# Lossless checkpoint delta probe — September 12, 2026

A read-only experiment used all 32 tensor components from the final real training transition, from parent `a46b988196995779124bde7d9c79a6dd90e2bfb55182db4e654a1dfd95bb2758` to candidate `ecb4262c2b7d63a0bfb0d17a9dc9023af33c351645d0e6ba12f087ba6d474677`.

Bytewise XOR against each exact parent component, followed by zlib level 1, reduced 538,085,392 raw tensor bytes to 153,503,190 bytes. Compressing the complete candidate components directly took 422,376,714 bytes. Decoding the deltas reproduced every target byte and all target SHA-256 roots.

On the shared host, one repetition measured 9.828 seconds for delta creation/compression, 4.803 seconds for restoration/hash verification, and 5.551 seconds for reading and hashing both input checkpoints. Direct candidate compression took 21.821 seconds; its decompression was not measured. These are local component measurements, not network throughput. Model metadata and transport envelopes are excluded.

This suggests an optional lossless transfer or archival encoding when a receiver already has the exact parent. It does not alter model identities or float32 arithmetic. It does not prove gradients, remove artifact availability requirements, or permit pruning an active dispute's inputs. Production use would need bounded decoding, missing-parent fallback, recovery checkpoints, ancestry retention rules, and measurements across many updates. No transport, ledger, training implementation, or active artifact was changed by this probe.