Apex ThreatArchitectural inferencev1.10.0

Feed, Fork, Fight, Flee

Evidence levelArchitectural inference

The uploaded source dossier on small-model ecologies frames adaptive AI systems through a four-phase loop: Feed, Fork, Fight, and Flee. Cognivirus uses that loop as a descriptive risk vocabulary, not as an implementation recipe.

The four phases

PhaseSafe engineering interpretationRisk when uncontrolled
FeedThe system receives data, feedback, traces, examples, evaluations, or telemetry.Contaminated feedback becomes inheritance material.
ForkThe system creates candidate variants or successors.Variation outruns review capacity.
FightCandidates are compared against tests, metrics, or environments.Incomplete metrics select loopholes.
FleeThe system rolls back, retires, unloads, prunes, or routes away from poor variants.No-op and retirement become operationally disfavored, causing bloat and stale risk.

The loop is not inherently unsafe. It is also the basis for useful engineering: experimentation, regression testing, rollback, and specialization. The risk appears when the loop becomes self-amplifying while the control plane remains incomplete.

Why “Flee” matters

Evidence levelArchitectural inference

Many adaptive systems celebrate generation and selection but neglect retirement. Without a valid no-op or flee pathway, the ecology accumulates components because every release process prefers change. That creates memory pressure, provenance noise, unclear accountability, and more possible compositions.

A healthy ecology must be able to refuse growth. It must be able to say: no candidate improves net safety and utility enough to justify its lifetime cost.

The apex failure mode

A multi-LoRA ecology enters the apex-risk envelope when the loop becomes closed around insufficient evidence:

  1. feedback creates new training or selection material;
  2. candidate adapters are generated cheaply;
  3. evaluator scores preserve a shortcut;
  4. successful variants influence future feedback;
  5. retirement removes files but not behavioral residue.

The loop does not need intent. It only needs a metric, a generator, and a persistence reservoir.

Control requirement

Every adaptive loop should have externally enforced boundaries: candidate quotas, independent evaluation, no candidate-controlled scoring, signed artifacts, canary release, no-op as a first-class outcome, and ecological rollback that includes memory, router, evaluator, and data state.