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

Shared vector composition and temporal replay

Registered 9 Oct 2026 · Snapshot 12 Oct 2026

Execution: completedScientific verdict: bounded temporal replay pass

The question

Use qualified acquired vectors and retained temporal context to retrieve and replay a continuing life independently of row order or observer timestamps.

Why this experiment?

Remembered content needs a role in current activity and a relation to earlier and later events. Supplied controls let those mechanisms be tested separately.

How it works

Supply image fragments and visibility masks to recall, then test working-state hold/release and temporal replay. The addresses, clocks and control commands are supplied; encountered images provide the content.

Follow this part of the investigation: Recall something, hold it, then recover what followed →
Implementation and comparison details

Continue the frozen 40-encounter life through six irregularly timed observations, context-driven replay, row permutation and context-swap controls, restart and a matched repeated external observation.

What happened

Bounded 1.5 passed: all 276 temporal choices were correct across original, shuffled and context-swapped storage; all 1,263 independent audit checks passed. Seven forward and six backward steps decoded actual retained working content. Full and restarted replay matched exactly, and subsequent acquisition matched the no-replay control. Qualification changed decoded pixels by at most 2.98e-8. Noisy cleanup remained ambiguous (margin 0.0159 < 0.02), so no prototype was substituted. The verified continuation contains 47 encounters and holds recalled episode 45. Full 1.3 and native visual movement remain open.

Measurements
Measurements — values preserved from this run
Recorded fieldValue
episodes47
extended episodes46
checks
{
  "intake": true,
  "source_denied": true,
  "temporal_selection": true,
  "replay_branches": true,
  "split_replay_exact": true,
  "split_post_observation_exact": true,
  "post_acquisition_exact": true,
  "post_cursor_exact": true
}

Results and visual artifacts

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

This test asks whether stored temporal relationships can guide a sequence of recollections. Cliste starts from a held memory and is asked to move forward or backward through its own encounter order. The pictures are decoded from the selected stored vectors. Encounter numbers identify observations; they are not class labels. Replay requests and temporal coordinates are supplied, so successful recall does not demonstrate a learned decision about what to remember or predict what happens next.

Remembering 7 encounters after a held memory
Remembering 7 encounters after a held memory
Read left to right, then continue on the next row. Starting from encounter 38, each supplied forward request retrieves the next memory in that direction; these panels show encounters 39–45. Each picture is reconstructed from the vector actually placed in working memory, with the original image unavailable to retrieval. Δ is the elapsed interval from the preceding recalled encounter, in 10 ms ticks; observer t is the recorded encounter time, added only for checking the answer. The order is the order Cliste encountered the inputs, not the order of events in a movie.

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

Remembering 6 encounters before a held memory
Remembering 6 encounters before a held memory
Read left to right, then continue on the next row. Starting from encounter 38, each supplied backward request retrieves the next memory in that direction; these panels show encounters 37–32. Each picture is reconstructed from the vector actually placed in working memory, with the original image unavailable to retrieval. Δ is the elapsed interval from the preceding recalled encounter, in 10 ms ticks; observer t is the recorded encounter time, added only for checking the answer. The order is the order Cliste encountered the inputs, not the order of events in a movie.

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

Adding and removing vector context preserves the remembered image
Adding and removing vector context preserves the remembered image
Left: decoding before adding structured context. Middle: decoding after nested role bindings are applied and then reversed for one retained image region. Right: four times the absolute pixel difference between those two decodings; black means unchanged. The largest measured difference is 2.98 × 10⁻⁸. This demonstrates reversible composition with supplied roles; it does not move an object within the scene or infer spatial relationships.

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

Does recall follow stored time relationships or array position?
Does recall follow stored time relationships or array position?
Each row counts correct next/previous choices, including requests at the ends of the stored sequence. Native means the original storage layout. Permuted means memory rows were shuffled while their content and temporal context stayed paired. Context swap exchanges the temporal associations of two encounters without swapping their visual content. All three conditions score 92/92: recall follows the temporal association, including the deliberately changed association, rather than a fixed row index.

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

Supplied roles, coordinate bases, temporal horizon and replay requests; no autonomous selection, spatial perception or replay learning benefit. Six irregularly timed natural observations extend a retained life. Full 1.3 remains open; final 32-image panel remains sealed.

Next question. Assess temporal replay and shared composition before planning further continuing-life work; full 1.3 requirements remain open.

Inspect the evidence

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