Research library
Measurements are shown with method, source, scope, and limitation so case evidence cannot outgrow its protocol.
Figure 1.The voice case reports 0.961 mean WavLM SECS for one speaker on held-out English sentences under an encoder-matched protocol.
The figure shows a one-speaker voice protocol. A 50-clip centroid enrolls the target. For each held-out English sentence, the pipeline generates 12 candidates and selects by WavLM scoring. The reported result is 0.961 mean WavLM SECS under an encoder-matched protocol; cross-encoder triangulation remains pending.
Figure 2.The Shakespeare LoRA case is a text-side meaning-compression story with explicit verification scope.
The figure shows a small public-domain Shakespeare corpus on the left and derived training records on the right. It frames the case as evidence that structured labels can be materialized from fixed real text while keeping the source data grounded.
Figure 3.Full State Verification requires a second read from the source of truth after execution.
The figure represents the Full State Verification pattern. A trigger runs the system and writes durable artifacts. A separate verifier then reads persisted files and reported records from their source of truth, rather than relying on the original process return value.
The case-study registry is typed by evidence status. A measured card cannot silently become a protocol card, and a protocol card cannot render as measured.
measured
The published case reports 0.961 mean WavLM SECS on 10 held-out English sentences for one speaker under an encoder-matched protocol.
measured
The current manuscript reports five manually verified prompts on the trained LoRA, including cross-lingual period-style transfer observed on a Spanish prompt.
protocol-level
The manuscript treats the identity-locked talking-head case as architecturally complete but measurement-pending.
measured
The private-preview paper describes a JEPA-style runtime whose reported outputs are read back from persisted source-of-truth artifacts.
00:03:27 / JGFFOTL35wA
Case-study video for the one-speaker WavLM SECS measurement and its scope limits.
The formal site claim is 0.961 mean WavLM SECS under an encoder-matched protocol for one speaker and 10 held-out English sentences.
00:11:49 / mXAJgE2G87Q
Concrete public-domain text example for how a small corpus can produce many derived training records.