AGILLM-4.3-EGGROLL-Experimental

AGILLM 4.3 EGGROLL Experimental

<!-- V25-UNIFIED-START -->

v25 unified EGGROLL experimental runtime — 10 August 2026

Measured verdict: genuine MoE-router fitness signal, but no demonstrated pretraining speedup. The controlled fixed-token active branch took 41.7% longer than the mean baseline training window. Production v22 remains unchanged, and no long-running EGGROLL checkpoint fork was started.

v25 unifies strict fixed-token reproducibility and calibrated router EGGROLL:

• agillm43_singlefile_intelligence_v25_eggroll_unified_experimental.py
• launch_v25_calibrated_dry_run.sh
• RESULTS_20260810.md
• EXPERIMENT_RESULTS_20260810.json
• immutable bundle: validation/2026-08-10-v25/

SHA-256: 8b9e62eaa9dde68c624ea9f3d4c9311b4a841ec43a89761453754e0673e25674

<!-- V25-UNIFIED-END -->

Default-off EGGROLL experiments for AGILLM 4.3's top-1 Mixture-of-Experts routers, inspired by Evolution Strategies at the Hyperscale. Normal AdamW, AR, SAT, NAT and DiffusionBlock training remain intact.

Experiment verdict, 10 August 2026

A paid, fixed-token Hugging Face A100 probe from checkpoint step 1,777,105 found mechanistic antithetic signal but no demonstrated pretraining speedup.

• Exact tensor hashes matched across two baselines and one active branch.
• Active EGGROLL was 41.7% slower over the measured training window.
• Mean local-loss delta was -0.000786, median zero, 95% interval crossed zero widely, sign-test p 0.727.
• Better than both baselines on 3/14 commits, worse on 5/14, tied on 6/14.
• Ten updates passed non-regression, but none improved guard CE measurably.

Decision: keep it disabled by default and do not fork an active checkpoint from this evidence. See HF_EXPERIMENT_RESULTS_20260810.md and results/2026-08-10-hf-fixed-probe/.

Locations

• GitHub: Marxist-Leninist/AGILLM4.3-EGGROLL-Experimental
• Canonical Hugging Face: MarxistLeninist/AGILLM-4.3-EGGROLL-Experimental
• Legacy mirror: OpenTransformer/AGILLM-4.3-EGGROLL-Experimental

Versions

v23 original optional sidecar

agillm43_singlefile_intelligence_v23_eggroll_experimental.py, SHA-256 f487cf2fbb38b69bf9f19322759176c3e4b66636a6375fc30759434be22aab80.

Includes antithetic low-rank perturbations, candidate-major evaluation, search and guard crops, clipping, rollback, dry-run mode and JSONL receipts. Disabled when --eggroll_every_steps 0.

v24 strict fixed-token runtime

agillm43_singlefile_intelligence_v24_eggroll_fixed_tokens.py, SHA-256 e3c89682ed04e8d5419cf30e814577b50def3346d3546a4eb27f37acd78f19cf.

Adds --fixed_tokens_file, --fixed_tokens_sha256, --[no-]fixed_tokens_cycle, bypasses presets and hot dataset reload, and prints file/token-stream SHA-256 receipts. Test-only deterministic CUDA settings use AGILLM43_EXPERIMENTAL_DETERMINISTIC=1.

Frozen base

• Step 1,777,105
• Tokens 87,348,264,960
• Checkpoint SHA-256 72ee3cf3ae4e3dedb22f16d8c2d21c80c66e0bdcd2001f606f030886cc98e2af

Production v22 was not modified.

Files

• agillm43_singlefile_intelligence_v23_eggroll_experimental.py
• agillm43_singlefile_intelligence_v24_eggroll_fixed_tokens.py
• experiments/hf-fixed-probe/fixed_active_probe.py
• HF_EXPERIMENT_RESULTS_20260810.md
• results/2026-08-10-hf-fixed-probe/
• EXPERIMENT.md, EXPERIMENT_MANIFEST.json, and canonical HF base_checkpoint/

Limits

The valid probe covered 14 commits and 114,688 tokens per branch, population 16, rank 1, sigma 0.1, router-only. It argues against enabling this configuration as a pretraining accelerator; it does not rule out every EGGROLL target, engine or non-differentiable objective.