diff options
| -rw-r--r-- | COLLABORATOR_ONBOARDING.md | 36 | ||||
| -rw-r--r-- | CROSS_ARCHITECTURE_CROSSOVER.md | 23 | ||||
| -rw-r--r-- | README.md | 8 | ||||
| -rwxr-xr-x | experiments/run_a6000_matrix.py | 3 |
4 files changed, 48 insertions, 22 deletions
diff --git a/COLLABORATOR_ONBOARDING.md b/COLLABORATOR_ONBOARDING.md index 79b1091..f0c5b92 100644 --- a/COLLABORATOR_ONBOARDING.md +++ b/COLLABORATOR_ONBOARDING.md @@ -123,13 +123,30 @@ residualization is necessary. ## Frozen-run state -The Plain-CNN selector and its 27-cell endpoint are frozen. As of 2026-07-30, -the ResNet selector has 18/19 records and the Transformer selector has not yet -opened; therefore the current commit can start the Plain-CNN 27 cells but the -complete launcher intentionally refuses to start ResNet or Transformer -endpoints until their passed selector JSON files are committed. Wait for the -maintainer to announce the exact A6000 run-packet commit/tag before launching -`--family all`. Never construct or edit a selector manually. +All three validation-only selectors are frozen, and the complete A6000 matrix +is open at tag `a6000-x1-v1`. The ResNet selector passes 19/19 records at +[`results/resnet_crossover/p1_selector.json`](results/resnet_crossover/p1_selector.json) +(SHA-256 `0f0fcdfcd0056b034a9cd95e57970782bb2c33785832ab1cbe27e5dbe212111b`), +and the Transformer selector passes 27/27 records at +[`results/transformer_crossover/p1_selector.json`](results/transformer_crossover/p1_selector.json) +(SHA-256 `5bd65b2d22ff666c4f720f2094cd1c7d9a5eff68ceee27ece19da60a6a62d11b`). +Both use validation only, contain all nine methods, and report no missing +experiments. Never construct, edit, or retune a selector manually. + +| method | ResNet selected LR | Transformer selected LR | +|---|---:|---:| +| BP | 0.1 | 0.001 | +| FA | 0.1 | 0.001 | +| DFA | 0.01 | 0.003 | +| PEPITA | 0.003 | 0.001 | +| Forward-Forward | 0.03 | 0.0001 | +| EP | 0.0003 | 0.001 | +| Dual Propagation | 0.05 | 0.001 | +| clean KP | 0.1 | 0.003 | +| SDIL | 0.1 | 0.001 | + +These rates were selected on the smallest model of each family under the +precommitted bounded grid and are copied unchanged to all three family sizes. ## Machine and environment setup @@ -207,9 +224,8 @@ Start the formal run inside `tmux` or another persistent session: --family all 2>&1 | tee a6000-matrix.log ``` -Before the selector-release commit, the collaborator may start only the -already-frozen Plain-CNN block by replacing `--family all` with -`--family plain`. Later, rerunning `--family all` preserves every completed +The full `--family all` queue is now open. If the collaborator already started +the Plain-CNN block, rerunning `--family all` preserves every completed manifest and continues with ResNet and Transformer. A manifest is the resume boundary. If an output exists without its manifest, the runner stops and asks for an audit; do not delete the output just to make the queue continue. diff --git a/CROSS_ARCHITECTURE_CROSSOVER.md b/CROSS_ARCHITECTURE_CROSSOVER.md index cc27097..1101806 100644 --- a/CROSS_ARCHITECTURE_CROSSOVER.md +++ b/CROSS_ARCHITECTURE_CROSSOVER.md @@ -199,15 +199,20 @@ families. The panel does not prove asymptotic scaling, biological implementation, clean-setting superiority over BP, or universal superiority over EP/Dual Propagation. -## Execution ledger (2026-07-27) - -- Plain-CNN P1 is complete and its selector is frozen. The 27-cell P2 is - active on the only authorized physical GPUs, timan107 GPU5 and GPU7. -- ResNet X0, the 19-cell ResNet-20 selector, and its analyzer are frozen. - Formal R1 waits for an authorized GPU rather than sharing memory with P2. -- Transformer X0 passes in both development PyTorch and the Pascal PyTorch - 2.3 environment. Its 27-cell T1 registry and analyzer are frozen. Formal - T1 likewise waits for an authorized GPU. +## Execution ledger (2026-08-01) + +- Plain-CNN P1 is complete, its selector is frozen, and the complete 27-cell + P2 audit passes. +- ResNet X0 and the 19-cell ResNet-20 selector are complete. The passed + selector is tracked at `results/resnet_crossover/p1_selector.json`; formal + 27-cell R2 is active on timan107 GPU5 and GPU7. +- Transformer X0 and the 27-cell Transformer-4 selector are complete. The + passed selector is tracked at + `results/transformer_crossover/p1_selector.json`; formal T2 remains queued + behind R2 on the authorized GTX-1080 pair. +- A separate single-A6000 hardware profile, restart-safe launcher, and frozen + author-code patch stack are available at run-packet tag `a6000-x1-v1` for + an independent complete 81-cell replication. GPU serialization changes calendar time, not the cell registry. No pending family may borrow an unapproved GPU, and no method or size may be removed to @@ -153,8 +153,12 @@ nonfinite at VGG16. Dual Propagation reaches `92.38%` on VGG16 but takes `5.97` hours versus SDIL's `1.24`; clean KP slightly dominates SDIL there at `90.86%` in `1.08` hours. Thus the first family supports scaling and a cost advantage over iterative strong baselines, but not global Pareto -dominance. ResNet and Transformer crossover cells remain gated on their -complete validation-only selectors. +dominance. The ResNet and Transformer crossover cells were gated on their +complete validation-only selectors; both gates have now passed and are +tracked. ResNet is complete at 19/19 candidates and Transformer at 27/27, +with all nine selected rates frozen before the full-size endpoints. The +GTX-1080 ResNet R2 panel is active, and the independent A6000 81-cell run +packet is open at tag `a6000-x1-v1`. ## Publication-facing artifacts diff --git a/experiments/run_a6000_matrix.py b/experiments/run_a6000_matrix.py index e5330ef..15933ec 100755 --- a/experiments/run_a6000_matrix.py +++ b/experiments/run_a6000_matrix.py @@ -165,7 +165,8 @@ def main(): ) if args.family == "all": - require_file(plain_audit, "complete plain-CNN P2 audit") + if not args.dry_run: + require_file(plain_audit, "complete plain-CNN P2 audit") run( [ sys.executable, |