Browser simulation · evidence-reading educational exercise
Can you prove the AI is stopped?
Choose the strongest claim that fictional evidence supports.
Can you prove the AI is stopped?Research
Articles and research from noqt on specialist models, security tools and control of AI systems.
Interactive exercises
Read the boundary before you trust the result.
Fixed synthetic browser educational exercise
Memory Control Lab
Review provenance, scope, correction, rejection and forgetting in one synthetic example.
Memory Control Lab· noqt Research
Lumi: a proposed frontier micro-model programme for security localisation
Lumi asks whether a small local specialist model can rank the source locations relevant to an existing Python security finding materially better than a cheap lexical baseline—by enough to justify the added complexity.
· noqt Research
What CVSS v4 Recovery does—and does not—represent
A synthetic CVSS v4.0 comparison holds the score at 8.7 while Recovery changes from R:U to R:I, then sets out what one fixed two-day NVD observation can—and cannot—show.
· noqt Research
95, 152, 3,611: three government AI transparency numbers that count three different things
It is tempting to rank the countries from those totals. That would be wrong. Each number counts a different object.
· noqt Research
A green dashboard can be stale
How to distinguish a fresh health report from a freshly scraped old one—and keep stale, missing, unreachable and unhealthy evidence separate.
· noqt Research
The AI shutdown claim card: six questions behind every “kill switch”
A source-by-source reference for separating proposed AI shutdown policy, bounded risk assessment, observed incident response and the evidence each source actually establishes.
· noqt Research
Can software physically break hardware? A field guide to what “damage” actually means
Software has permanently damaged specific hardware, but burned CPUs, flipped bits, locked FPGAs and “bricked” modems are four different claims.
· noqt Research
One session, two generated frames, one stored observation, then restart
Inspect one retained Zumi owner-path record: two simulated phone exposures, their hashes and provenance, and a stored technical observation retrieved from the same persisted session after restart.
· noqt Research
A stop request is not a containment result
How to read a local AI-workload containment receipt: identify the boundary, check the observed result, and keep the claim no broader than the test.