{"name":"Can you use Jev for embeddings? Clean-slate question vectors","model":"jev-1.13.0","denseModel":"voyageai/voyage-4-large","denseResponseModel":"voyage-4-large","widths":[64,128,256,384,512,768,1024,1536],"maxWidth":1536,"blockSize":256,"bank":"Fixed authored/reworded generic yes/no questions, seed-hashed within groups and proportionally interleaved. No Cartesian object-color, actor-action or spatial products. No feature selection or learned weights. Nested prefixes frozen before dev inference.","prompt":"Shared state.policy once per request, not a protected system prompt. First 256 dimensions also run in a fresh repeated-policy control on every dev input, interleaved with shared blocks. Same sentence, model and question meanings.","selection":"Choose smallest listed width whose dev Spearman is within 0.01 absolute of the maximum across listed widths. This choice freezes before any target inference. Always complete all planned dev vectors. No bank changes after dev.","efficiencyTolerance":"Report relative Spearman loss shared versus repeated policy and actual billed input-token ratio. User target: less than 5% relative quality loss and roughly 2x token/cost saving. This is not classification accuracy. Shared policy remains the predeclared candidate even if the target fails; disclose failure, do not target-tune.","test":"Full pinned MTEB SICK-R split: 9927 pairs, 6077 unique strings. All three methods freshly encoded. Main three bars are selected Jev width, full Voyage 4 Large 2048d and fresh training-fitted TF-IDF. This is semantic textual similarity, not retrieval or clustering, and not a suite aggregate.","nearParity":"Descriptive near-parity only if relative point-score loss is at most 5%. Do not call statistically equal or generally as good. Paired bootstrap interval reported whether favorable or not. No benchmark-specific tuning.","metric":"Raw probability vectors, unit weights, ordinary cosine, Spearman with average ranks for ties. No centering, thresholds, fitted similarity or projection.","bootstrapDraws":2000,"bootstrapSeed":20261007,"concurrency":4,"inputTokensPerSecond":70000,"requestTimeoutMs":45000,"maxAttempts":4,"wallDeadlineMinutes":120,"operationalAmendment":"After dev selected 1536, extend the target phase deadline from 75 to 120 minutes. The short initial target attempt was stopped unscored and excluded in full; no vectors reused. Questions, width rule, model, batching, rate, inputs and scoring unchanged. Original dev protocol/freeze retained separately.","pricePerMillionInputUsd":0.042,"densePricePerMillionInputUsd":0.12,"denseBatchSize":96,"timing":"Matched serial synthetic sentences; each Jev sentence fans out independent blocks in parallel. Real unbatched Voyage request. No corpus batch/time division presented as latency.","preserve":"No old inference, scores, or audits overwritten. Fresh directory. Retain exact immutable request snapshots, receipts and retries. Public exports exclude corpus, raw requests, credentials and private caps. Dataset and code fingerprints freeze before inference. Repeat API noise and public-data training overlap are unknown."}