One exact neural realization. Verifiable claims, explicit limits.
AMF transformed Mistral 7B for billing/technical classification. The recorded campaign demonstrates physical parameter removal and bounded development results, not independent qualification or superiority over the unmodified parent.
parameters removed
ROB · development
OOD · development
recorded Ledger events
What was actually measured
The ROB/OOD results belong to the development campaign. The 12-case SEARCH and local holdout are separate checks, not an independent final evaluation. They must not be pooled to imply a certification.
Trace the exact candidate
The following identifiers were recorded for the campaign. They identify artifacts; they do not independently verify the underlying measurements. Artifact access is not claimed on this page.
20261022T210311Z_25979c170c708c41dac9275d15a8fff4eca08d52bab718abc643336f326b39b2a7773385dbf5ad4956ebfa9cfbb269e889ef2758d250357cb884c432c509b95b41274f508b12c171bc778fd9bc3ca0189e89e5b3c7ec8Honest limits make evidence useful.
No independent QIR campaign or independent attestation has been completed for FP2.
The original, pruning-only, LoRA-only and combined variants have not all been compared in one controlled, paired causal attribution protocol.
12.04% fewer parameters is not a measured 12.04% speedup, memory saving or financial saving.
Persistent Scientist memory retrieval and score influence are observed in CPU replays; better decisions across distinct future missions remain unproven.