szl-lambda-gate / README.md
betterwithage's picture
mirror v0.2.2 from szl-holdings/szl-lambda-gate@1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63
5fb6a10 verified
|
Raw History Blame
24.6 kB
---
library_name: kernels
license: apache-2.0
tags:
- kernel
- governance
- lambda
- gate
- provenance
- torch
- surrogate
- pytorch
- doi:10.5281/zenodo.19944926
---
> **SOFTWARE / ADVISORY KERNEL · SURROGATE WEIGHTS UNAVAILABLE · NO PROMOTION**
>
> This model-type repository is a legacy source/card mirror for the weighted-geometric-mean
> kernel. The receipt-named `model.safetensors` and `config.json` are absent.
> [MODEL_PROVENANCE.json](./MODEL_PROVENANCE.json) records `trained_weights_present: false`;
> [TRAINING_RECEIPT.json](./TRAINING_RECEIPT.json) retains historical surrogate metrics.
> [OPERATIONAL.json](./OPERATIONAL.json) and [BENCH.laptop-blackwell.json](./BENCH.laptop-blackwell.json)
> retain the documented compile failures. No new surrogate release, deployment readiness,
> model promotion, or autonomous authority follows from this card.
>
> Λ uniqueness remains **Conjecture 1: OPEN and advisory**. A signature or `selfcheck`
> does not establish proven trust. The predecessor and compatibility relationship is
> [szl-governed-norm](https://huggingface.co/SZLHOLDINGS/szl-governed-norm);
> canonical kernel source is [szl-lambda-gate](https://github.com/szl-holdings/szl-lambda-gate).
> Kernel Hub migration references below describe a separate repository type.
> ### CORRECTION 2026-08-30 — the surrogate weights are NOT in this repo
>
> A callout below stated that this repository "also ships" `model.safetensors`, `config.json` with MEASURED
> fidelity **0.9670**. They are not here — every one returns **HTTP 404** on
> `resolve/main`. `MODEL_PROVENANCE.json` also asserted `trained_weights_present: true`
> with a sha256 for the missing file; that attestation has been corrected in the same
> commit and now reads `false`, with the digest retained as the *expected* value for
> when the weights are pushed.
>
> What remains true: the kernel is real, `get_kernel` is import-LIVE, and
> `TRAINING_RECEIPT.json` documents a genuine training run, so the fidelity figures
> keep their provenance. What was false: the claim that the resulting artifact is
> downloadable from this repo. **Do not build against the surrogate here — there is
> nothing to load.** The kernel was always the declared ground truth; that part of the
> card was correct and is unchanged.
<!-- SZL-KERNEL-OPERATIONAL:START -->
## Operational (MEASURED laptop-Blackwell)
> **STATUS:** tests **FAIL**. `get_kernel` **import-LIVE**. Unsloth/LoRA is the wrong tool. Receipted kernels, not silent CUDA.
| Thing | Label | Method / N / date / what-NOT |
|---|---|---|
| tests (`PYTHONPATH=torch-ext`) | **FAIL** | MEASURED 2026-08-29T15:53:47Z host `betterwithage` Windows-10-10.0.26200-SP0. torch `2.10.0+cu128`. GPU `NVIDIA GeForce RTX 5050 Laptop GPU` arch `Blackwell`. pytest `4 failed, 54 passed, 14 warnings in 52.57s`. Failed nodes: `tests/test_lambda.py::test_torch_compile_friendly; tests/test_lambda.py::test_fullgraph_lambda_aggregate; tests/test_lambda.py::test_fullgraph_lambda_gate_score; tests/test_lambda.py::test_fullgraph_lambda_gate_batch_score`. What-NOT: not a leaderboard. torch.compile fullgraph failures on Windows Blackwell (`cl is not found`) are MEASURED, not hidden. |
| Kernel Hub `get_kernel` | **import-LIVE** | kernels `0.16.1`. Default: `get_kernel("SZLHOLDINGS/szl-lambda-gate", revision="main", trust_remote_code=True)` → `True`. `backend="cpu"` → `True`. trust_remote_code=False → `ValueError` (SZLHOLDINGS is not a trusted publisher). repo_type=kernel required (kernels 0.16). What-NOT: not a weight load; do not pickle/joblib.load. |
| formula-tax | **ADVISORY** | locked-8 `F1 F4 F7 F11 F12 F18 F19 F22`. registry_count=21. Λ geomean `0.316227766016838`. uniqueness **Conjecture 1** (never a theorem). |
| I1–I8 | **catalog** | `I1 receipt-chain-continuity; I2 ledger-failure-shape; I3 served-run-has-model; I4 signed-columns-atomic; I5 loop-steps-positive; I6 receipt-ed25519-verify; I7 receipt-columns-consistent; I8 flywheel-lineage`. Executed by `SZLHOLDINGS/szl-invariants`. Statuses never coerced. Λ untouched. |
| CUDA speedup / tokens/s / joules | **UNAVAILABLE** | Not claimed. Receipted kernels, not silent CUDA. |
GitHub source: [`szl-holdings/szl-lambda-gate`](https://github.com/szl-holdings/szl-lambda-gate) @ `3fb5bb65dfdba006ae917e10b7703a535ea303e0`. Artifacts: [`BENCH.laptop-blackwell.json`](./BENCH.laptop-blackwell.json), [`OPERATIONAL.json`](./OPERATIONAL.json).
```python
from kernels import get_kernel
k = get_kernel("SZLHOLDINGS/szl-lambda-gate", revision="main", trust_remote_code=True)
```
<!-- SZL-KERNEL-OPERATIONAL:END -->
## Source-only CPU follow-up — 2026-09-24
An owner-run local check tested [GitHub source `f7e290a197922134076ff190118d7fca2e4975db`](https://github.com/szl-holdings/szl-lambda-gate/tree/f7e290a197922134076ff190118d7fca2e4975db)
on Windows 10 build 26200, Python 3.11.9, CPU PyTorch 2.14.0, and
MSVC 14.44.35207. No GPU was used. The unmodified source suite passed **62 tests, with no failures or skips**,
including all four compilation tests, after activating the installed MSVC
developer environment. A separate explicit run of the folded normalization
tests passed **110 tests, with no failures or skips**. Those
`governed_norm_test_*.py` files are not collected by default pytest discovery.
Kernel code and test thresholds were unchanged.
The [source-only test record](./SOURCE_CPU_TESTS.2026-09-24.json) names the exact
source, scope, run times, and retained log/JUnit files with their hashes. It is
unsigned owner-run evidence. The dated operational failures above remain
unchanged. These new results do **not** qualify the current Hub model build,
the separate first-class Kernel Hub repository, CUDA execution, production
deployment, or autonomous authority. Separate Hub-build probes were incomplete
after disk-space exhaustion and provide no runtime pass or diagnosed failure.
<!-- SZL-ESTATE-CARD:v2:START -->
<p align="center"><a href="https://a-11-oy.com/"><img src="https://huggingface.co/spaces/SZLHOLDINGS/README/resolve/main/assets/estate-banner-v2.svg" alt="SZL Holdings — governed, receipted, verifiable" width="100%"></a></p>
<p align="center">
<a href="https://github.com/szl-holdings/.github/tree/main/doctrine"><img src="https://img.shields.io/badge/doctrine-v11%20LOCKED-0B1F3A?style=flat-square" alt="doctrine v11"></a>
<a href="https://a-11-oy.com/"><img src="https://img.shields.io/badge/evidence%20wall-LIVE%20%C2%B7%20verify%20in%20browser-3AF4C8?style=flat-square" alt="live evidence wall"></a>
<a href="https://huggingface.co/datasets/SZLHOLDINGS/szl-lake"><img src="https://img.shields.io/badge/szl--lake-offline%20verifiable-C9B787?style=flat-square" alt="szl-lake offline verifiable"></a>
<a href="https://szlholdings-szl-constellation.hf.space/fabric"><img src="https://img.shields.io/badge/estate%20map-holographic-5B8DEE?style=flat-square" alt="holographic estate map"></a>
</p>
<p align="center"><sub>Part of the <a href="https://huggingface.co/SZLHOLDINGS">SZL Holdings</a> governed estate — claims are designed to carry checkable receipts. Verification proves integrity &amp; origin, never accuracy or performance.</sub></p>
<!-- SZL-ESTATE-CARD:v2:END -->
<!-- SZL-ARTIFACT-NOTICE:v1:START — honesty plate: repo semantics, no fake model tags. -->
> **🟥 Kernel real; surrogate weights NOT PUBLISHED — see the correction at the top.** The Λ governance kernel (pure-torch, differentiable) is UNCHANGED and remains the sole ground truth. Since **surrogate v1** this repo was described as shipping (IT DOES NOT — 404) `model.safetensors` + `config.json` — a real trained tiny torch MLP that predicts the ADVISORY gate decision `lambda_gate(axes, threshold).passed` over the 13-axis Yuyay space, with **MEASURED** fidelity **0.9670** (agreement vs the kernel on a held-out split). The surrogate approximates the gate DECISION only; the kernel Λ stays authoritative and `get_kernel`-discoverable. **Λ is the weighted geometric mean, NOT proven trust — uniqueness = Conjecture 1 (OPEN).**
<!-- SZL-ARTIFACT-NOTICE:v1:END -->
<p align="center">
<a href="https://huggingface.co/SZLHOLDINGS/szl-lambda-gate/tree/main/build/torch-universal/szl_lambda_gate"><img src="https://img.shields.io/badge/kernel%20hub-torch--universal-5b8dee?style=flat-square" alt="kernel hub"></a>
<a href="https://huggingface.co/SZLHOLDINGS/szl-lambda-gate/blob/main/MODEL_PROVENANCE.json"><img src="https://img.shields.io/badge/provenance-MODEL_PROVENANCE.json-3af4c8?style=flat-square" alt="provenance"></a>
<a href="https://huggingface.co/SZLHOLDINGS/szl-lambda-gate/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-7e8aa3?style=flat-square" alt="license"></a>
</p>
# szl-lambda-gate
> **Kernel Hub migration (verified 2026-07-15):** `get_kernel(...)` now resolves
> the matching first-class [Kernel Hub repository](https://huggingface.co/kernels/SZLHOLDINGS/szl-lambda-gate).
> Its `main` and stable `v1` refs both pin verified revision
> `47c7eb2db8859507d4115adbbff20c65de66dbb5`. This model-type repository is
> retained as the legacy source/card mirror.
**Λ — a governance aggregator as a Hugging Face kernel.** A differentiable weighted-geometric-mean aggregator with an ADVISORY non-compensatory gate and runtime axiom self-checks, from [SZL Holdings](https://huggingface.co/SZLHOLDINGS). Its torch.compile support is limited by the documented failures above.
> Companion to [`szl-governed-norm`](https://huggingface.co/SZLHOLDINGS/szl-governed-norm). Where that kernel makes a normalization *auditable*, this one makes a *governance decision* computable and checkable at the tensor layer.
## Interactive demo
<!-- SZL-ATELIER-CUT:v1:START -->
## The cut
A gate you can `import` in PyTorch. Fail-closed by default.
The smallest possible governed op.
### Silhouette → leave → SZL
| Leader | Take, then tweak |
|---|---|
| Anthropic | Refuse compiled to autograd. |
| NVIDIA | Custom op, NVIDIA-shaped packaging. |
| Unsloth | No. |
Nobody else ships this combination. That is the point of a one-of-one.
## Intended use
Torch forward-pass gate.
## Limitations
- Surrogate. Not Conjecture-1 solved.
Canonical GitHub: [`szl-holdings/szl-khipu`](https://github.com/szl-holdings/szl-khipu/blob/main/szl_khipu/lambda_gate.py)
<!-- SZL-ATELIER-CUT:v1:END -->
> **Live demos (in-browser, nothing to install)** — [`lambda-gate-holo`](https://szlholdings-lambda-gate-holo.static.hf.space) (this kernel's holographic gate demo) · [`szl-kernels-live`](https://szlholdings-szl-kernels-live.static.hf.space) (unified suite demo).
>
> The quickstart above runs fully locally. For a full governed-kernel suite demo, see [szl-kernels](https://huggingface.co/SZLHOLDINGS/szl-kernels). For the live a11oy substrate, see [a11oy Space](https://huggingface.co/spaces/SZLHOLDINGS/a11oy).
## What Λ is — and is NOT (read this first)
Λ is the **weighted geometric mean** over axis scores in [0,1]:
\[ \Lambda(x) = \prod_i x_i^{w_i}, \quad \sum_i w_i = 1, \; w_i > 0, \; x_i \in [0,1] \]
It is a **non-compensatory, ADVISORY** roll-up: any single zeroed (or non-finite) axis drives the whole aggregate to 0 — a conservative "one bad axis fails the gate" signal. **Λ is NOT "proven trust" and NOT a closed theorem.** Its *uniqueness* (that the weighted geometric mean is the only aggregator satisfying the carried axioms) remains **Conjecture 1 — OPEN**. A gate "pass" is an advisory signal, never a guarantee. We label this honestly everywhere.
## Quickstart
```bash
pip install kernels torch
```
```python
import torch
from kernels import get_kernel
# Current `kernels` (>=0.15) requires an explicit revision/version + trust flag for org kernels:
lg = get_kernel("SZLHOLDINGS/szl-lambda-gate", revision="main", trust_remote_code=True)
# (once a tag is published you can pin it, e.g. revision="v0.2.0")
axes = torch.tensor([0.9, 0.8, 0.95]) # axis scores in [0,1]
score = lg.lambda_aggregate(axes) # Λ(x) ∈ [0,1]
res = lg.lambda_gate(axes, threshold=0.5)
print(res.score, res.passed, res.advisory) # advisory is always True
print(lg.selfcheck()) # empirical A1–A4 checks + version
```
## API
| Function | Notes |
|---|---|
| `lambda_aggregate(axes, weights=None)` | Λ over the last dim. Differentiable, batched, torch.compile-friendly. |
| `lambda_gate(axes, weights=None, threshold=0.5)` | Advisory gate → `LambdaGateResult(score, passed, threshold, advisory)`. |
| `lambda_gate_batch(candidates, weights=None, threshold=0.5)` | Score many candidate vectors `(..., N, k)` in one call; returns the advisory pass mask. |
| `selfcheck()` | Empirical A1–A4 axiom checks + adversarial falsification search + version. NOT a uniqueness proof. |
| `is_monotone / is_homogeneous / is_egyptian_exact / is_bounded_by_max` | The four carried axioms as real runtime checks. |
| `yuyay_weights()`, `YUYAY_AXES`, `YUYAY_FLOORS` | Canonical 13-axis Yuyay preset (advisory). |
| layers: `LambdaGate`, `LambdaAggregate` | Pure `nn.Module` for the Kernel Hub layer-mapping mechanism. |
## Carried axioms (verifiable, not a proof)
- **A1 IsMonotone** — Λ is non-decreasing in each axis.
- **A2 IsHomogeneous (deg 1)** — Λ(t·x) = t·Λ(x).
- **A3 IsEgyptianExact** — Λ(c,…,c) = c.
- **A4 IsBounded** — Λ(x) ≤ maxᵢ xᵢ.
`selfcheck()` verifies these empirically on sampled inputs and runs a random falsification search. A clean run is **evidence, not proof** — Λ-uniqueness is Conjecture 1 (open).
## Provenance
Backed by the Lean 4 formalization [szl-holdings/lutar-lean](https://github.com/szl-holdings/lutar-lean) (749 declarations / 14 axioms / 163 tracked sorries), DOI [10.5281/zenodo.20434308](https://doi.org/10.5281/zenodo.20434308). Λ uniqueness = Conjecture 1 (open).
## Honesty
- Pure-Python universal kernel — a correctness reference, not a CUDA speed record. No fabricated benchmarks (50 passing tests).
- Λ is advisory; never "proven trust."
- Prior art honestly attributed: the weighted geometric mean as a less-compensatory composite indicator is established practice (UN HDI 2010, OECD Composite Indicators Handbook 2008); the veto/cut-off idea is ELECTRE. The 13-axis conjunctive form is SZL's own yuyay_v3 gate.
## Compatibility
Python 3.9+, `torch>=2.5`, standard library + torch only.
## License
Apache-2.0. Copyright 2026 SZL Holdings.
## Historical Λ-gate surrogate v1 (receipt-reported metrics; weights unavailable)
[TRAINING_RECEIPT.json](./TRAINING_RECEIPT.json) reports a **tiny torch MLP**
(3 hidden ReLU layers, 64 units; the absent `model.safetensors` + `config.json`)
trained on **40,000 axis-score vectors** synthesized and **labeled by this kernel itself**
(`lambda_gate(axes, weights=yuyay_uniform_1/13, threshold=0.5).passed`, seed 20260721; 800
samples re-audited by independent full kernel replay during generation — all agreed). Inputs are
the 13 Yuyay axis scores in [0,1], including non-compensatory zero-route rows (a single zeroed
axis must fail the gate).
| metric | value |
|---|---|
| fidelity vs kernel (held-out agreement) | **0.9670** |
| recall GATE_PASS | 0.9912 |
| recall GATE_FAIL | 0.9469 |
**Honest boundary:** the surrogate learns the *decision boundary* of an ADVISORY, non-compensatory
aggregator; it is a fast approximation, NOT the exact Λ and NOT proven trust. Residual disagreement
lives near the Λ=threshold surface — the exact kernel `lambda_gate` remains authoritative. Class
counts: GATE_FAIL=21828, GATE_PASS=18172. Λ uniqueness = Conjecture 1 (open).
**Historical evidence only — surrogate loading is unavailable.**
[TRAINING_RECEIPT.json](./TRAINING_RECEIPT.json) reports the metrics above and names
`model.safetensors` plus `config.json`; neither file is present in this repository.
[MODEL_PROVENANCE.json](./MODEL_PROVENANCE.json) preserves the expected weight digest
without attesting that the bytes are published. Do not attempt to load either
missing file. At the [revision inspected on 2026-09-24](https://huggingface.co/SZLHOLDINGS/szl-lambda-gate/blob/c6562ab0d4600cce5c26cd8e747e6b69176269ba/MODEL_PROVENANCE.json),
`claims.trained_model=SURROGATE_PRESENT_MEASURED_FIDELITY` contradicted the
false presence fields. The [2026-09-24 metadata correction](./MODEL_PROVENANCE_CORRECTION.2026-09-24.json)
changes only that label to `SURROGATE_UNAVAILABLE_RECEIPT_REPORTED_FIDELITY`.
Correction-file SHA-256: `731bd283b2ee7ae0732b8468b7c13e56e87d713a35d3e8a4d5e7e9d9cdeb49fb`. The immutable original
record and historical metrics remain available; this correction does not publish
or evaluate surrogate weights.
These historical surrogate metrics cannot be replayed from this
repository until the named artifacts are published and the weight hash is verified.
The retained `scripts/eval.py` is historical verification tooling, not a working
surrogate replay path for the current file set. The exact kernel remains the
reference for the advisory decision; the documented compile failures remain open.
---
## SZL Kernels Suite
Part of the [`szl-kernels`](https://huggingface.co/SZLHOLDINGS/szl-kernels) governed-kernel suite — the hub links every member, and each member links back to the hub so no leaf is orphaned:
| Kernel | Lane |
|---|---|
| [`szl-kernels`](https://huggingface.co/SZLHOLDINGS/szl-kernels) | **hub** — unified suite, cross-kernel `UnifiedReceiptChain` |
| [`szl-governed-norm`](https://huggingface.co/SZLHOLDINGS/szl-governed-norm) | RMSNorm/LayerNorm + SHA3-256 receipts |
| **`szl-lambda-gate`** (this repo) | **advisory Λ gate (Conjecture 1, OPEN)** |
| [`governed-inference-meter`](https://huggingface.co/SZLHOLDINGS/governed-inference-meter) | MEASURED-joule energy accounting (NVML) |
| [`szl-govsign`](https://huggingface.co/SZLHOLDINGS/szl-govsign) | signed governance attestation (DSSE / in-toto) |
| [`szl-blocked`](https://huggingface.co/SZLHOLDINGS/szl-blocked) | honest-BLOCKED state + EU AI Act Annex IV DRAFT |
| [`szl-provctl`](https://huggingface.co/SZLHOLDINGS/szl-provctl) | provenance-DAG verify + in-toto/SLSA interop |
**Live Spaces:** [a11oy](https://huggingface.co/spaces/SZLHOLDINGS/a11oy) · [hatun-mcp](https://huggingface.co/spaces/SZLHOLDINGS/hatun-mcp).
**Related — Governed Kernels collection:** [Governed Kernels & Verifiers](https://huggingface.co/collections/SZLHOLDINGS/governed-kernels-and-verifiers-6a542ad83a4b75151bf5eae3) groups the whole family in one page. **Live console:** [a11oy](https://szlholdings-a11oy.hf.space) · [a-11-oy.com](https://a-11-oy.com) · [llm-router](https://szlholdings-llm-router-live.hf.space) · [receipt verifier](https://szlholdings-governed-receipt-verifier.static.hf.space) · [receipt spec (hub)](https://github.com/szl-holdings/governed-receipt-spec).
---
<sub><b>SZL Holdings</b> · Λ governance aggregator · advisory, not proven trust · <a href="https://a-11-oy.com">a-11-oy.com</a> · <a href="https://github.com/szl-holdings">github.com/szl-holdings</a> · <a href="https://huggingface.co/SZLHOLDINGS">huggingface.co/SZLHOLDINGS</a></sub>
---
[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.19944926.svg)](https://doi.org/10.5281/zenodo.19944926)
## Citation
**Cite this.** Part of the SZL Holdings *Ouroboros Thesis* (Governed Post-Determinism).
Concept DOI (always-latest): [10.5281/zenodo.19944926](https://doi.org/10.5281/zenodo.19944926).
Author: Stephen P. Lutar Jr. · [ORCID 0009-0001-0110-4173](https://orcid.org/0009-0001-0110-4173) · License CC-BY-4.0.
Full DOI-pinned lineage (v1→v26) + the 8 papers: [szl-papers PAPERS_INDEX](https://github.com/szl-holdings/szl-papers/blob/main/PAPERS_INDEX.md).
No artifact-specific DOI is minted for this model; the concept DOI above covers the program.
Honesty (Doctrine v11): Λ unconditional uniqueness is **Conjecture 1** (machine-checked FALSE as stated) — never a theorem; conditional uniqueness is **Theorem U** (axiom-free). Locked-proven formulas = **exactly 8** {F1,F4,F7,F11,F12,F18,F19,F22}; ~185 experimental theorems are a separate CI-green tier; Khipu BFT safety = Conjecture 2. Trust never 100%.
```bibtex
@misc{lutar_szl_ouroboros,
author = {Lutar, Stephen P., Jr.},
title = {SZL Holdings --- The Ouroboros Thesis (Governed Post-Determinism)},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.19944926},
url = {https://doi.org/10.5281/zenodo.19944926},
note = {Concept DOI --- always resolves to the latest version. ORCID 0009-0001-0110-4173. CC-BY-4.0.}
}
```
*Signed-off-by: Stephen Lutar <stephenlutar2@gmail.com>*
## Files in this repo
| Path | What it is |
|---|---|
| `build/torch-universal/szl_lambda_gate/__init__.py` | public API — `lambda_aggregate`, `lambda_gate`, `selfcheck()` |
| `build/torch-universal/szl_lambda_gate/_lambda.py` | the weighted-geometric-mean aggregator + A1–A4 empirical checks |
| `build/torch-universal/szl_lambda_gate/layers.py` | `nn.Module` wrapper |
| `build.toml` · `build/torch-universal/szl_lambda_gate/metadata.json` | Kernel Hub build/metadata manifests |
| `LICENSE` · `SECURITY.md` | Apache-2.0 · security policy |
---
<p align="center">
<a href="https://huggingface.co/SZLHOLDINGS">SZL Holdings</a> ·
<a href="https://a-11-oy.com">a-11-oy.com</a> ·
<a href="https://huggingface.co/SZLHOLDINGS/szl-kernels">szl-kernels</a>
</p>
<p align="center"><sub>SLSA: L1 honest · L2 attested · L3 roadmap. Λ = Conjecture 1 (advisory, never a theorem). Trust ceiling 0.97 — never 100%. Labels honest by default: MEASURED / REPORTED / MODELED / HEURISTIC / UNKNOWN / UNAVAILABLE. locked-proven = exactly 8 {F1,F4,F7,F11,F12,F18,F19,F22}.</sub></p>
<!-- SZL-ESTATE-WIRING-V2 -->
## Estate wiring
[Source](https://github.com/szl-holdings/szl-lambda-gate) | [Estate hologram](https://szlholdings-szl-constellation.hf.space/fabric) | [Receipt ledger](https://huggingface.co/datasets/SZLHOLDINGS/szl-lake) | [Product](https://a-11-oy.com) | [Proof](https://a11oy.net)
Verification proves integrity and declared origin only; it does not prove accuracy, readiness, or performance.
<!-- SZL-HF-MIRROR:START -->
## Mirrored release v0.2.2
# SZL Lambda Gate v0.2.2
Release-tooling patch from canonical GitHub source commit `1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63`.
- Propose new Hugging Face releases through an additive, exact-byte-verified Hub pull request instead of attempting a direct commit to protected `main`.
- Bind a reviewed Hub PR merge to its source, proposal artifact, complete file hashes, and release assets before creating an immutable tag and measured publication receipt.
- Audit the latest measured Hub release with read-only OIDC verification.
The kernel runtime version remains 0.2.0. This release changes publication controls and does not claim new model qualification. GitHub publication starts the SBOM producer and Hub proposal; Hub publication remains pending until the Hub PR is reviewed, merged, finalized, and independently read back.
| Field | Value |
|---|---|
| GitHub source | https://github.com/szl-holdings/szl-lambda-gate |
| GitHub release | https://github.com/szl-holdings/szl-lambda-gate/releases/tag/v0.2.2 |
| Source commit | `1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63` |
| Hub revision | `v0.2.2` |
### szl.lambda/v1
| Field | Value |
|---|---|
| Contract | https://github.com/szl-holdings/szl-lambda-gate/blob/1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63/spec/szl.lambda.v1.json |
| Golden vectors (60) | https://github.com/szl-holdings/szl-lambda-gate/blob/1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63/spec/lambda_v1_vectors.json |
| Vectors SHA-256, canonical JSON (recorded in the contract) | `61bfb0410b9f0eaab0eb9f22f29cb7cb13cfde8c083fe308d895565d6ba9ebd4` |
| Vectors SHA-256, file bytes in this revision | `460bb966ff2bec4fd99e03d267b721f22c5fad731e0667723407a6909bd4dcec` |
| Stdlib reference | https://github.com/szl-holdings/szl-lambda-gate/blob/1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63/reference/szl_lambda_v1.py |
| Admit contract (τ) | https://github.com/szl-holdings/szl-lambda-gate/blob/1dab9bbe0f8c702f7ff8af4a426bb4ca2459fc63/frontier/model_admit_contract.v1.json |
Λ uniqueness: Conjecture 1 (open); nothing in the contract depends on it.
This is an advisory kernel; the gate is floors ∧ Λ with τ from the admit contract.
Mirror direction: GitHub is authoritative for code; Hub-only artifacts in this
repo are preserved and never deleted by the mirror. Truth state is MEASURED only
after the workflow re-reads this repo from the Hub.
<!-- SZL-HF-MIRROR:END -->