Every embedding method the research org has tried, with measured verdicts and its stage in the Research → Business Development → Execution pipeline.
Measured model versions from this workspace's core-api /v1/models, ranked by effective rank.
Based on 891 runs and 1,190 trainings · evidence: EVIDENCE.md §3.1–§3.7, SWEEP_FINDINGS.md, SWEEP_REPORT.md · updated 2026-06-28
The research org tests every embedding method on measured benchmarks. Methods that prove their worth are promoted to Business Development for real-world tenant pilots, then the Execution team implements the winners in production. Dead ends are archived with the evidence that killed them.
Under active investigation in the research org: measured, but not yet promoted.
Loss-space rank floor. Prevents collapse where it occurs (non-contrastive / low-negative regimes) but is neutral on standard InfoNCE real-text training. Kept as insurance, not a default.
Adds a differentiable effective-rank maximization term to InfoNCE. Works but ranked below Barlow/InfoNCE in the search.
Earned its keep in research. Handed to Business Development for real-world tenant pilots.
Cheap per-tenant fine-tuning on the tenant's domain. The defensible product moat: beats general commercial embeddings in-domain, with no out-of-domain forgetting at this scale.
Redundancy-reduction objective. Top method in the Bayesian search; keeps full-dim quality while resisting collapse. Real-text confirmation vs VICReg in progress.
Trains nested prefixes of the served representation so it degrades gracefully when truncated. The real lever for cheap truncated serving (decorrelation is not).
Validated and shipped. Part of the production serving / training path today.
Standard contrastive objective. Inherently resists collapse with adequate negatives (batch ≥16); the production default training path.
Per-vector int8 quantization for edge serving. Essentially lossless; ship it as the default cheap-serving tier.
Tested honestly and rejected. Kept here so we don't re-litigate dead ends.
Periodic weight-space SVD surgery shrinking the weak singular band. Fights anisotropy, not collapse. In a real collapse regime it made collapse worse (rank 3.4→1.0).
Redesigned surgery that lifts weak singular values. Fixes the harm of three-tier but still loses to doing nothing; weight-space surgery can't outrun the collapse gradient.
Bio-inspired wake/sleep phasing with synaptic downscaling + dream pruning + replay. Phasic consolidation fails; only a continuous in-loss rank floor works. Downscaling survives as a collapse-neutral renormalizer.
Self-distillation with centering + sharpening. Worst method in the Bayesian search on this task.