run record~/research/cliste/runsAll Cliste runs

Type a section name and press Enter. Use “help” for the list.

Figures: softened for dark viewingDisplay only; image files are unchanged.
← All Cliste runs

Cliste experiment

Frozen encoder compatibility diagnostic

Registered 8 Oct 2026 · Snapshot 12 Oct 2026

Execution: completedScientific verdict: diagnostic complete

The question

Distinguish cross-version encoder mismatch from partial-cue matching within an updated representation, using a frozen two-by-two diagnostic.

Why this experiment?

A memory can only preserve what its sensory interface encodes. These studies separate encoder quality, decoder quality and compatibility with earlier stored vectors.

How it works

Change one component or input representation at a time and evaluate immediate reconstruction on separate images. Keep old-vector readback visible so gains on new inputs do not hide compatibility losses.

Follow this part of the investigation: Sharper perception can also change the language of memory →
Implementation and comparison details

Freeze both cue and bank encoder generations in four matched cells. Reuse exact parent cues, corruption mask, complete region geometry and original acquired memories; regenerated banks use original native pixels. No training, endpoint selection or checkpoint installation.

What happened

The frozen 2 × 2 diagnostic supports a cross-version matching cost. Re-encoding the comparison bank raises occluded-cue top-1 accuracy from 65.28% to 81.94% for ordinary continuation and from 76.04% to 81.25% for replay; the original baseline is 81.25%. Paired gains are +16.67 percentage points [95% interval 9.72, 24.31] and +5.21 [1.39, 9.72]. Within-version top-1 change remains unresolved, and correct-memory cosine margins are lower. All 38 cells and parent repeats pass; independent statistics checks pass 61,578/61,578 and bank replay 475/475. Counterfactual banks require original RGB. No memory migration, training or checkpoint adoption occurs. The worker took 240.0 seconds, peaked at 8.41 GiB CPU resident memory and 225.5 MiB PyTorch GPU allocation; each transient full bank uses 1.948 GiB.

Results and visual artifacts

Open a figure for its input, comparison, and result. Click an image to enlarge it.

The encoder turns a viewed image into vectors; the memory bank holds vectors created earlier. This diagnostic asks whether continued training makes new cue vectors incompatible with old memories. Each comparison crosses the original or updated cue encoder with the original or re-encoded bank. “Ordinary” trains on new images alone; “replay” adds rehearsal of old-vector decoding. Numbers 401–403 identify three training seeds, not image classes. Re-encoding the bank uses the original photographs, so it diagnoses compatibility but is not a way to migrate memory without source images.

Does a new encoder still match memories made by the old encoder?
Does a new encoder still match memories made by the old encoder?
Rows show training with or without replay across three seeds. Columns cross the original or updated cue encoder with memories encoded before or after training. Left uses complete local cues; right hides 20% of the same cue pixels. Each percentage is the fraction of 96 queries whose first-ranked answer is the correct source image; higher is better, with the same 0–100% scale in both panels. Compare the two updated-encoder columns to see the cost of retaining older vectors. Re-encoding the memory bank requires the original pixels and is a diagnostic, not a demonstrated migration procedure. The original-encoder/original-bank baseline repeats in every row. Complete-cue results using the same encoder for cues and memories mostly test exact self-matching, not general recognition.

Original recorded figure ↗

Replotted from saved outputs with clarified labels. The original recorded figure remains available.

Does a new encoder still match memories made by the old encoder? — original layout
Does a new encoder still match memories made by the old encoder? — original layout
Rows are the two training methods across three seeds. Columns cross the original/updated cue encoder with the original/updated memory bank. Left uses complete local cues; right masks 20% of the same cue pixels. Each percentage is the fraction of 96 queries whose first-ranked answer is the correct source image; higher is better. Compare “updated cue / original bank” with “updated cue / updated bank” to isolate the cost of mismatched encoder versions. The original/original baseline is repeated in every row. Complete-cue same-version results are mostly exact self-match checks, not broad recognition evidence. The original colorbar overlaps the last data column; use the corrected matrix for reading values.

Display derivative: WebP quality 90; maximum edge 2400 px. Original scientific pixels remain in the source run; do not measure this preview.

Training on new images only — first six partial-cue queries
Training on new images only — first six partial-cue queries
These are the first six queries in the fixed candidate order, chosen without looking at whether they succeed. Read across each row: the 80 × 80 cue with 20% of pixels hidden, the correct memory, then the four original/updated encoder-and-bank combinations labeled above the columns. Every answer is displayed using its original stored rendering, so image-decoder changes do not masquerade as a different retrieval choice. “Correct” means the expected source memory was selected. All examples use the prespecified display seed 402.

Display derivative: WebP quality 90; maximum edge 2400 px. Original scientific pixels remain in the source run; do not measure this preview.

Training on new images only — examples where re-encoding changes the answer
Training on new images only — examples where re-encoding changes the answer
These are the first six queries whose chosen memory changes when the updated encoder searches the updated bank instead of the old bank. They are selected for diagnostic contrast, not to estimate accuracy. Read across each row: the 80 × 80 cue with 20% of pixels hidden, the correct memory, then the four original/updated encoder-and-bank combinations labeled above the columns. Every answer is displayed using its original stored rendering, so image-decoder changes do not masquerade as a different retrieval choice. “Correct” means the expected source memory was selected. All examples use the prespecified display seed 402.

Display derivative: WebP quality 90; maximum edge 2400 px. Original scientific pixels remain in the source run; do not measure this preview.

Training with memory rehearsal — first six partial-cue queries
Training with memory rehearsal — first six partial-cue queries
These are the first six queries in the fixed candidate order, chosen without looking at whether they succeed. Read across each row: the 80 × 80 cue with 20% of pixels hidden, the correct memory, then the four original/updated encoder-and-bank combinations labeled above the columns. Every answer is displayed using its original stored rendering, so image-decoder changes do not masquerade as a different retrieval choice. “Correct” means the expected source memory was selected. All examples use the prespecified display seed 402.

Display derivative: WebP quality 90; maximum edge 2400 px. Original scientific pixels remain in the source run; do not measure this preview.

Training with memory rehearsal — examples where re-encoding changes the answer
Training with memory rehearsal — examples where re-encoding changes the answer
These are the first six queries whose chosen memory changes when the updated encoder searches the updated bank instead of the old bank. They are selected for diagnostic contrast, not to estimate accuracy. Read across each row: the 80 × 80 cue with 20% of pixels hidden, the correct memory, then the four original/updated encoder-and-bank combinations labeled above the columns. Every answer is displayed using its original stored rendering, so image-decoder changes do not masquerade as a different retrieval choice. “Correct” means the expected source memory was selected. All examples use the prespecified display seed 402.

Display derivative: WebP quality 90; maximum edge 2400 px. Original scientific pixels remain in the source run; do not measure this preview.

Limits and next question

This is a post-result diagnostic on 96 previously examined memories and one fixed 20% occlusion mask, not a fresh held-out test. Regenerating banks requires original pixels; it is a counterfactual comparison, not source-free memory migration. Geometry and cue selection are supplied structure. Retrieval is appearance-instance matching, not semantic cleanup. Intervals are conditional on the fixed fits and images; overlap with zero does not establish equivalence. Bank recovery does not prove an invertible or useful bridge, and failure does not prove information destruction. Report regional bank bytes, peak CPU/GPU memory, durable storage and elapsed time. Parent evidence and operative memories stay unchanged; no training or checkpoint adoption occurs.

Next question. Use the fixed within-version and cross-version contrasts to justify a later encoder intervention; further training requires its own plan and new final sources.

Inspect the evidence

Download the methods and measurements for this experiment. Large training and saved-state files are not included.