4P3X AI™ / INSIDE APEX™
A product story should show what improved, what failed and what remains to be tested. APEX is the spoken name of 4P3X AI™, not a second AI system.
Selected entries from the public build dataset, with their evidence boundaries intact.
SEPTEMBER 2026 / RECORDED RELEASE CHANGE
A September RWS revision changed a preference question from asking what helps someone learn to asking what feels comfortable today. The release record describes updates across six languages and 31 associated regression checks.
Explore Read • Write • SpeakSEPTEMBER 2026 / RECORDED DESIGN LESSON
Deaf-access work established transcript-first presentation, visual feedback and a non-voice route. The general accessibility lesson can be reused without transferring client-owned content.
Explore the access principlesSEPTEMBER 2026 / STANDALONE TESTBED
The agent-assurance testbed passed bundled tests but failed four reviewer-defined hardening checks. A revised standalone testbed subsequently passed 106 core checks. The private APEX runtime was not connected.
Read the failure and retestAUGUST 2026 / FEEDBACK RECORD
A voice workflow could pass automated routing checks while a human still found the delivery unnatural. Automated checks did not replace listening, judgement or clear attribution.
Read the evidence boundariesEvidence basis: SRC-INSIDE-1909. Historical results retain their original test scope.
The documents separate principles, specified components and demonstrated behaviour.
Approved project context, scoped tools and evidence guide the work. Confidence, knowledge and agreement do not themselves grant permission. Irreversible actions require the appropriate human authority.
The knowledge architecture calls for source dates, provenance, revision tracking and appropriate language or jurisdiction context. A library entry is not proof of complete knowledge or an implemented live update service.
Deterministic App controls, saved progress and approved local packs can continue without cloud access. A complete local reasoning runtime needs its own models, hardware, permissions and validation.
The documented loop is Forecast → Challenge → Resolve → Score → Calibrate → Learn. It calls for dated evidence, explicit uncertainty, outcome resolution and error tracking. No forecasting performance is claimed here.
The public site explains the approach. It does not expose the private runtime or client knowledge.
START WITH SOMETHING SPECIFIC
