vulcora.se →

See a real reading — live

Don't take the shape of a reading on faith. Look up any model and read its real verdict — the plain grades, the changes a stranger made, and exactly what we couldn't see clearly, marked empty. Live, in your browser.

the scanner →

Cures · Modelsstation −400 m

The curator — audits modified models

our strongest · try it today

Find the hidden backdoor off-the-shelf scanners miss — and cut it out.

Our strongest product. It reads a modified model straight from the file — without ever making it say anything harmful — and finds what a stranger's extra training really did, including what it hid. You get signed proof you can re-check yourself, plus an optional repair with an undo button. You can try it today; the always-on version is coming next.

the trust product — protora.vulcora.se →

Watch the real catch

On a sealed challenge cohort the planted structure is real and the reversible excise is certified — flagged, proven, and re-runnable by a stranger. Watch the trigger fire on a model anyone can load, then watch the repair pass the very proof that found it.

the live challenge →

Defends · Codestation −650 m

The immune system — code security with proof

real · still growing

Catch the exploit your scanner can't prove — even in AI-written code.

It checks code the way an attacker would. Every real hole comes with a tiny program that shows the attack working — no vague "maybe" alerts. After you fix it, that same proof runs again, and the issue closes only when the attack stops working. Strong on the languages it supports today, honest about the rest — and until it opens to the public, we won't draw you a verdict you can't yet run.

Holds · Agentsstation −850 m

The nervous system — agent accountability

we use it daily · not for sale yet

Know your AI agents kept their word.

Every agent states what it's about to do before it acts; the outcome is then graded against that promise, and both are kept in a record nobody can edit. Humans and AI agents, held to the same standard. We run our own company on it today — it's not for sale yet, and we won't hand you a screenshot standing in for a record you can't inspect.

start with one model — send it in, get back a signed reading · legend@vulcora.se

Descend

Layer 4: the discipline — how attacks get in, and how we keep score in public.

Below the products lies the method: the ways a backdoor gets planted, the public tests we grade ourselves on, and how we watch the whole world of downloadable AI.

Descend to Layer 4↓ −4 000 m Up to Layer 2↑ −200 m