# Full-size restart readiness: exact state, small GPU execution differences

V1 remains failed under its exact-execution criterion. V2 separately tested
state fidelity and numerical execution using bounds written before its run.
V2 manifest9cb4b2d32cfa964ea2d8af868d46bcbc55251292cd445c9b4e3bb4e181026071.

All live model and AdamW tensor bytes, shapes, dtypes and non-tensor values were
identical immediately after restoration. Full records are in
notes/latent_rl_restart_preflight_v2.json. The subsequent repeated update was
not bitwise identical:790781/751632384 coordinates differed, maximum absolute
error1.4901161193847656e-8, RMS2.4315625962448493e-11. The objective was identical;
preclip-norm relative difference2.7314278469196036e-6. Every coordinate passed
atol1e-8/rtol2.5e-7, aggregate RMS<=1e-9 and the scalar bounds. This demonstrates
faithful state restoration and bounded numeric execution for this check, not
identical future stochastic trajectories or population/seed equivalence.

Synthetic shape: all original751632384 FP32parameters, K32,960prompttokens,
eight scored completions128–512tokens, denominator8, two updates plus replay of
the second. Also sampled8synthetic16-token outputs with AdamW buffers resident.
Peakallocated22833771008bytes (21.27GiB), reserved24270340096bytes (22.60GiB).
Checkpoint raw9.01996GB, compressed6.58426GB. Worker267.87s/allocated272.2s;
GPU5released. Temporary synthetic checkpoint was deleted. No math questions,
verifier supervision or research checkpoint were used.

Decision: readiness supports implementing the broader fixed learning comparison
with explicit numeric reproducibility limits. Preserve V1 failure and both
manifests; do not relabel the first run as a PASS. Maintain exact source/order,
state hashes, frozen code and full work records despite GPU rounding differences.
