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

Can Cliste recall an image after a small movement?

Execution: completedScientific verdict: Acquired-piece alignment search meets the prospective target; independent arrangement search fails preservation tolerances. Optional slow-recall candidate, with a concrete layout regression; no default promotion.

The question

Test whether looking at the same pixels at different sampling offsets helps Cliste reuse its acquired visual patterns when an image moves slightly.

Implementation and comparison details

Keep the learned pattern banks, visual encoder, decoder and memory operations fixed. Compare ordinary matching, one sampling offset with equally restricted visible support, and a search across 64 offsets. Each offset is ranked separately; favorable pieces from different views are never mixed into one match.

What happened

Completed 56 queries from 16 fresh sources. Acquired-piece search raised moved and partly hidden recall from 8/16 with ordinary pieces to 15/16 each. Moved writing improved from 2/8 to 8/8; direct visual recall was 4/8. The piece method met the prospective criteria, with one unchanged-photo regression and no grid-aligned loss. Independent arrangement search failed the preservation tolerances. All four state-preservation checks passed.

Recorded checks
Recorded checks — values preserved from this run
Recorded fieldValue
hidden pixels invariantPass
ordinary recurrence unchangedPass
retained state unchangedPass
vocabulary unchangedPass

Results and visual artifacts

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

Read each example as a memory-selection task: the query is what Cliste receives now; the expected original is the complete image stored earlier; the result is a selected memory decoded from its retained vectors. Gray regions are unavailable input. The matched single-alignment control uses the search scoring and restricted support, so comparison with search isolates trying more alignments. Correct means the original source image was selected, not that its meaning was understood. Four examples were fixed before evaluation; the extra writing and photograph failures are explicitly labeled as selected afterward.

How often did each method select the correct original?

Artifact pending or unavailable in this snapshot.

Final outcomes for all 56 queries: unchanged, moved, partly hidden and grid-aligned. Ordinary matching, a matched single-alignment control and 64-alignment search are separated. Small-movement results split writing from photographs. Search timing covers pieces and arrangements together; ordinary timing covers four existing methods. Counts refer to query outcomes, not independent sources or confidence scores.
Individual pieces: query, expected image and selected memories

Artifact pending or unavailable in this snapshot.

Four preregistered examples show the query, expected original and actual decoded choices from ordinary matching, the matched single-alignment control and search. An additional unchanged sled-dog photograph, selected after seeing the results, shows a real regression: ordinary and matched recall return the intact image, while search selects its exchanged-halves copy. Gray areas are unavailable pixels. Reversed sentences and exchanged halves are stored alternatives, not generated text or edits.
Arrangements: selected memories and an unchanged-input regression

Artifact pending or unavailable in this snapshot.

The same four preregistered examples compare arrangement recall under three methods. An additional unchanged writing image is explicitly labeled as selected after seeing the results: ordinary and matched recall select the correct memory, while search chooses another source's reversed sentence. Movement gains therefore appear alongside lost behavior. Every returned image is an actual stored-memory decode.

Limits and next question

The same sled-dog photograph selects its swapped-halves alternative in three conditions. Arrangement search adds other regressions. Direct visual recall still outperforms search on moved photographs (8/8 versus 7/8). Search reuses fixed learned content; it does not establish learned invariance, letters, reading, meaning or generation. The combined 64-view search averaged 15.8 seconds for writing and 59.8 seconds for photographs. Only 16 independent sources and 32 candidate memories were tested.

Next question. Connect acquired-piece search as an explicitly requested slow association operation alongside ordinary recall, preserve the layout regression, and measure its own cost. A spatial-consistency check may help choose between competing layouts; it has not been tested here. Milestone 1.3 remains open.

Inspect the evidence

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