A2b feasibility: G0/G1 endpoint-recovery probe for input-level re-embedding

[samscarrow]

Scope — feasibility only (no experiment, no detector changes)

Narrow gate for the A2b input-level re-embed fidelity study (follow-on to #88). This PR adds only the G0/G1 endpoint-recovery probe and the design scope. It does not build ReembedDriftStream, run the subset experiment, or touch forward(), associative_core.py, the detector, or any existing analyzer.

What this proves

The faithful input-level arm needs to map each cached ImageNet-R ViT-L/14 row back to its exact source image and re-embed it, sharing endpoints with the cache-linear arm. G0/G1 de-risks that before any harness is written:

cache row i -> ImageFolder sample train_indices[i] -> exact image path -> fresh DINOv2 embedding -> cosine(cached, fresh) > 0.999

Verdict: PASS (gentoo, CUDA, train split):

Check	Result
Label-match rate	100% (35/35)
cosine(cached, fresh) min/median/mean/max	1.0 / 1.0 / 1.0 / 1.0
count cos > 0.999	35/35
Rows below threshold	none
Split replay length vs cache rows	23918 == 23918
Cache features raw vs unit-normalized	raw (L2 norm 39.17–48.67), confirmed not unit

Sample: classes {166, 63, 77, 156} + attractor 134 + 5 out-of-class negatives, seed 0. Split/transform/model taken from extract_imagenet_r_vitl14 itself (not reimplemented from memory). Cosines are exactly 1.0 — the row↔image mapping is bit-exact (deterministic encoder, shuffle=False cache).

Consequence for the next step

Because the cache stores raw (~45-norm) features, the input-level arm must feed raw DINOv2 output to the engine (engine normalizes internally, as the text RealDriftStream does); unit-normalize only for geometry diagnostics. The next risk is no longer endpoint identity — it is whether input-level interpolation yields a meaningful drift path vs. visual artifacts / a degenerate trajectory.

Files

• probe_g0g1_reembed.py — read-only probe over the merged feature_cache_inr_vitl14 cache
• results/issue_input_reembed_fidelity/G0_G1_PROBE.md — verdict + run metadata
• results/issue_input_reembed_fidelity/SCOPE.md — A2b paired-experiment design (diagnostic / not-production framing preserved)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2b6d7a3a4d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Guard split-length mismatches before indexing

When the replayed split is shorter than the cache (one of the G0 failure modes this probe is meant to diagnose), the script still samples cache rows and indexes split_indices[i] before it reaches the later split_len_ok verdict logic. In that environment the probe raises IndexError instead of writing the intended FAIL report, so the length check needs to abort/report before any per-row recovery is attempted.

Useful? React with 👍 / 👎.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.