Representations
Compressed forms competing on measured preservation. An envelope is a fact about a cohort and a k, earned by an executed trial; a declared representation that has never been trialled shows none.
- representation
- exact-f32
- axis
- latent_geometry
- availability
- available
- bytes_per_body
- 1536
- envelope
- recall@5 1.000 over 47 institutional bodies, 0 false prunes, by construction
- cannot_infer
- an exact representation -> a correct answer; it is the definition of agreement
- representation
- int8
- axis
- latent_geometry
- availability
- available
- bytes_per_body
- 384
- envelope
- recall@5 0.987 at 4x, 3 false prunes and 3 false keeps over 47 queries
- cannot_infer
- this envelope -> the same envelope on a learned model, whose geometry differs from a hashed-token backend; recall at 5 -> recall at any other k
- representation
- sign-1bit
- axis
- latent_geometry
- availability
- available
- bytes_per_body
- 48
- envelope
- recall@5 0.170 at 32x, 195 false prunes over 47 queries
- cannot_infer
- poor preservation here -> poor preservation generally; a hashed-token embedding spreads signal uniformly across dimensions, which is the worst case for sign quantization
- representation
- halved-f32
- axis
- latent_geometry
- availability
- available
- bytes_per_body
- 768
- envelope
- recall@5 0.528 at 2x, 111 false prunes over 47 queries
- cannot_infer
- a crude projection -> a learned one; discarding dimensions is not dimensionality reduction
- representation
- fpq-v8
- axis
- numerical_body
- availability
- declared_untrialled
- bytes_per_body
- envelope
- cannot_infer
- a cosine reported on model tensors -> a preservation envelope for candidate retrieval; these are different axes and different questions
- representation
- cms-quantized-ann
- axis
- latent_geometry
- availability
- declared_untrialled
- bytes_per_body
- envelope
- cannot_infer
- a shape that works on three fixtures -> an envelope; three examples establish that the path exists, not what it preserves