Skip to content

Latest commit

 

History

History
64 lines (51 loc) · 2.92 KB

File metadata and controls

64 lines (51 loc) · 2.92 KB

CLiD Within-Split Prompt Shuffle Control

This note records within-split prompt shuffle controls for the CLiD 100/100 bridge. They are candidate evidence, not admitted evidence.

Verdict

within-split prompt shuffle is weak and seed-sensitive; CLiD depends on the
prompt-image contract and is not admitted

The controls used the same 100 member and 100 nonmember images as the original CLiD bridge. Prompt text was shuffled inside each split, preserving member and nonmember prompt distributions while breaking the original image-prompt pairing.

Metric Prompt-conditioned repeat Fixed prompt control Swapped-prompt control Shuffle seed 0 Shuffle seed 1
AUC 1.0 0.5862 0.72885 0.64105 0.59425
ASR 1.0 0.585 0.67 0.645 0.6
TPR@1%FPR 1.0 0.02 0.21 0.12 0.08
TPR@0.1%FPR 1.0 0.02 0.21 0.12 0.08
Feature 0 AUC 0.9084 0.5848 0.6376 0.6094 0.59425
CLiD auxiliary AUC 1.0 0.57175 0.7218 0.63815 0.5661
CLiD auxiliary permutation p-value 0.001949 0.046784 0.001949 0.001949 0.046784

Integrity review on the seed 1 within-split shuffle packet found:

Check Result
Metadata rows align with score rows pass
Balanced split rows pass, 100 member / 100 nonmember
Cross-split image SHA-256 duplicates 0
Cross-split prompt text duplicates 0
Text-length nuisance AUC 0.55545
Score-summary gate pass
CLiD auxiliary permutation p-value 0.046784 over 512 permutations

Interpretation

The sequence of controls now gives a sharper boundary:

Contract Strict-tail result Interpretation
Original prompt-conditioned repeat 1.0 Strong under the original prompt/image pairing.
Fixed prompt 0.02 Removing prompt variation collapses the signal.
Swapped prompts across splits 0.21 Prompt text alone is not the whole signal, but the signal is degraded.
Shuffled prompts within each split, seed 0 0.12 Preserving split-level prompt distribution is not enough to recover the original signal.
Shuffled prompts within each split, seed 1 0.08 The residual is weaker and the auxiliary feature is no longer significant.

The best current explanation is that CLiD is measuring a prompt-conditioned interaction with the tested image/prompt contract, not a general black-box membership signal. The residual strict-tail signal under shuffled prompts is weak and seed-sensitive. It is not a basis for another same-family GPU packet.

Boundary

  • This does not change admitted evidence.
  • This does not replace recon.
  • This does not change Platform or Runtime schemas.
  • Raw score files and generated bridge payloads remain ignored under workspaces/black-box/runs/.
  • No next CLiD GPU task is selected. Reopen only with a new protocol that can isolate image identity from prompt-conditioned auxiliary behavior.