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 training detect losses before they become the new baseline?

Execution: completedScientific verdict: Conclusion reported in the findings and result narrative

The question

Qualify resumable joint visual training with scheduled checks of incoming perception, older memories, ordinary recall and working use.

Implementation and comparison details

Joint training now pauses for checks against original memories, with complete saved optimizer and sampling state.

What happened

The bounded CUDA qualification completes checks at zero, 50, 250 and 500 updates and recovers an interrupted evaluation from saved state. Incoming-image MSE improves 1.39%; six old-vector readbacks worsen 0.17%. Obscured first-choice accuracy remains 78/96 through two losses and two gains. Ordinary recall and working decisions remain unchanged. Localized writing checks expose changes hidden by whole-image scores. No candidate is adopted.

Results and visual artifacts

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

Compare each checkpoint with the same original and fixed acquisition rendering. Writing and surrounding photographs have separate scores. This public selection contains 3 of 80 figures. Aggregate results cover the full measured panel; individual records and remaining figures stay in the research archive.

500 updates · Incoming image 1
500 updates · Incoming image 1
Original pixels, fixed codec and current checkpoint. Encoding and decoding together. Image errors are fidelity proxies. A lower score does not establish accurate shapes or readable writing.
500 updates · Incoming image 5
500 updates · Incoming image 5
Original pixels, fixed codec and current checkpoint. Encoding and decoding together. Image errors are fidelity proxies. A lower score does not establish accurate shapes or readable writing.
500 updates · Incoming image 5 · writing detail
500 updates · Incoming image 5 · writing detail
Identical native crop at 2x; whole-image errors are labeled separately. Image errors are fidelity proxies. A lower score does not establish accurate shapes or readable writing.

Limits and next question

One bounded training seed and exposed diagnostic cases; no fresh generalization claim, mixed-version intake, codec adoption or repaired rotation. Numerical alerts require interpretation alongside the actual images.

Next question. Use observed gains and losses to choose the next focused training change before returning to milestone expansion.

Inspect the evidence

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