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

Native continuation: completed evaluation recovery

Registered 8 Oct 2026 · Snapshot 12 Oct 2026

Execution: completedScientific verdict: acuity and retention gain

The question

Improve native visual encoding and decoding through broader experience while measuring whether continued learning preserves previously stored VSA percepts.

Why this experiment?

A memory can only preserve what its sensory interface encodes. These studies separate encoder quality, decoder quality and compatibility with earlier stored vectors.

How it works

Change one component or input representation at a time and evaluate immediate reconstruction on separate images. Keep old-vector readback visible so gains on new inputs do not hide compatibility losses.

Follow this part of the investigation: Sharper perception can also change the language of memory →
Implementation and comparison details

Continue the development-selected joint codec using broader photographs at native dimensions and aspect ratios. Compare ordinary joint refinement with matched continuation plus replay toward archived readback of stored VSA vectors. Separate acquisition, replay training, development selection and held-out retained memories. Measure fresh quality, old-vector readback, retrieval, exposure and resource costs; original acquisition identities and earlier lives remain intact.

What happened

Completed and independently audited: replay lowers fresh Flickr MSE 5.28% and held-out stored-memory MSE 0.35%; ordinary training improves fresh Flickr 10.11% but worsens retained readback 21.92%. The registered verdict is acuity_and_retention_gain. Occluded-cue retrieval still falls from 81.25% to 76.04% with replay (ordinary 65.28%). Fine textures remain soft; larger-video MSE worsens 4.80% on average. Run-001's evaluation failed at a CUDA device-name check; run-002 is evaluation recovery of the same comparison, with exact inherited inputs and locked endpoints, no retraining and no reselection. Repeated baseline tensors are byte-identical. No candidate is installed.

Measurements
Measurements — values preserved from this run
Recorded fieldValue
verdictacuity_and_retention_gain

Results and visual artifacts

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

This experiment trains an image encoder and decoder on more visual experience while checking what happens to older vector memories. “Ordinary” learns from new images only; “replay” also rehearses decoding old vectors toward their earlier decoded images. Replay targets are imperfect reconstructions, not original photographs. Seeds 401–403 are repeated training trials; seed 402 was chosen in advance for final illustrations. An optimizer update is one training step. MSE means mean squared RGB pixel error, where lower is better. Development images guide method choices; final test images measure the locked comparison. Error panels use a fixed fourfold display scale, separate from the numerical scores. Final evaluation was recovered after an execution failure without retraining or choosing new endpoints. The progress figures below are the same retained development history, not additional experimental replications.

COCO test photographs — before and after continued training
COCO test photographs — before and after continued training
These are the first six sources in fixed order, rather than examples chosen for their results. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

COCO test photographs — enlarged central detail
COCO test photographs — enlarged central detail
The central crop of the first source is at most 128 × 128 original pixels and is enlarged twofold. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Flickr test photographs — before and after continued training
Flickr test photographs — before and after continued training
These are the first six sources in fixed order, rather than examples chosen for their results. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Flickr test photographs — enlarged central detail
Flickr test photographs — enlarged central detail
The central crop of the first source is at most 128 × 128 original pixels and is enlarged twofold. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Text on a plain background — before and after continued training
Text on a plain background — before and after continued training
These are the first six sources in fixed order, rather than examples chosen for their results. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds. This tests visual reconstruction of writing, not reading or language understanding.

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

Text on a plain background — enlarged central detail
Text on a plain background — enlarged central detail
The central crop of the first source is at most 128 × 128 original pixels and is enlarged twofold. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds. This tests visual reconstruction of writing, not reading or language understanding.

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

Text over photographs — before and after continued training
Text over photographs — before and after continued training
These are the first six sources in fixed order, rather than examples chosen for their results. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds. This tests visual reconstruction of writing, not reading or language understanding.

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

Text over photographs — enlarged central detail
Text over photographs — enlarged central detail
The central crop of the first source is at most 128 × 128 original pixels and is enlarged twofold. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds. This tests visual reconstruction of writing, not reading or language understanding.

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

Video frames at larger resolutions — before and after continued training
Video frames at larger resolutions — before and after continued training
These are the first six sources in fixed order, rather than examples chosen for their results. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Video frames at larger resolutions — enlarged central detail
Video frames at larger resolutions — enlarged central detail
The central crop of the first source is at most 128 × 128 original pixels and is enlarged twofold. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. Each model encodes and decodes the same new input image. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Old memories excluded from rehearsal — before and after continued training
Old memories excluded from rehearsal — before and after continued training
These are the first six sources in fixed order, rather than examples chosen for their results. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. The same old vectors are decoded by each model; original photographs supply comparison references only. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Old memories excluded from rehearsal — enlarged central detail
Old memories excluded from rehearsal — enlarged central detail
The central crop of the first source is at most 128 × 128 original pixels and is enlarged twofold. Columns show the original image, starting model, training on new images only, training with memory rehearsal, then the error of each model in the same order. The same old vectors are decoded by each model; original photographs supply comparison references only. 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. The final trained examples use the prespecified seed 402; aggregate results include all three seeds.

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

Partial-cue recall after training — first six queries
Partial-cue recall after training — first six queries
These are the first six fixed held-out-memory queries, selected without checking outcomes. Read across: an 80 × 80 local view with 20% of pixels hidden, the correct old memory, and the memories selected by the starting encoder, the encoder trained on new images alone, and the encoder trained with memory rehearsal. All answers are shown with their frozen acquisition renderings so decoder quality cannot change which answer appears correct. The bank and cue pixels stay fixed; only the cue encoder changes. This is matching a seen source photograph, not semantic recognition. Trained examples use seed 402.

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

Partial-cue recall after training — selected error cases
Partial-cue recall after training — selected error cases
These are the first six queries with an error in at least one displayed encoder; this selection illustrates failures rather than estimating their frequency. Read across: an 80 × 80 local view with 20% of pixels hidden, the correct old memory, and the memories selected by the starting encoder, the encoder trained on new images alone, and the encoder trained with memory rehearsal. All answers are shown with their frozen acquisition renderings so decoder quality cannot change which answer appears correct. The bank and cue pixels stay fixed; only the cue encoder changes. This is matching a seen source photograph, not semantic recognition. Trained examples use seed 402.

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

A very dark video frame — reconstruction comparison
A very dark video frame — reconstruction comparison
Columns show the original frame, the starting model, training on new images only, training with memory rehearsal, and each model’s error. This is the first recorded Alien clip frame, using saved model outputs and the prespecified display seed 402. 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. Almost the entire input is dark with sparse stars. A small average pixel error is therefore weak evidence for recovering fine detail; the near-black appearance of these panels is expected.

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

A very dark video frame — enlarged central detail
A very dark video frame — enlarged central detail
Columns show the original frame, the starting model, training on new images only, training with memory rehearsal, and each model’s error. This is the central crop of the first Alien clip frame, enlarged twofold. 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. Almost the entire input is dark with sparse stars. A small average pixel error is therefore weak evidence for recovering fine detail; the near-black appearance of these panels is expected.

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

The fixed development images used to track training
The fixed development images used to track training
These six source images are evaluation references, not model outputs. They keep the same order throughout the training-progress sheets so a change in the reconstruction can be compared against the same scene. The dimensions describe the original rectangular input; each picture is resized only to fit this page.

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

Training with memory rehearsal: new-image reconstruction after 500 updates (replay-weight pilot 1)
Training with memory rehearsal: new-image reconstruction after 500 updates (replay-weight pilot 1)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results. Pilot weight 1 multiplies the replay loss relative to the fresh-image loss; it is not a number of replayed images.

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

Training with memory rehearsal: old-memory readback after 500 updates (replay-weight pilot 1)
Training with memory rehearsal: old-memory readback after 500 updates (replay-weight pilot 1)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results. Pilot weight 1 multiplies the replay loss relative to the fresh-image loss; it is not a number of replayed images.

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

Learning sharper new images while preserving old vector memories
Learning sharper new images while preserving old vector memories
Horizontal axes count training updates; each line is one method and training seed, with short pilot lines testing replay weights 1, 4 and 16. Left: reconstruction MSE on new development images, where lower is better. Middle: percentage change in old-memory reconstruction error relative to its acquisition value; zero preserves the original error, negative improves it, and the dashed red line is the provisional +2% development tolerance. Right: MSE between current and original decodings of the same old vectors, where zero means unchanged rendering. Low rendering drift can also preserve old decoding errors. These are development curves, not final test results. Replay rehearses saved vectors against their earlier decodings. The interrupted attempt and recovery share these recorded training histories; they are not independent training repetitions.

Original recorded figure ↗

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

Training with memory rehearsal: new-image reconstruction after 500 updates (replay-weight pilot 4)
Training with memory rehearsal: new-image reconstruction after 500 updates (replay-weight pilot 4)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results. Pilot weight 4 multiplies the replay loss relative to the fresh-image loss; it is not a number of replayed images.

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

Training with memory rehearsal: old-memory readback after 500 updates (replay-weight pilot 4)
Training with memory rehearsal: old-memory readback after 500 updates (replay-weight pilot 4)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results. Pilot weight 4 multiplies the replay loss relative to the fresh-image loss; it is not a number of replayed images.

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

Training with memory rehearsal: new-image reconstruction after 500 updates (replay-weight pilot 16)
Training with memory rehearsal: new-image reconstruction after 500 updates (replay-weight pilot 16)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results. Pilot weight 16 multiplies the replay loss relative to the fresh-image loss; it is not a number of replayed images.

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

Training with memory rehearsal: old-memory readback after 500 updates (replay-weight pilot 16)
Training with memory rehearsal: old-memory readback after 500 updates (replay-weight pilot 16)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results. Pilot weight 16 multiplies the replay loss relative to the fresh-image loss; it is not a number of replayed images.

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

Training on new images only: new-image reconstruction after 1,000 updates (trial seed 401)
Training on new images only: new-image reconstruction after 1,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 1,000 updates (trial seed 401)
Training on new images only: old-memory readback after 1,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 2,000 updates (trial seed 401)
Training on new images only: new-image reconstruction after 2,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 2,000 updates (trial seed 401)
Training on new images only: old-memory readback after 2,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 3,000 updates (trial seed 401)
Training on new images only: new-image reconstruction after 3,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 3,000 updates (trial seed 401)
Training on new images only: old-memory readback after 3,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 4,000 updates (trial seed 401)
Training on new images only: new-image reconstruction after 4,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 4,000 updates (trial seed 401)
Training on new images only: old-memory readback after 4,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 5,000 updates (trial seed 401)
Training on new images only: new-image reconstruction after 5,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 5,000 updates (trial seed 401)
Training on new images only: old-memory readback after 5,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 6,000 updates (trial seed 401)
Training on new images only: new-image reconstruction after 6,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 6,000 updates (trial seed 401)
Training on new images only: old-memory readback after 6,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 1,000 updates (trial seed 401)
Training with memory rehearsal: new-image reconstruction after 1,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 1,000 updates (trial seed 401)
Training with memory rehearsal: old-memory readback after 1,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 2,000 updates (trial seed 401)
Training with memory rehearsal: new-image reconstruction after 2,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 2,000 updates (trial seed 401)
Training with memory rehearsal: old-memory readback after 2,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 3,000 updates (trial seed 401)
Training with memory rehearsal: new-image reconstruction after 3,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 3,000 updates (trial seed 401)
Training with memory rehearsal: old-memory readback after 3,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 4,000 updates (trial seed 401)
Training with memory rehearsal: new-image reconstruction after 4,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 4,000 updates (trial seed 401)
Training with memory rehearsal: old-memory readback after 4,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 5,000 updates (trial seed 401)
Training with memory rehearsal: new-image reconstruction after 5,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 5,000 updates (trial seed 401)
Training with memory rehearsal: old-memory readback after 5,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 6,000 updates (trial seed 401)
Training with memory rehearsal: new-image reconstruction after 6,000 updates (trial seed 401)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 6,000 updates (trial seed 401)
Training with memory rehearsal: old-memory readback after 6,000 updates (trial seed 401)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 1,000 updates (trial seed 402)
Training with memory rehearsal: new-image reconstruction after 1,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 1,000 updates (trial seed 402)
Training with memory rehearsal: old-memory readback after 1,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 2,000 updates (trial seed 402)
Training with memory rehearsal: new-image reconstruction after 2,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 2,000 updates (trial seed 402)
Training with memory rehearsal: old-memory readback after 2,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 3,000 updates (trial seed 402)
Training with memory rehearsal: new-image reconstruction after 3,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 3,000 updates (trial seed 402)
Training with memory rehearsal: old-memory readback after 3,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 4,000 updates (trial seed 402)
Training with memory rehearsal: new-image reconstruction after 4,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 4,000 updates (trial seed 402)
Training with memory rehearsal: old-memory readback after 4,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 5,000 updates (trial seed 402)
Training with memory rehearsal: new-image reconstruction after 5,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 5,000 updates (trial seed 402)
Training with memory rehearsal: old-memory readback after 5,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 6,000 updates (trial seed 402)
Training with memory rehearsal: new-image reconstruction after 6,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 6,000 updates (trial seed 402)
Training with memory rehearsal: old-memory readback after 6,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 1,000 updates (trial seed 402)
Training on new images only: new-image reconstruction after 1,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 1,000 updates (trial seed 402)
Training on new images only: old-memory readback after 1,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 2,000 updates (trial seed 402)
Training on new images only: new-image reconstruction after 2,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 2,000 updates (trial seed 402)
Training on new images only: old-memory readback after 2,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 3,000 updates (trial seed 402)
Training on new images only: new-image reconstruction after 3,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 3,000 updates (trial seed 402)
Training on new images only: old-memory readback after 3,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 4,000 updates (trial seed 402)
Training on new images only: new-image reconstruction after 4,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 4,000 updates (trial seed 402)
Training on new images only: old-memory readback after 4,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 5,000 updates (trial seed 402)
Training on new images only: new-image reconstruction after 5,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 5,000 updates (trial seed 402)
Training on new images only: old-memory readback after 5,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 6,000 updates (trial seed 402)
Training on new images only: new-image reconstruction after 6,000 updates (trial seed 402)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 6,000 updates (trial seed 402)
Training on new images only: old-memory readback after 6,000 updates (trial seed 402)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 1,000 updates (trial seed 403)
Training on new images only: new-image reconstruction after 1,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 1,000 updates (trial seed 403)
Training on new images only: old-memory readback after 1,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 2,000 updates (trial seed 403)
Training on new images only: new-image reconstruction after 2,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 2,000 updates (trial seed 403)
Training on new images only: old-memory readback after 2,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 3,000 updates (trial seed 403)
Training on new images only: new-image reconstruction after 3,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 3,000 updates (trial seed 403)
Training on new images only: old-memory readback after 3,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 4,000 updates (trial seed 403)
Training on new images only: new-image reconstruction after 4,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 4,000 updates (trial seed 403)
Training on new images only: old-memory readback after 4,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 5,000 updates (trial seed 403)
Training on new images only: new-image reconstruction after 5,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 5,000 updates (trial seed 403)
Training on new images only: old-memory readback after 5,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training on new images only: new-image reconstruction after 6,000 updates (trial seed 403)
Training on new images only: new-image reconstruction after 6,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training on new images only: old-memory readback after 6,000 updates (trial seed 403)
Training on new images only: old-memory readback after 6,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 1,000 updates (trial seed 403)
Training with memory rehearsal: new-image reconstruction after 1,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 1,000 updates (trial seed 403)
Training with memory rehearsal: old-memory readback after 1,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 2,000 updates (trial seed 403)
Training with memory rehearsal: new-image reconstruction after 2,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 2,000 updates (trial seed 403)
Training with memory rehearsal: old-memory readback after 2,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 3,000 updates (trial seed 403)
Training with memory rehearsal: new-image reconstruction after 3,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 3,000 updates (trial seed 403)
Training with memory rehearsal: old-memory readback after 3,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 4,000 updates (trial seed 403)
Training with memory rehearsal: new-image reconstruction after 4,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 4,000 updates (trial seed 403)
Training with memory rehearsal: old-memory readback after 4,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 5,000 updates (trial seed 403)
Training with memory rehearsal: new-image reconstruction after 5,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 5,000 updates (trial seed 403)
Training with memory rehearsal: old-memory readback after 5,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: new-image reconstruction after 6,000 updates (trial seed 403)
Training with memory rehearsal: new-image reconstruction after 6,000 updates (trial seed 403)
Each row shows a fixed development image, its current encode-and-decode reconstruction, and its pixel error. The whole original-size image is processed; the thumbnail is only a display. Compare the same scene across training steps to see acuity change. 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 are development examples, not unseen final-test results.

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

Training with memory rehearsal: old-memory readback after 6,000 updates (trial seed 403)
Training with memory rehearsal: old-memory readback after 6,000 updates (trial seed 403)
Each row holds the old vector fixed. Columns show the original photograph, its reconstruction at acquisition, the current decoder’s reconstruction of that same vector, then the two errors against the photograph. This tests whether decoder training changes the meaning of a stored representation. 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 are development examples, not unseen final-test results.

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

COCO test photographs, example 1 — starting-model reconstruction
COCO test photographs, example 1 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 1 — reconstruction after new-image training
COCO test photographs, example 1 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 1 — reconstruction after training with memory rehearsal
COCO test photographs, example 1 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 1 — original input
COCO test photographs, example 1 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 2 — starting-model reconstruction
COCO test photographs, example 2 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 2 — reconstruction after new-image training
COCO test photographs, example 2 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 2 — reconstruction after training with memory rehearsal
COCO test photographs, example 2 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

COCO test photographs, example 2 — original input
COCO test photographs, example 2 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 1 — starting-model reconstruction
Flickr test photographs, example 1 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 1 — reconstruction after new-image training
Flickr test photographs, example 1 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 1 — reconstruction after training with memory rehearsal
Flickr test photographs, example 1 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 1 — original input
Flickr test photographs, example 1 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 2 — starting-model reconstruction
Flickr test photographs, example 2 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 2 — reconstruction after new-image training
Flickr test photographs, example 2 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 2 — reconstruction after training with memory rehearsal
Flickr test photographs, example 2 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Flickr test photographs, example 2 — original input
Flickr test photographs, example 2 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Text on a plain background, example 1 — starting-model reconstruction
Text on a plain background, example 1 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 1 — reconstruction after new-image training
Text on a plain background, example 1 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 1 — reconstruction after training with memory rehearsal
Text on a plain background, example 1 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 1 — original input
Text on a plain background, example 1 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 2 — starting-model reconstruction
Text on a plain background, example 2 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 2 — reconstruction after new-image training
Text on a plain background, example 2 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 2 — reconstruction after training with memory rehearsal
Text on a plain background, example 2 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text on a plain background, example 2 — original input
Text on a plain background, example 2 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 1 — starting-model reconstruction
Text over photographs, example 1 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 1 — reconstruction after new-image training
Text over photographs, example 1 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 1 — reconstruction after training with memory rehearsal
Text over photographs, example 1 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 1 — original input
Text over photographs, example 1 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 2 — starting-model reconstruction
Text over photographs, example 2 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 2 — reconstruction after new-image training
Text over photographs, example 2 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 2 — reconstruction after training with memory rehearsal
Text over photographs, example 2 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Text over photographs, example 2 — original input
Text over photographs, example 2 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. Images keep their original proportions; resizing on this page is for display only. Visible words are part of the image; this is not a test of reading comprehension.

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

Video frames at larger resolutions, example 1 — starting-model reconstruction
Video frames at larger resolutions, example 1 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 1 — reconstruction after new-image training
Video frames at larger resolutions, example 1 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 1 — reconstruction after training with memory rehearsal
Video frames at larger resolutions, example 1 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 1 — original input
Video frames at larger resolutions, example 1 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 2 — starting-model reconstruction
Video frames at larger resolutions, example 2 — starting-model reconstruction
This is the decoded output for the same source shown in the corresponding original-input image. The starting model has not received continuation training. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 2 — reconstruction after new-image training
Video frames at larger resolutions, example 2 — reconstruction after new-image training
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 2 — reconstruction after training with memory rehearsal
Video frames at larger resolutions, example 2 — reconstruction after training with memory rehearsal
This is the decoded output for the same source shown in the corresponding original-input image. The model shown is the prespecified display trial, seed 402. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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.

Video frames at larger resolutions, example 2 — original input
Video frames at larger resolutions, example 2 — original input
This is the source image used as the comparison reference, not a reconstruction. Compare corresponding images by example number to inspect fine structure and color. This single-image view has no error overlay; the comparison sheet shows errors on a common scale. 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

Both arms refine the same joint native codec. Replay targets are archived old-decoder outputs, not original RGB truth, and preserving those outputs can preserve their errors. Retention training, development and final cohorts have separate roles; held-out retention is known acquired experience, while fresh final photographs measure new-input quality. Source-free cross-version readback explicitly retains acquisition and current decoder identities; it does not establish automatic codec compatibility. Geometry and spatial bindings are supplied sensory structure. Each complex128 region costs 256 KiB before overlap and other state; report bytes, exposure, replay overhead, latency and peak memory. Earlier lives retain their original models; no candidate is installed. Reconstruction, retention and measured retrieval remain separate from reading, learned meaning or general cognition.

Next question. Test encoder compatibility during continued learning, retaining fresh-acuity and stored-readback gates and adding a retrieval-preservation criterion. The exact mechanism needs its own justified plan; the present study does not identify a learning ceiling or justify checkpoint adoption.

Inspect the evidence

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