Experiment trajectory: 9 October 2026

Paired generalization progress in R2

A read-only R2 check at 11:37 a.m. America/New_York found 2 completed jobs, 1 resumable job, and 42 pending jobs in paired_generalization_v2. The check read the saved study plan, progress pointers, and completed reports, and validated the referenced artifact sizes and recorded hashes in checkpoint metadata. It downloaded no models and changed no R2 objects. Local evidence is in outputs/r2-review/progress-2026-10-09/progress.json.

The completed jobs are sources/order_01/seed_0, with all eight source lessons complete, and controls/order_01/seed_0/learn_15, with the learning-only task-15 lesson complete. The resumable job is branches/order_01/seed_0/forget_3/learn_15. Its saved active request is forgetting task 3 at step 0, with the eight source requests complete. This pointer does not establish any saved unlearning updates. Work in a running session after this boundary is not reflected in this snapshot.

One of nine sources and one of nine controls are complete. None of the 27 unlearn-then-learn branches is complete, so none of the 27 paired comparisons is ready. The 2-of-45 job count is 4.4%, but jobs require different amounts of work, so it is not a GPU-time completion estimate. No accuracy or averaged effect was assessed in this check.

Next: continue notebook 19 with the same study ID, fixed revision, and settings on A100. Keep only one writer active. Notebook 20 can inspect partial reports, but paired averages require completed branches and matching controls. Wiki publication remains on its scheduled days.

Deadline queue and saved step-60 progress

The user requested a smaller run because the study must be wrapped up today, authorized reuse of existing progress, and confirmed that the notebook was stopped. A fresh read-only R2 check at 11:51 a.m. America/New_York confirmed that the source and control remain complete and that the task-3 branch now has a verified resumable boundary at unlearning step 60. No branch has completed yet. The local progress audit now contains this latest snapshot.

Notebook 19 now selects only the order-01, seed-0 source, its L15 control, U3 followed by L15, and U14 followed by L15. The source and control are reused. U3/L15 runs before U14/L15. All other 41 jobs remain in the saved study but are excluded from execution. The study plan, job contracts, fixed training revision, and checkpoints are unchanged. Selection is applied only to the session scheduler, so the existing step-60 state remains resumable. MAX_JOBS is now two; the 120-minute soft budget and 15-minute save reserve remain. An unfinished branch stops the session and resumes in a later session.

Five mocked notebook checks passed, including exact selected-job execution, source/control reuse, unchanged plan and original scheduler, missing-source refusal, multipart retry, and existing CPU flows. No GPU training was performed by this change. The resulting two comparisons, if completed, are a single-seed pilot. Notebook 20 still reports the full study as partial and does not treat this subset as the completed multi-seed matrix. The next step is to reopen the updated notebook 19 on A100, enable training, and run all cells with the same study ID. Do not run PREPARE again.

User-reported results for the two selected comparisons

The user supplied notebook-20 output for both order-01, seed-0 comparisons. The review output reports one valid seed for each setting. These results were supplied by the user and have not been independently reread from R2 for this update. The supplied local export path is /content/cark-paired-study/outputs/paired_study/paired_generalization_v2/summaries/20261009T175649556604Z; this names an export directory and does not independently establish run completion time or remote export persistence.

Both branches reuse the source sequence L3, L0, L9, L5, L17, L1, L7, L14 and the same L15-only control. Differences below are final validation accuracy in the unlearn-then-learn branch minus the learning-only control, in percentage points. The retained-task mean excludes the forgotten task.

Continuation	Task-15 difference	Mean retained-task difference	Largest retained-task drop
U3 then L15	0.0	-1.086	Task 9: 11.0
U14 then L15	-2.6	-1.6	Task 17: 14.2

Retained task	U3 then L15 difference	U14 then L15 difference
0	+10.6	+18.6
3	Forgotten, excluded	-2.2
5	+4.8	-2.4
7	+2.2	-3.0
9	-11.0	-7.0
14	0.0	Forgotten, excluded
17	-8.0	-14.2
1	-6.2	-1.0

The selected two-comparison pilot now has results, while the original multi-seed study remains incomplete. Complete setting: False means the planned seeds 1 and 2 are missing, not that the seed-0 comparison failed. Standard deviation is unavailable because each setting has one seed. Average retained-task differences are small relative to some individual drops; positive changes on other tasks offset the losses. The two comparisons share a control and have different retained-task sets, so they are not independent repetitions of one setting.

The supplied excerpt omits absolute task-15 accuracies, the forgotten task's accuracy after unlearning and after L15, and the forgetting gate. It therefore does not establish successful forgetting or persistence of forgetting through L15. No erasure, recovery, privacy, or across-seed robustness claim follows from these results. Next: inspect those missing fields in the saved CPU review reports and finalize the single-seed pilot report. Additional GPU training is not required to inspect saved evidence. Wiki publication remains on its scheduled days.

Seed-only execution configuration

The user requested that notebook 19 require only a seed change followed by Run all. Its configuration now exposes SEED, defaulting to 0 and accepting the saved seeds 0, 1, and 2. It derives the order-01 source, L15 control, U3/L15 branch, and U14/L15 branch IDs from that seed. RUN_TRAINING now defaults to True. The previous False default was a guard against unintended GPU execution; the user has explicitly requested direct execution. MAX_JOBS is four so a new seed can proceed through all four selected jobs within the unchanged soft time budget. Completed jobs are skipped; unfinished work resumes with the same seed. New seeds require independently trained eight-lesson sources and may take multiple sessions. The study plan, fixed implementation revision, and existing results are unchanged.

Five mocked notebook tests passed, including selecting correct jobs for all three seeds, rejecting an unsupported seed or CPU execution, and the existing selection, resume-contract, upload-retry, and CPU-review checks. No new GPU training or R2 modification was performed by this configuration change.

Automatic R2 summary exports

The user requested durable storage of the experiment summaries. Notebook 20 previously made R2 export optional and disabled it by default, even though training reports were already saved remotely. It now enables R2 export by default and uses study ID paired_generalization_v2. Its final cell saves summary JSON, the study configuration, both CSV tables, and PNG/SVG figures when generated to a timestamped folder under the study's summaries/ prefix. Every artifact is content-verified, with progress bars. The final success message includes the R2 location and appears only after all uploads finish. A disabled export flag produces an explicit local-only message.

Five mocked notebook checks passed. The CPU export test now verifies the default-on behavior without manually enabling the flag. No live summary upload was performed by this implementation change. Earlier local summaries can be regenerated from the saved R2 job reports using notebook 20 on CPU; no additional training is needed.

Compact overall progress display

The user supplied a screenshot showing dataset-download progress printed repeatedly as separate lines and requested a clear view of overall experiment progress. Notebook 19 now uses a single HTML display that updates in place beside the running cell. It counts the narrowed order-01 study across all three saved seeds: 12 jobs, six comparisons, and 39 requests. The selected seed has four jobs and 13 requests. These counts exclude deferred orders and the task-5 branch; they are not percentages of expected GPU time.

The panel shows saved job statuses, current job and request, current learning or unlearning steps, epoch/loss, validation or transfer progress, and the latest verified R2 boundary. It does not credit inherited source requests again in branches. Live training and transfer updates are throttled to about once per second, with immediate stage and save notifications. Dataset downloads and R2 upload/download/verification use the same panel. Display hooks are restored when the session ends or setup fails. The fixed training revision, saved plan, optimizer, random state, and checkpoints are unchanged.

Three display tests and seven notebook-flow tests passed. Checks covered inherited-request exclusion, cross-seed and per-seed counts, throttling, nested-stage restoration, actual display-hook installation/restoration, and forwarding the original training arguments with verified-boundary callbacks. These were CPU and mocked checks, not a live Colab/GPU run. The user can use the updated notebook at the next restart with the same seed; no PREPARE run is needed. No new experiment result was supplied by the screenshot.

Seed 1 session saved at the budget boundary

The user supplied session record 20261009T222425633613Z, with status saved and plan hash a51f342dfdf1faef9dd22f06426db5ad1e5ac7800a6b9318df055400cf310171. The order-01, seed-1 source and L15-only control completed. The U3/L15 branch was attempted and returned resumable. The U14/L15 branch was not listed among attempted jobs. Two of four selected seed-1 jobs are complete. The saved excerpt does not show the active request, epoch, unlearning step, accuracy, or checkpoint hash; those details have not been independently read from R2 for this update.

The session ran for 6,321.9689 seconds, approximately 105 minutes 22 seconds, with MAX_JOBS four, a 120-minute soft budget, and a 15-minute save reserve. This is consistent with stopping at the 105-minute work deadline and saving at a boundary. The normal completion messages report saved progress followed by Colab runtime release; no error was supplied. These messages do not independently establish actual GPU release.

Next: reopen notebook 19 on A100, keep study ID paired_generalization_v2 and SEED 1, and run all cells. Completed source and control jobs are skipped. U3/L15 resumes from its verified boundary, then U14/L15 runs if the next session budget permits. Keep SEED 1 until both branches complete. Notebook 20 can then review seed-0 and seed-1 comparisons on CPU. No new seed-1 paired effect can be reported from this session excerpt.

Verified recorded session durations

A read-only R2 audit at 10:39 p.m. America/New_York on 9 October downloaded and content-verified all six JSON session logs under uncle/paired_generalization/paired_generalization_v2/sessions/. All six contained valid elapsed durations. Local evidence is in outputs/r2-review/session-times-2026-10-09/sessions.json.

Session start, Eastern daylight time	Work recorded	Outcome	Elapsed time
8 October, 19:11:32	Seed 0 source	Complete	1 h 33 m 33 s
8 October, 21:57:48	Seed 0 L15 control	Interrupted: S3UploadFailedError	25 m 10 s
8 October, 22:46:40	Seed 0 L15 control	Complete	19 m 28 s
9 October, 12:02:20	Seed 0 U3/L15 and U14/L15	Both complete	1 h 38 m 26 s
9 October, 14:07:49	Seed 1 source	Resumable	1 h 47 m 54 s
9 October, 18:24:25	Seed 1 source, L15 control, U3/L15	Source and control complete; branch resumable	1 h 45 m 22 s

The sum of the unrounded elapsed times is 26,993.584321 seconds, rounded to 7 h 29 m 54 s. The interrupted upload session is included. This is logged runner wall time, including data preparation inside the runner, training, evaluation, and checkpoint transfers. It excludes earlier notebook setup, gaps between sessions, active sessions not yet logged, and sessions that disconnected before their final log reached R2. It is not a complete measure of billed GPU time. This audit independently verifies the latest supplied seed-1 session record and seed-0 branch completion; it does not assess additional accuracy results.