Adaptive education
Change subject, language or learning level and follow the proposed teaching adjustment. The learner’s next step stays within the authorised learning scope.
CORE AI & SAFETY / PROTOTYPE
4P3X Guardian AI™
Executable local prototypes and simulation research for agent control, containment, verification and auditability.
New Guardian film in preparation
The replacement film has not been created yet. The five interactive demonstrations are available to explore now.
Follow APEX and Guardian through learning, missing context, compromised-agent containment, reviewed updates, Laws 0–4 and evidence of the build.
See APEX and Guardian working together
Five interactive scenarios make context, permission, reviewed learning and evidence visible. These are bounded demonstrations.
Rogue or compromised agents
Inspect a suspicious action, the containment decision and the human review. The scenario explains the intended control path.
Reviewed learning and updates
Follow a proposed lesson through review, approval and a controlled update. Inspect rollback and retained evidence rather than allowing an agent to promote its own change.
Laws 0–4 in decisions
Use concrete decision cases to see which constraint changes the outcome. Agreement, confidence and personal preference do not create authority.
Proof of the build
Separate the artefact, the simulated behaviour, the recorded check and the remaining acceptance step. The guided tour and architecture map explain how the parts relate.
Human control is a system requirement.
4P3X Guardian AI™ — Prototype / R&D. A developing safety and assurance architecture for tool-using and multi-agent systems.
Permission and scope
Designed to check what an agent is authorised to do, detect unexpected action chains and preserve human approval.
Containment and override
Explore how a compromised, rogue or unsafe agent can be isolated, stopped and reviewed in a controlled simulation.
Reviewable evidence
Test trace and provenance integrity, audit trails, failures and false positives. Record the repair and repeat the scenario.
The next proof point
Independent security review, measurable benchmarks and production hardening are still required.
A contained pilot
Start with a representative non-production workflow, synthetic or explicitly approved data and agreed success criteria.
Report the misses
Adversarial testing is useful only when detection gaps, failed controls and unresolved risks remain visible.
Current limits
Guardian is not represented as a certified security product or approved critical-infrastructure system. Protection is not guaranteed.
Beyond the architecture: executable local work
The reviewed standalone candidate is v1.3.0. Production remains on hold.
Standalone prototype and simulation lab
Versioned source, executable local services, verification tools and scenario work provide concrete development evidence.
Recorded checks and retained failures
Core, transport, witness and verifier checks are recorded within local scope. Repair/retest campaigns preserve earlier failures. Overlapping test groups are not added into a misleading total.
Four separate development tracks
Deterministic controls, governed integration, learned-model research and adversarial assurance progress separately. A passing control test does not establish complete learned intelligence.
What the next review must resolve
Local success is one stage in the assurance work.
Containment and owner control
The case study retains an availability limit affecting timely owner stop, plus untested authority-separation boundaries.
External and integrated acceptance
Independently administered off-host witnessing, rendered browser/mobile testing and real confined-agent integration remain outstanding.