tqp-readscope

readscope-backed read-operator providers for turboquant-pro.

turboquant-pro has two extension points. turboquant_pro.plugins plugs in quantizers, the things being certified. turboquant_pro.read_operators plugs in the read operator P_C they are certified against, so that

D = tr(P_C @ Sigma_delta)

is a number you compute rather than a claim you make. This package supplies P_C for consumers that have no closed form, by measuring it.

from turboquant_pro.read_operators import (
    create_read_operator, consumer_distortion, error_covariance,
)

P = create_read_operator("readscope_jacobian", consumer=my_head).operator(keys)
sigma = error_covariance(keys, codec.decompress(codec.compress(keys)))
consumer_distortion(P, sigma)      # what the consumer actually loses

Two providers

Prefer readscope_jacobian where the consumer emits a vector: each probe direction then returns m numbers instead of one, at the same call cost.

The budget is not a tuning knob

readscope's calibration found that recovery against the direction budget k/d is a cliff at k = d, and that the cliff is rank independent — asking for one direction costs the same as asking for sixteen, because below full dimension the estimate is a projection onto a random subspace and a projected operator's leading eigenvector is not the operator's.

So both providers default to k = d and warn when given less, rather than quietly returning a one-direction reading that looks like a measurement. Full specification, including what is still unmeasured: readscope SPEC.md.

Neither package depends on the other

turboquant-pro declares the protocol. readscope measures the operator. This adapter is the only code that knows about both, and it is about a hundred lines. readscope stays numpy-only and usable by people with no interest in compression, which is worth more than the convenience of merging them.

Install

pip install -e plugins/tqp-readscope

Discovered automatically through the turboquant_pro.read_operators entry point. MIT.