szl-lambda-gate

Λ, a governance aggregator, as a pure-PyTorch (noarch) kernel. A differentiable, torch.compile-friendly weighted-geometric-mean aggregator with an ADVISORY non-compensatory gate, the strict szl.lambda/v1 entry point and runtime axiom self-checks, from SZL Holdings.

This repository is published from GitHub szl-holdings/szl-lambda-gate by its hf-kernel-publish workflow, which is its only writer. GitHub is the source of truth: every file under build/ is a byte copy of torch-ext/szl_lambda_gate at the commit named in SZL_KERNEL_SOURCE.json.

Load it

# pip install "kernels>=0.17.2" torch
import torch
from kernels import get_kernel

lg = get_kernel("SZLHOLDINGS/szl-lambda-gate", version=1, trust_remote_code=True)

axes = torch.tensor([0.9, 0.8, 0.95], dtype=torch.float64)   # axis scores in [0,1]
weights = torch.tensor([0.5, 0.25, 0.25], dtype=torch.float64)
gate = lg.lambda_v1_gate(axes, weights, tau=0.8)              # GO / NO_GO / ABSTAIN / BLOCK
print(gate.verdict, gate.code)

print(lg.lambda_aggregate(axes))   # legacy Λ(x) in [0,1]
print(lg.selfcheck())              # empirical A1-A4 checks, not a proof

• version=1 resolves the v1 branch, the stable API channel. It is republished on every release. The kernels library refuses a call that names neither a version nor a revision.
• trust_remote_code is required: SZLHOLDINGS is not a trusted kernel publisher on the Hub. The narrowest form is trust_remote_code=["SZLHOLDINGS/szl-lambda-gate"].
• CPU hosts load build/torch-cpu; CUDA and other hosts load build/torch-universal. Both variants carry identical code, built from the same GitHub commit.

Pin an exact build (production)

from huggingface_hub import HfApi
from kernels import get_kernel

refs = HfApi().list_repo_refs("SZLHOLDINGS/szl-lambda-gate", repo_type="kernel")
pin = next(b.target_commit for b in refs.branches if b.name == "v1")
lg = get_kernel("SZLHOLDINGS/szl-lambda-gate", revision=pin, trust_remote_code=["SZLHOLDINGS/szl-lambda-gate"])

Record pin (a commit of this kernel repository) in your lockfile. Commits of the model repository https://huggingface.co/SZLHOLDINGS/szl-lambda-gate are not kernel revisions and do not resolve here.

What Λ is, and is not

Λ is the weighted geometric mean over axis scores in [0,1]: Λ(x) = ∏ xᵢ^wᵢ with Σ wᵢ = 1, wᵢ > 0. It is a non-compensatory, ADVISORY roll-up: any zero (or, in the legacy entry, non-finite) axis drives the aggregate to 0. A gate pass is an advisory signal that the host must enforce; it is not proven trust. Λ uniqueness remains Conjecture 1 (open). The A1-A4 self-checks are empirical evidence on sampled inputs, not a proof.

API

Files

No speedup, throughput or energy figure is claimed for this kernel. Test results live in GitHub Actions for the source commit.

License and citation

Apache-2.0. Copyright 2026 SZL Holdings. Concept DOI 10.5281/zenodo.19944926; see CITATION.cff in the GitHub repository.