jakeatx/ream-288-rank32-3m-adapter

Add base-model retrace gate

Seal targeted termination repair v1

Add isolated 5e-7 repair fallback arm

Add targeted termination repair training package

Export rejected 2e-6 and 5e-6 loop-repair canaries

Make checkpoint 1158 control portable from verified Hub export

Export checkpoint 1158 promotion control

Seal portable loop-repair prep release

Prepare checkpoint 3489 promotion and stabilized LR canaries

Remove superseded failed REAM-288 repair checkpoints

Export last-known-good REAM-288 2e-5 checkpoint 3489

Add sealed REAM-288 loop-repair training package

Add full-context v2 run provenance at 5001448

Export full-context pass4 checkpoint 5001448

Default one-GPU continuation to short cohort before deferred long tail

Export full-context pass4 checkpoint and unified remainder 2230317

Export full-context pass4 checkpoint 1620271

Record verified budget-boundary checkpoint redownload

Export budget-boundary full-context checkpoint budget-checkpoint-1018003 (part 2)

Export budget-boundary full-context checkpoint budget-checkpoint-1018003

Record verified budget-boundary checkpoint redownload

Export budget-boundary full-context checkpoint budget-checkpoint-766283 (part 2)

Export budget-boundary full-context checkpoint budget-checkpoint-766283

Record verified checkpoint redownload

Export audited full-context 25K+ checkpoint checkpoint-462864

Add validated grouped NVFP4 training runtime

Add Astra RTX runtime canary receipts

Add corrected full-context resume handoff

Add exact heldout validation receipt for corrected 256461-token checkpoint

Upload corrected full-context R2 pilot checkpoint 256461

Reject incorrect window-trained first third: remove adapter and optimizer

Remove transient Python bytecode from recovery bundle

Add validated second-third continuation runner and deterministic window order

Add targeted long-context R2 first-third recovery at 1,833,333 tokens

Document Qwen3.8 Perian lineage and training provenance

Add 9.336M high-speed Terminal-Bench audit

Add stopped 9,336,692-token checkpoint

Add 9,000,725-token milestone

Add verified 8M-token final-epoch checkpoint

Add verified 7.0M-token final-epoch checkpoint

Recover 6.25M continuation from CUDA allocator fragmentation

Add verified 6.25M supervised-token continuation checkpoint

Update final epoch profile test for measured floor

Update RTX PRO continuation floor documentation

Continue RTX PRO final epoch to $1.50 reserve

Archive final-epoch RTX PRO throughput gate and port evidence

Rebuild causal-conv1d against exact paid-run CUDA ABI

Fix final-epoch tokenized gate path and make launcher resumable

Fix final epoch resume layout and portable private data staging

Allow verified pre-staged private dataset transfer