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 edited working content support another operation after new experience?

Registered 10 Oct 2026 · Snapshot 12 Oct 2026

Execution: completedScientific verdict: bounded working sequence candidate

The question

Use the frozen learned readout for two overlapping exchanges, holding the first change through fresh incoming photographs and a restart while preserving original memories.

Why this experiment?

A useful construction must remain usable after it changes. The second operation should depend on the edited state, including after new experience and a restart.

How it works

Exchange positions A and B, hold the result while four more photographs arrive, save and reload it, then exchange B and C. Compare reading the updated state with reading the original or another photograph, and verify exact undo.

Follow this part of the investigation: Use changed working content after new experience →
Implementation and comparison details

Carry a first learned exchange through held native intake, persist actual edited work, then compare a second operation with stale and shuffled-content controls.

What happened

Held edited content supported useful second operations after new experience and restart under the prospective provisional tolerances. Supplied requests, errors and the experimental persistence boundary remain explicit.

Measurements
Measurements — values preserved from this run
Recorded fieldValue
operations32
sources16

Results and visual artifacts

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

This public record includes 4 of 13 saved figures: summary comparisons and fixed examples, including every figure referenced by the guide. Reported aggregates include every case; individual records and complete outputs remain in the research archive.

The complete task and its four intervening photographs
The complete task and its four intervening photographs
Recall a supplied region, exchange A with B, hold that edited work during four ordinary admissions, reload it and exchange B with C. The overlap makes the second request depend on the first. Acquired memory remains protected; this does not test distractors superposed into working content or autonomous object selection.

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

Useful work after two exchanges: mean error and every request
Useful work after two exchanges: mean error and every request
Final conditions share the same two-operation target. Read-original and read-donor controls choose the second exchange from different content but apply it to the actual edited workspace. Separate first-step and marginal second-step counts distinguish a good initial edit from useful continued work.

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

First photograph: central region, both requested exchanges
First photograph: central region, both requested exchanges
Fixed example chosen independently of outcome. The feature grid labels the supplied A/B/C addresses, and the crops separate original memory, the exact first and final reference, and actual working content. References use original features for scoring, not operative selection.

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

Bright window region — unchanged
Bright window region — unchanged
The fixed comparison crop lies on a bright window, marked on the original-photo inset. The white panels are real scene content, not missing images. Neither exchange changed the representation; this request is unchanged, not an improvement. The example remains because it was selected before the outcome.

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

32 supplied feature-exchange sequences from 16 fresh photographs, four shared incoming photographs, one frozen codec/readout/bank. Protected working storage is not neural distraction resistance. Edited-work persistence is experimental; no default-runtime promotion or learned task selection.

Next question. Use measured cumulative utility, marginal second-operation effects and visible errors to choose the next bounded downstream task or a targeted optimization round.

Inspect the evidence

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