Continued interface learning and batched audit measurement
The continuation completed and was rejected. Single answers stayed at 13/16, combined answers stayed at 2/16, and structured answers stayed at 7/8. General retention failed: the upper confidence bound was +0.02127 against +0.02. Both 1,536-update arms and both exact optimizer replay checks completed. The final remained unopened. Mixed training response loss fell from 0.27065 to 0.01163; this did not transfer to better held-out combined answers. More steps on this interface are not the next experiment. See the complete result.
Eight GPU token checks took 0.640–0.661 seconds each. The three 128-token cached responses took 15.44–15.68 seconds to generate. Each fixed-context check sent 83,886,320 tensor bytes, versus 3,968,048–19,314,688 for cached generation across the eight cases. Six cached responses matched; two differed (one and two token positions). Both deliberately forged responses were rejected. These measured disagreements confirm that native acceptance cannot silently substitute one numerical method for the other.
The next frozen streaming trial uses those same weights and eight exposed cases. It checks the implemented canonical correction procedure, complete-response re-execution, two different chunk-boundary comparisons, and two first-token forgeries. It measures first checked-chunk latency, complete time and traffic under a one-hour/$10 bound. There is no training or quality search in this run. The candidate remains rejected regardless of an execution-method pass.
The first owned-interface trial improved single-fact answers to 13/16 and combined answers to 2/16. General retention passed (upper bound +0.01412 against +0.02), and structured answers stayed at 7/8. It failed the combined-answer and control-gain gates; the final remains unopened. Both three-owner optimizer restart checks passed. The complete result includes the public evidence and retired-resource accounting.
The continuation tests a specific remaining possibility: insufficient optimization of the mixed conversation interface. Mixed training response loss fell from 1.8643 to 0.3072 over four epochs; the auxiliary source loss ended at 0.8829. These training measurements do not guarantee generalization. They motivate one bounded run from the exact preceding gate and adapter roots.
Both arms receive 1,536 further updates on the unchanged training documents. Each starts from its own committed terminal weights with explicitly fresh Adam state, a lower learning rate and the same quality gates. This is a newly prescribed optimization trajectory, not a restart of the former optimizer. Checkpoint spacing is 64 updates, with an additional boundary for replaying the last 16. No intermediate selection is permitted. The two-hour/$25 limit and automatic retirement remain in force.
After quality scoring, eight frozen development cases also measure a distinct inference verification program. A reported response is included as input to one causal forward pass across all source owners. At each response position, the checker compares its predicted next token with the reported token. Two cases also change the first output token to check rejection and causal independence from future input. All original and tampered observations are retained.
The one-pass idea is described in the Verde production discussion. NeuroShard's diagnostic uses one document padded to a fixed context, with every position marked valid and a causal attention mask. Under a deterministic causal implementation with this fixed execution shape, earlier predictions cannot depend on later reported tokens. This gives a well-defined batched check of the greedy response. It is our proposed numerical contract, not an assertion that ordinary cached decoding follows the same floating-point program.
The trial therefore measures cached-response mismatches as well as time and transport cost. Passing examples do not establish universal equivalence. The existing native chain still uses its admitted replay method. A deployment of this alternative would need its own execution profile, provider behavior on numerical disagreements, funded audits and native acceptance integration. Cross-hardware reproducibility remains a separate constraint; see the Verde paper.