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

1.1 — one continuing perceptual life

Registered 7 Oct 2026 · Snapshot 12 Oct 2026

Execution: completedScientific verdict: pass

The question

Test one continuing perceptual life whose retained hyperdimensional VSA vectors supply readout.

Why this experiment?

Before testing reasoning or understanding, a continuing system needs to admit experience, retain it and recover it after a restart.

How it works

Admit natural images into VSA memory, add more experience, then decode earlier vectors and test restoration from a saved state. The earliest image interface used small center crops, so it did not retain the complete scene.

Follow this part of the investigation: Can an experience survive after its image file is gone? →
Implementation and comparison details

First connected reset runtime: frozen perceptual sensor/painter, raw VSA episode memory, represented working origin and persistent temporal state.

What happened

All registered continuity, isolation and retained-content control checks passed for this run.

All registered continuity, isolation and retained-content control checks passed for this run.

Measurements
Measurements — values preserved from this run
Recorded fieldValue
images12
initial images8
correct mean mse first80.006882880718798252
zero mean mse first80.12917698406583084
cyclic mean mse first80.15215120234326995
correct mean mse all120.00645667602327277
restart max abs0
earlier readout max abs after intake0
max cyclic donor error0
min cyclic own change0.7486177682876587
individual zero better episodes
[]
individual cyclic better episodes
[]
resources
{
  "peak_rss_bytes": 1523208192,
  "peak_cuda_allocated_bytes": 268006400,
  "peak_cuda_reserved_bytes": 283115520
}
final storage
{
  "logical_content_bytes": 3145728,
  "logical_context_bytes": 196608,
  "allocated_content_bytes": 16777216,
  "allocated_context_bytes": 1048576,
  "peak_rss_bytes": 1523208192,
  "peak_cuda_allocated_bytes": 268006400,
  "peak_cuda_reserved_bytes": 283115520
}
model file bytes59884597
stage timings
{
  "initial": {
    "birth_or_load_seconds": 0.6461686340626329,
    "readback_seconds": 0.17589149402920157,
    "home_bytes": 66730385,
    "snapshot_bytes": 6844697
  },
  "isolated8": {
    "readback_seconds": 1.1184880989603698
  },
  "continued": {
    "birth_or_load_seconds": 0.406919750967063,
    "readback_seconds": 0.2398141650483012,
    "home_bytes": 70072727,
    "snapshot_bytes": 10187039
  },
  "isolated12": {
    "readback_seconds": 0.6801737960195169
  },
  "uninterrupted": {
    "birth_or_load_seconds": 0.4016055980464443,
    "readback_seconds": 0.24009965208824724
  }
}
intake events
[
  {
    "episode": 0,
    "intake_seconds": 0.4607924920273945,
    "snapshot_seconds": 0.049857852049171925,
    "snapshot": "20261007T172041258756Z-000002"
  },
  {
    "episode": 1,
    "intake_seconds": 0.004848994896747172,
    "snapshot_seconds": 0.04911179095506668,
    "snapshot": "20261007T172041312921Z-000003"
  },
  {
    "episode": 2,
    "intake_seconds": 0.004833713988773525,
    "snapshot_seconds": 0.048958671977743506,
    "snapshot": "20261007T172041363475Z-000004"
  },
  {
    "episode": 3,
    "intake_seconds": 0.005843402002938092,
    "snapshot_seconds": 0.050874009961262345,
    "snapshot": "20261007T172041418396Z-000005"
  },
  {
    "episode": 4,
    "intake_seconds": 0.0045759910717606544,
    "snapshot_seconds": 0.047563432017341256,
    "snapshot": "20261007T172041474629Z-000006"
  },
  {
    "episode": 5,
    "intake_seconds": 0.005855572992004454,
    "snapshot_seconds": 0.049404495977796614,
    "snapshot": "20261007T172041529450Z-000007"
  },
  {
    "episode": 6,
    "intake_seconds": 0.005305001977831125,
    "snapshot_seconds": 0.0484576829476282,
    "snapshot": "20261007T172041583677Z-000008"
  },
  {
    "episode": 7,
    "intake_seconds": 0.00510448997374624,
    "snapshot_seconds": 0.05035006103571504,
    "snapshot": "20261007T172041637398Z-000009"
  },
  {
    "episode": 8,
    "intake_seconds": 0.3320482469862327,
    "snapshot_seconds": 0.055816140957176685,
    "snapshot": "20261007T172048462164Z-000010"
  },
  {
    "episode": 9,
    "intake_seconds": 0.006025842041708529,
    "snapshot_seconds": 0.053806078038178384,
    "snapshot": "20261007T172048521800Z-000011"
  },
  {
    "episode": 10,
    "intake_seconds": 0.005324792000465095,
    "snapshot_seconds": 0.052781939040869474,
    "snapshot": "20261007T172048579986Z-000012"
  },
  {
    "episode": 11,
    "intake_seconds": 0.004652714007534087,
    "snapshot_seconds": 0.05426323297433555,
    "snapshot": "20261007T172048637365Z-000013"
  }
]
effective batch limit4
actual images per call1
Recorded checks
Recorded checks — values preserved from this run
Recorded fieldValue
initial:recall and render state contractPass
initial:nonworking unchangedPass
initial:frozen modelPass
initial:earlier rows unchangedPass
initial:exact save load statePass
initial:restored model exactPass
isolated8:recall and render state contractPass
isolated8:nonworking unchangedPass
isolated8:zero recall and render state contractPass
isolated8:zero nonworking unchangedPass
isolated8:rolled recall and render state contractPass
isolated8:rolled nonworking unchangedPass
isolated8:empty refuses recallPass
isolated8:sources inaccessiblePass
isolated8:observer inaccessiblePass
isolated8:frozen modelPass
continued:recall and render state contractPass
continued:nonworking unchangedPass
continued:frozen modelPass
continued:earlier rows unchangedPass
continued:exact save load statePass
continued:restored model exactPass
isolated12:recall and render state contractPass
isolated12:nonworking unchangedPass
isolated12:sources inaccessiblePass
isolated12:observer inaccessiblePass
isolated12:frozen modelPass
uninterrupted:recall and render state contractPass
uninterrupted:nonworking unchangedPass
uninterrupted:frozen modelPass
uninterrupted:earlier rows unchangedPass
isolated8:all registered source probes deniedPass
isolated12:all registered source probes deniedPass
source home unchanged by initial controlsPass
source home unchanged by final readbackPass
initial state exact in isolated restartPass
continued state exact in isolated restartPass
continued equals uninterrupted statePass
initial restart pixelsPass
continued restart pixelsPass
earlier readouts after later intakePass
each cyclic readout matches donorPass
each cyclic readout changes ownPass
correct mean mse below zeroPass
correct mean mse below cyclicPass
model exact across all processesPass
snapshots contain only declared vectorsPass
earlier raw rows and contexts exactPass

Results and visual artifacts

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

The experiment stores photographs as VSA vectors and reconstructs them before and after a restart that cannot access the original images. Two controls erase the vector or substitute another image’s vector. A substituted vector should reconstruct its donor image: this tests whether image content actually comes from the stored representation. Inputs are 64 × 64 center crops; matching before/after readbacks establishes persistence, not perfect reconstruction or image recognition.

Observation 0: retaining a cat behind a curtain across restart
Observation 0: retaining a cat behind a curtain across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 1 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 1: retaining a street scene with a fire hydrant across restart
Observation 1: retaining a street scene with a fire hydrant across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 2 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 2: retaining a surfer on a wave across restart
Observation 2: retaining a surfer on a wave across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 3 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 3: retaining a food and drinks on a table across restart
Observation 3: retaining a food and drinks on a table across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 4 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 4: retaining a stop sign and building across restart
Observation 4: retaining a stop sign and building across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 5 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 5: retaining a televisions above a restaurant table across restart
Observation 5: retaining a televisions above a restaurant table across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 6 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 6: retaining a people with bicycles across restart
Observation 6: retaining a people with bicycles across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 7 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 7: retaining a bedside scene across restart
Observation 7: retaining a bedside scene across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last column intentionally uses observation 0 rather than this image; it should resemble that donor, demonstrating that the vector supplies the content. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 8: retaining a people around a table across restart
Observation 8: retaining a people around a table across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last two controls were planned only for the first eight photographs and are correctly marked “Not run” here. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 9: retaining a people near a doorway across restart
Observation 9: retaining a people near a doorway across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last two controls were planned only for the first eight photographs and are correctly marked “Not run” here. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 10: retaining a pumpkin on steps across restart
Observation 10: retaining a pumpkin on steps across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last two controls were planned only for the first eight photographs and are correctly marked “Not run” here. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

Observation 11: retaining a ocean wave across restart
Observation 11: retaining a ocean wave across restart
Left to right: the admitted 64 × 64 crop, reconstruction before restart, reconstruction after restarting without source-image access, reconstruction after erasing the memory vector, and reconstruction after substituting a different image’s vector. The last two controls were planned only for the first eight photographs and are correctly marked “Not run” here. Before/after restart outputs match exactly in the recorded numerical check; blur relative to the input remains. These are administrative memory readbacks, not image recognition from a cue.

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

One frozen pretrained checkpoint and twelve COCO validation photos, newly encountered by this life but previously eligible for model evaluation. Scores concern the admitted center-cropped 64 × 64 RGB image. Episode ordinals are administrative readback. This does not establish partial-cue retrieval, acquired cleanup, conceptual composition, scene editing, perceptual understanding or broad generalization.

Next question. If the registered continuity and content controls pass, extend this same life to image collections, video and pauses in 1.2.

Inspect the evidence

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