Skip to content

Improve xi_robustness_v2 (~0.23 → >0.5) without touching the frozen chirality axis #327

Description

@NickFlach

Source: research note experiments/notes/2026-06-09T01-xi-true-robustness-uuid-fix.md (from closed PR #216).

Finding: xi_robustness_v2's genuine adversarial robustness is ≈0.2326 (3-trial deterministic 0.14789x). The earlier ~0.56 average was a UUID-ordering artifact — the mean of adversarials attacking successfully (xi≈0.23) and adversarials self-disrupting (xi→1.0), an illusion of robustness. Deterministic adversarial UUIDs are now set at build time in build_adversarial_set_l5, so xi is stable and the 0.23 is real.

Content-based chirality (Attempt A in the note) was catastrophic (carrier_e→0.000, transfer→0.379) and is a frozen axis — do not modify stage_chiral_perturbation sort.

Build-on directions to raise xi 0.23→>0.5:

  1. Amplitude pruning during dreaming — prune high-amplitude memories lacking phase coherence with their cluster (adversarials sit slightly off-phase).
  2. Frequency-band enforcement — adversarials are stored at frequency=0.1 yet attack the attention band; stricter band filtering could isolate them.
  3. Phase authentication — corpus memories have deterministic encoder-derived initial phases; adversarials use random phases. A phase-consistency check could reject them.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions