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

Preserving useful distinctions with compact perceptual memory

Registered 8 Oct 2026 · Snapshot 12 Oct 2026

Execution: completedScientific verdict: bounded storage screen completed

The question

Compare stricter reuse and quantized residual detail with matched-bit dense VSA storage, measuring retained distinctions and complete cost.

Why this experiment?

Reusable patterns could reduce repeated storage and support later operations. The relevant test is useful readback at an honest storage cost, not vocabulary size alone.

How it works

Acquire a bank from encountered visual features and compare storage and readback against simpler vector encodings. Pattern IDs are addresses of acquired vectors, not supplied object categories.

Follow this part of the investigation: Can experience share a vocabulary without losing useful distinctions? →
Implementation and comparison details

Ten fixed storage conditions reuse the unchanged codec and previously acquired banks. Sixteen fresh development photographs join the original twenty-four memories; supplied-visibility cues, failure traces and measured prefix costs distinguish storage effects from candidate-set expansion. No learning or final-input opening.

What happened

Eight-bit dense storage preserves every baseline retrieval choice on all three assays while reducing tensor payload from 616.75 to 176.71 MiB (71.35%). Eight-bit residuals improve numerical readback at higher cost without changing recall; four-bit alternatives introduce paired regressions. All 17,055 independent checks and 400 source-denied readbacks pass. Dense8 is the candidate for larger confirmation; the final 32 inputs remain sealed and integration is pending.

Measurements
Measurements — values preserved from this run
Recorded fieldValue
conditions10
fresh sources16
total sources40
eligible conditions3

Results and visual artifacts

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

This experiment compares ways to reduce stored vector size while preserving recall. Dense4 and dense8 store vector coordinates with 4- and 8-bit precision. Shared080/shared095 use prototypes accepted at cosine thresholds 0.80/0.95; r4/r8 additionally stores a 4-/8-bit residual, the difference from the prototype approximation. Tight002/tight0005 keep exact vectors whenever relative squared vector error would exceed 0.02/0.005. Reconstruction and recall are separate tests: an image can look similar yet select the wrong memory.

How much storage can be saved without changing recall?
How much storage can be saved without changing recall?
Rows are storage methods, aligned across all three panels. Left: memory use in MiB (1,048,576 bytes), including the entire prototype bank; the three bars distinguish tensor content, saved files, and saved files plus the shared model. Shorter is smaller. Middle and right: correct source recalls out of eight previously examined and 16 fresh development queries; darker blue is better. “Visible features only” excludes features affected by hidden pixels; “complete local view” uses the full field; “gray-filled occlusion” encodes hidden pixels as gray. Compare methods within the same cue and cohort. Equal totals can conceal different mistakes; these small development groups are not a final generalization test.

Original recorded figure ↗

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

How storage grows as more encounters are retained
How storage grows as more encounters are retained
The horizontal axis is the number of retained encounters. Left: total stored tensor bytes in MiB. Right: that cost divided by the uncompressed vector baseline at the same encounter count; below the dashed line at 1 means less storage. The entire acquired prototype bank is charged at every point, even before many memories reuse it. Lines connect measured counts only. Model files and saved-file overhead are excluded here and included separately in the storage comparison. Prototype numbers are cosine thresholds; residuals store differences from the prototype approximation. Max-error policies use exact vectors when the limit would be exceeded.

Original recorded figure ↗

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

4-bit vector storage: reconstructed photographs
4-bit vector storage: reconstructed photographs
Each row shows the original photograph, exact-vector reconstruction, the compressed reconstruction, then the error for each reconstruction. The first two sources caused recall failures in the earlier prototype-only test; that label does not mean the method shown here failed on them. The other two are the first fresh development sources. Error panels show the absolute RGB difference from the reference, multiplied by four and clipped for display: black means a match; brighter colors mean larger pixel differences. The scale is the same across panels. These fixed examples reveal visual detail; recall accuracy comes from the separate query test. Images keep their original proportions; resizing on this page is for display only.

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

8-bit vector storage: reconstructed photographs
8-bit vector storage: reconstructed photographs
Each row shows the original photograph, exact-vector reconstruction, the compressed reconstruction, then the error for each reconstruction. The first two sources caused recall failures in the earlier prototype-only test; that label does not mean the method shown here failed on them. The other two are the first fresh development sources. Error panels show the absolute RGB difference from the reference, multiplied by four and clipped for display: black means a match; brighter colors mean larger pixel differences. The scale is the same across panels. These fixed examples reveal visual detail; recall accuracy comes from the separate query test. Images keep their original proportions; resizing on this page is for display only.

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

Prototype reuse at 0.80 similarity plus 4-bit residuals: reconstructed photographs
Prototype reuse at 0.80 similarity plus 4-bit residuals: reconstructed photographs
Each row shows the original photograph, exact-vector reconstruction, the compressed reconstruction, then the error for each reconstruction. The first two sources caused recall failures in the earlier prototype-only test; that label does not mean the method shown here failed on them. The other two are the first fresh development sources. Error panels show the absolute RGB difference from the reference, multiplied by four and clipped for display: black means a match; brighter colors mean larger pixel differences. The scale is the same across panels. These fixed examples reveal visual detail; recall accuracy comes from the separate query test. Images keep their original proportions; resizing on this page is for display only.

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

Prototype reuse at 0.95 similarity plus 4-bit residuals: reconstructed photographs
Prototype reuse at 0.95 similarity plus 4-bit residuals: reconstructed photographs
Each row shows the original photograph, exact-vector reconstruction, the compressed reconstruction, then the error for each reconstruction. The first two sources caused recall failures in the earlier prototype-only test; that label does not mean the method shown here failed on them. The other two are the first fresh development sources. Error panels show the absolute RGB difference from the reference, multiplied by four and clipped for display: black means a match; brighter colors mean larger pixel differences. The scale is the same across panels. These fixed examples reveal visual detail; recall accuracy comes from the separate query test. Images keep their original proportions; resizing on this page is for display only.

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

Prototype reuse at 0.95 similarity without residuals: reconstructed photographs
Prototype reuse at 0.95 similarity without residuals: reconstructed photographs
Each row shows the original photograph, exact-vector reconstruction, the compressed reconstruction, then the error for each reconstruction. The first two sources caused recall failures in the earlier prototype-only test; that label does not mean the method shown here failed on them. The other two are the first fresh development sources. Error panels show the absolute RGB difference from the reference, multiplied by four and clipped for display: black means a match; brighter colors mean larger pixel differences. The scale is the same across panels. These fixed examples reveal visual detail; recall accuracy comes from the separate query test. Images keep their original proportions; resizing on this page is for display only.

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 component comparison uses fixed acquired banks and supplied visibility. Examined and fresh development queries are reported separately; repairing known failures is not general noninferiority. Count the entire bank, references, fallbacks, quantized coordinates, scales, envelopes, serialization and model costs. Residuals restore the original VSA space before existing operations. No candidate is installed, the final panel remains sealed, and milestone 1.3 integration remains open.

Next question. Assess whether retained detail improves the full cost/utility tradeoff over dense quantization before separately locking larger confirmation.

Inspect the evidence

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