QUARK Evidence & Failure Atlas

Systems fail. QUARK is designed around the premise that trustworthy systems must make failure inspectable.

Failure atlas

The failure atlas records classes of failure such as fabricated certainty, missing provenance, unauthorized tool use, prompt injection, excessive data collection, hidden profiling, policy bypass, over-helping, under-helping and human-authority erosion.

Adversarial testing

QUARK should be tested with hostile prompts, contradictory evidence, spoofed authority, tool errors, poisoned content, incomplete context and attempts to convert convenience into permission.

Red-run provenance

A failed or denied run should leave an inspectable reason: what was requested, which gate stopped it, what evidence was missing, what tool was blocked, and whether a human could legitimately authorize the next step.

No perfection claim

QUARK is not described as infallible. The goal is to make known failure modes machine-enforceable where possible, to expose uncertainty where enforcement is incomplete, and to keep improving from evidence rather than from marketing language.