betterwithage commited on
Commit
06cc46f
·
verified ·
1 Parent(s): fd819c5

Publish verified torch-cpu kernel v1

Browse files

Dedicated Kernel Hub migration from legacy model commit 883050fab65a1bb2096c77526dbfdabaf3fc4a3a. Locally loaded with kernels 0.16.0 on CPU and passed its bounded verification.

LICENSE ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ Copyright 2026 SZL Holdings
6
+
7
+ Licensed under the Apache License, Version 2.0 (the "License");
8
+ you may not use this file except in compliance with the License.
9
+ You may obtain a copy of the License at
10
+
11
+ http://www.apache.org/licenses/LICENSE-2.0
12
+
13
+ Unless required by applicable law or agreed to in writing, software
14
+ distributed under the License is distributed on an "AS IS" BASIS,
15
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16
+ See the License for the specific language governing permissions and
17
+ limitations under the License.
18
+
19
+ The full Apache License 2.0 text is available at the URL above.
README.md ADDED
@@ -0,0 +1,273 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - kernel
4
+ - governance
5
+ - provenance
6
+ - suite
7
+ - doi:10.5281/zenodo.19944926
8
+ library_name: kernels
9
+ license: apache-2.0
10
+ szl-governance:
11
+ verdict: ADVISORY
12
+ lambda: "Conjecture 1 (open) — uniqueness unproven; advisory only"
13
+ energy: MEASURED-only (real NVML delta; None when unavailable)
14
+ provenance: UnifiedReceiptChain (SHA3-256, op-agnostic, cross-kernel)
15
+ honest_blocked: "a failed check stays failed — never faked green"
16
+ ---
17
+
18
+ <p align="center">
19
+ <a href="https://huggingface.co/SZLHOLDINGS/szl-kernels/tree/main/build/torch-universal/szl_kernels"><img src="https://img.shields.io/badge/kernel%20hub-torch--universal-5b8dee?style=flat-square" alt="kernel hub"></a>
20
+ <a href="https://huggingface.co/SZLHOLDINGS/szl-kernels/blob/main/MODEL_PROVENANCE.json"><img src="https://img.shields.io/badge/provenance-MODEL_PROVENANCE.json-3af4c8?style=flat-square" alt="provenance"></a>
21
+ <a href="https://huggingface.co/SZLHOLDINGS/szl-kernels/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-7e8aa3?style=flat-square" alt="license"></a>
22
+ </p>
23
+
24
+ # szl-kernels — the unified governed-kernel suite
25
+
26
+ **The first kernel that governs provenance *across* ops, not just within one.** A `get_kernel`-discoverable suite that ties SZL Holdings' three governed kernels — [`szl-governed-norm`](https://huggingface.co/SZLHOLDINGS/szl-governed-norm), [`szl-lambda-gate`](https://huggingface.co/SZLHOLDINGS/szl-lambda-gate), and [`governed-inference-meter`](https://huggingface.co/SZLHOLDINGS/governed-inference-meter) — into **one shared, hash-chained `UnifiedReceiptChain`**, and anchors a governance/interop layer on top: [`szl-govsign`](https://huggingface.co/SZLHOLDINGS/szl-govsign) (signs the verdict), [`szl-blocked`](https://huggingface.co/SZLHOLDINGS/szl-blocked) (refuses honestly + derives EU AI Act Annex IV), and [`szl-provctl`](https://huggingface.co/SZLHOLDINGS/szl-provctl) (verifies the provenance DAG + bridges to in-toto/SLSA).
27
+
28
+ > Every Kernel Hub leader competes on **FLOPs per op**. They do not sign their build artifacts, they do not surface an honest BLOCKED verdict, and they do not stitch provenance *across* ops. `szl-kernels` opens that lane: a real forward pass touching norm + an advisory Λ gate + an energy reading produces **one auditable, tamper-evident log** — not three disconnected ones.
29
+
30
+ ## Quickstart
31
+
32
+ ```bash
33
+ pip install kernels torch
34
+ ```
35
+
36
+ ```python
37
+ import torch
38
+ from kernels import get_kernel
39
+
40
+ # Current `kernels` (>=0.15) requires an explicit revision/version + trust flag for org kernels:
41
+ suite = get_kernel("SZLHOLDINGS/szl-kernels", revision="main", trust_remote_code=True)
42
+
43
+ print(suite.list_kernels()) # the 3 numeric suite members + honest roles
44
+ print(suite.list_series()) # the governance/interop companions (govsign, blocked, provctl)
45
+ print(suite.selfcheck()) # one-shot CPU health: ALL checks pass
46
+
47
+ # ONE shared chain spanning multiple ops:
48
+ chain = suite.UnifiedReceiptChain()
49
+ x = torch.randn(4, 64)
50
+ y = suite.governed_rms_norm(chain, x, eps=1e-6) # governed_norm
51
+ gate = suite.governed_lambda_gate(chain, torch.tensor([0.9,0.8,0.95])) # lambda_gate (advisory)
52
+ e = suite.governed_measure_energy(chain) # energy_core (MEASURED-only)
53
+
54
+ ok, depth, brk = chain.verify() # the WHOLE pass verifies as ONE chain
55
+ print(ok, depth, chain.kernels_touched()) # True 3 ['governed_norm','lambda_gate','energy_core']
56
+ print(chain.to_json()) # export for offline third-party re-verification
57
+ ```
58
+
59
+ ### Flagship — a governed transformer sub-block
60
+
61
+ ```python
62
+ blk = suite.GovernedBlock()
63
+ res = blk.forward(x, gov_axes=torch.tensor([0.95, 0.9, 0.92]))
64
+ print(res["chain_ok"], res["chain_depth"], res["kernels_touched"])
65
+ # norm + advisory Λ gate + energy + binding receipt = 4 ops, one verifiable chain.
66
+ # The Λ gate is ADVISORY: it is recorded for audit, it does NOT alter the numerics.
67
+ ```
68
+
69
+ ## Cookbook
70
+
71
+ Three copy-paste recipes spanning the governed-kernel series. Every printed value is
72
+ labeled **expected shape (not executed here)** — the shapes are transcribed from each
73
+ kernel's committed API, not from a run on this card (SZL doctrine: never self-download to
74
+ inflate counters, never fabricate an output). Λ stays **Conjecture 1 (OPEN)**; energy stays
75
+ **MEASURED-only**; a BLOCKED verdict stays BLOCKED.
76
+
77
+ ### 1 — One receipt chain across three ops (suite)
78
+
79
+ ```python
80
+ import torch
81
+ from kernels import get_kernel
82
+
83
+ suite = get_kernel("SZLHOLDINGS/szl-kernels", revision="main", trust_remote_code=True)
84
+
85
+ chain = suite.UnifiedReceiptChain()
86
+ x = torch.randn(4, 64)
87
+ suite.governed_rms_norm(chain, x, eps=1e-6) # op 1: governed_norm
88
+ suite.governed_lambda_gate(chain, torch.tensor([0.9, 0.8, 0.95])) # op 2: lambda_gate (ADVISORY)
89
+ suite.governed_measure_energy(chain) # op 3: energy_core (MEASURED-only)
90
+
91
+ ok, depth, first_break = chain.verify()
92
+ print(ok, depth, chain.kernels_touched())
93
+ # expected shape (not executed here):
94
+ # True 3 ['governed_norm', 'lambda_gate', 'energy_core']
95
+ # -> one hash-chain, three ops, verifies as ONE ordered sequence.
96
+ # The Λ gate receipt is ADVISORY (Conjecture 1, OPEN): recorded, never proven trust.
97
+ # energy_core reports joules=None + UNAVAILABLE_NO_NVML on CPU — never a fabricated joule.
98
+ ```
99
+
100
+ ### 2 — honest-BLOCKED, not fake-green (szl-blocked)
101
+
102
+ ```python
103
+ from kernels import get_kernel
104
+
105
+ blk = get_kernel("SZLHOLDINGS/szl-blocked", revision="main", trust_remote_code=True)
106
+
107
+ chain = blk.UnifiedReceiptChain()
108
+ policy = blk.deny_if_action_in({"exfiltrate", "delete_all"})
109
+ work = lambda v: v * 2
110
+
111
+ allowed = blk.governed_call(work, policy, chain, request={"action": "summarize"}, args=(21,))
112
+ blocked = blk.governed_call(work, policy, chain, request={"action": "exfiltrate"}, args=(21,))
113
+
114
+ print(allowed.blocked, allowed.output)
115
+ print(blocked.blocked, blocked.output)
116
+ # expected shape (not executed here):
117
+ # False 42 -> ALLOWED path ran work(21); an ALLOW receipt is on the chain.
118
+ # True None -> BLOCKED path: work was NEVER called, output is None,
119
+ # a BLOCK receipt is recorded. Honest-BLOCKED, never faked green.
120
+ ```
121
+
122
+ ### 3 — Sign then verify a governance verdict (szl-govsign / DSSE)
123
+
124
+ ```python
125
+ from kernels import get_kernel
126
+
127
+ gs = get_kernel("SZLHOLDINGS/szl-govsign", revision="main", trust_remote_code=True)
128
+
129
+ priv = gs.generate_ephemeral_keypair() # production: Sigstore keyless / cosign key, out-of-band
130
+ pred = gs.build_governance_predicate(
131
+ lambda_verdict = gs.LambdaVerdict(score=0.92, notes="advisory only — Conjecture 1 (OPEN)"),
132
+ energy = gs.EnergyLabel(value=12.5, unit="joules"), # MEASURED-only
133
+ decision = gs.GovernanceDecision(status="ALLOWED", reason="passed gates"),
134
+ honest_blocked = False,
135
+ )
136
+ subjects = [gs.Subject(name="szl_kernels/UnifiedReceiptChain", digest={"sha256": "<chain-head>"})]
137
+ envelope = gs.attest(subjects, pred, priv)
138
+
139
+ print(gs.verify(envelope, priv.public_key()))
140
+ # expected shape (not executed here):
141
+ # True -> DSSE envelope (ECDSA P-256) verifies: authorship + integrity of the verdict.
142
+ # Any tamper -> verify() returns False (fails closed).
143
+ # The signature does NOT upgrade Λ to proven trust: proven_trust is locked False.
144
+ ```
145
+
146
+ > These recipes chain across three separately published, `get_kernel`-discoverable kernels.
147
+ > See [`szl-provctl`](https://huggingface.co/SZLHOLDINGS/szl-provctl) to turn any of these
148
+ > chains into a spec-exact in-toto v1 / SLSA v1 statement the wider supply-chain world verifies.
149
+
150
+ ## The governed-kernel series
151
+
152
+ Independently published, `get_kernel`-discoverable kernels that share one `UnifiedReceiptChain`. The first three are the **numeric core**; govsign + blocked + provctl are the **governance / interop layer**.
153
+
154
+ | Kernel | Lane | Live hologram |
155
+ |---|---|---|
156
+ | [`szl-governed-norm`](https://huggingface.co/SZLHOLDINGS/szl-governed-norm) | RMSNorm/LayerNorm + SHA3-256 receipts | [`governed-norm-holo`](https://szlholdings-governed-norm-holo.static.hf.space) ✅ **live** |
157
+ | [`szl-lambda-gate`](https://huggingface.co/SZLHOLDINGS/szl-lambda-gate) | advisory Λ gate (Conjecture 1, OPEN) | [`lambda-gate-holo`](https://szlholdings-lambda-gate-holo.static.hf.space) ✅ **live** |
158
+ | [`governed-inference-meter`](https://huggingface.co/SZLHOLDINGS/governed-inference-meter) | MEASURED-joule energy accounting | [`energy-attest-holo`](https://szlholdings-energy-attest-holo.static.hf.space) ✅ **live** |
159
+ | [`szl-govsign`](https://huggingface.co/SZLHOLDINGS/szl-govsign) | signed governance attestation (DSSE / in-toto, ECDSA P-256) | [`szl-govsign-live`](https://szlholdings-szl-govsign-live.static.hf.space) ✅ **live** |
160
+ | [`szl-blocked`](https://huggingface.co/SZLHOLDINGS/szl-blocked) | honest-BLOCKED first-class state + EU AI Act Annex IV DRAFT | [`szl-blocked-live`](https://szlholdings-szl-blocked-live.static.hf.space) ✅ **live** |
161
+ | [`szl-provctl`](https://huggingface.co/SZLHOLDINGS/szl-provctl) | provenance-DAG verify + in-toto v1 / SLSA v1 interop + per-kernel MEASURED energy | [`szl-provctl-live`](https://szlholdings-szl-provctl-live.static.hf.space) ✅ **live** |
162
+ | **`szl-kernels`** (this repo) | **unified suite — cross-kernel `UnifiedReceiptChain`** | [`szl-kernels-live`](https://szlholdings-szl-kernels-live.static.hf.space) ✅ **live** |
163
+
164
+ `suite.list_kernels()` returns the numeric core; `suite.list_series()` returns the govsign + blocked + provctl governance/interop layer.
165
+
166
+ ### The honest-model trio (offline replays of the live Alloy surface)
167
+
168
+ Published as HF **model** repos (NOT trained models, NO weights — pure-Python, stdlib-only offline replays). Each ships a `library_name: kernels` card and MEASURED local test counts:
169
+
170
+ | Model | Lane | Tests (MEASURED) |
171
+ |---|---|---|
172
+ | [`szl-invariants`](https://huggingface.co/SZLHOLDINGS/szl-invariants) | 8 falsifiable receipt/ledger invariants, offline | 14/14 |
173
+ | [`szl-ouroboros`](https://huggingface.co/SZLHOLDINGS/szl-ouroboros) | bounded-loop trace + MEASURED/DERIVED loop-tax accounting | 13/13 |
174
+ | [`szl-formulas`](https://huggingface.co/SZLHOLDINGS/szl-formulas) | the 21 canonical formulas + governed-loop composer, PROOF-STATUS mirrored verbatim (locked-proven = exactly 8) | 17/17 |
175
+
176
+ ## The gap this closes
177
+
178
+ Today even SZL's own governance is fragmented: `szl-governed-norm` keeps its own receipt chain, the energy meter keeps its own ledger, and `szl-lambda-gate` keeps none. So a single forward pass yields three logs no third party can re-walk as one ordered, tamper-evident sequence. `UnifiedReceiptChain` is that missing artifact — op-agnostic SHA3-256 receipts that hash-chain norm, Λ, and energy calls into **one** verifiable stream, in call order. `szl-govsign` then makes that chain head third-party-verifiable; `szl-blocked` makes a refusal a recorded, first-class state and derives the compliance paperwork from it; `szl-provctl` verifies the whole multi-run provenance DAG and bridges it to the in-toto/SLSA formats the rest of the supply-chain world reads.
179
+
180
+ ## API
181
+
182
+ | Symbol | What it does |
183
+ |---|---|
184
+ | `UnifiedReceiptChain` | Op-agnostic SHA3-256 hash chain. `emit`, `verify() -> (ok, depth, first_break)`, `kernels_touched()`, `to_json()`, `verify_json()` (offline). |
185
+ | `governed_rms_norm(chain, x, weight=None, eps=1e-6)` | RMSNorm + a receipt into the shared chain. Numerics match `szl-governed-norm`. |
186
+ | `governed_layer_norm(chain, x, ...)` | LayerNorm + receipt. |
187
+ | `governed_lambda_gate(chain, axes, weights=None, threshold=0.5)` | **Advisory** Λ gate; records an advisory receipt (`advisory=True`, never proven trust). |
188
+ | `governed_measure_energy(chain, measurement=None)` | Records an energy reading **verbatim** — `joules=None` + `UNAVAILABLE_NO_NVML` when no GPU. Never fabricated. |
189
+ | `GovernedBlock` | Pre-norm sub-block composing all three + a binding receipt into one auditable pass. |
190
+ | `list_kernels()`, `list_series()`, `get_member()`, `selfcheck()` | Numeric registry + governance-layer series + one-shot CPU health check. |
191
+
192
+ ## Honesty (SZL doctrine)
193
+
194
+ - **Λ is advisory.** Its uniqueness is **Conjecture 1 — OPEN**. A recorded gate "pass" is a non-compensatory advisory signal, **never proven trust**.
195
+ - **Energy is MEASURED-only.** Real NVML cumulative-energy delta when a GPU is present; otherwise `joules=None`, labeled `UNAVAILABLE_NO_NVML`. **No joule is ever fabricated.**
196
+ - **The digest is an integrity fingerprint, not a signature.** SHA3-256 over a canonical receipt body proves tamper-evidence + ordering — not authorship. Signing is a separate, out-of-band layer — see [`szl-govsign`](https://huggingface.co/SZLHOLDINGS/szl-govsign) for DSSE / in-toto attestation.
197
+ - **Honest BLOCKED beats fake green.** A failed verification stays failed — see [`szl-blocked`](https://huggingface.co/SZLHOLDINGS/szl-blocked) for refusal as a first-class, provenanced state.
198
+ - **Universal (pure-Python) suite — a correctness + provenance reference, not a CUDA speed record. No fabricated benchmarks.** Suite tests: 7/7 passing.
199
+
200
+ ## Provenance
201
+
202
+ 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).
203
+
204
+ ## See it live
205
+
206
+ - ✅ **Live now:** [a11oy](https://huggingface.co/spaces/SZLHOLDINGS/a11oy) (live governed inference) · [hatun-mcp](https://huggingface.co/spaces/SZLHOLDINGS/hatun-mcp). **All eight demo Spaces below are live** (static, in-browser).
207
+ - 📚 **Collection:** [Governed Kernels — verifiable AI building blocks](https://huggingface.co/collections/SZLHOLDINGS/governed-kernels-verifiable-ai-building-blocks-6a41d3936cfce4fba83ce378) — 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).
208
+ - ✅ Suite: [`szl-kernels-live`](https://szlholdings-szl-kernels-live.static.hf.space) ✅ **live** — holographic cross-kernel provenance graph with in-browser SHA3-256 + tamper / honest-BLOCKED demo.
209
+ - ✅ Members: [`governed-norm-holo`](https://szlholdings-governed-norm-holo.static.hf.space) ✅ **live** · [`lambda-gate-holo`](https://szlholdings-lambda-gate-holo.static.hf.space) ✅ **live** · [`energy-attest-holo`](https://szlholdings-energy-attest-holo.static.hf.space) ✅ **live** · [`receipt-chain-live`](https://szlholdings-receipt-chain-live.static.hf.space) ✅ **live**
210
+ - ✅ Governance layer: [`szl-govsign-live`](https://szlholdings-szl-govsign-live.static.hf.space) ✅ **live** · [`szl-blocked-live`](https://szlholdings-szl-blocked-live.static.hf.space) ✅ **live** · [`szl-provctl-live`](https://szlholdings-szl-provctl-live.static.hf.space) ✅ **live**
211
+ - 🔮 `szl-substrate` *(ROADMAP — not yet live)* — the hub tying the whole governed-compute substrate together.
212
+
213
+ ## Compatibility
214
+
215
+ Python 3.9+, `torch>=2.5`, standard library + torch only. Runs on CPU and CUDA.
216
+
217
+ ## License
218
+
219
+ Apache-2.0. Copyright 2026 SZL Holdings.
220
+
221
+ ---
222
+
223
+ <sub><b>SZL Holdings</b> · unified governed-kernel suite · cross-kernel provenance · Λ advisory (Conjecture 1) · energy MEASURED-only · <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>
224
+
225
+ ---
226
+
227
+ [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.19944926.svg)](https://doi.org/10.5281/zenodo.19944926)
228
+
229
+ ## Citation
230
+
231
+
232
+ **Cite this.** Part of the SZL Holdings *Ouroboros Thesis* (Governed Post-Determinism).
233
+ Concept DOI (always-latest): [10.5281/zenodo.19944926](https://doi.org/10.5281/zenodo.19944926).
234
+ Author: Stephen P. Lutar Jr. · [ORCID 0009-0001-0110-4173](https://orcid.org/0009-0001-0110-4173) · License CC-BY-4.0.
235
+ 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).
236
+ No artifact-specific DOI is minted for this model; the concept DOI above covers the program.
237
+
238
+ 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%.
239
+
240
+ ```bibtex
241
+ @misc{lutar_szl_ouroboros,
242
+ author = {Lutar, Stephen P., Jr.},
243
+ title = {SZL Holdings --- The Ouroboros Thesis (Governed Post-Determinism)},
244
+ year = {2026},
245
+ publisher = {Zenodo},
246
+ doi = {10.5281/zenodo.19944926},
247
+ url = {https://doi.org/10.5281/zenodo.19944926},
248
+ note = {Concept DOI --- always resolves to the latest version. ORCID 0009-0001-0110-4173. CC-BY-4.0.}
249
+ }
250
+ ```
251
+
252
+ *Signed-off-by: Stephen Lutar <stephenlutar2@gmail.com>*
253
+
254
+ ## Files in this repo
255
+
256
+ | Path | What it is |
257
+ |---|---|
258
+ | `build/torch-universal/szl_kernels/__init__.py` | public API — suite entry points + `selfcheck()` |
259
+ | `build/torch-universal/szl_kernels/_chain.py` | cross-kernel `UnifiedReceiptChain` (SHA3-256) |
260
+ | `build/torch-universal/szl_kernels/_ops.py` | the governed op set |
261
+ | `tests/test_suite.py` | suite test |
262
+ | `build.toml` · `metadata.json` | Kernel Hub build/metadata manifests |
263
+ | `LICENSE` · `SECURITY.md` | Apache-2.0 · security policy |
264
+
265
+ ---
266
+
267
+ <p align="center">
268
+ <a href="https://huggingface.co/SZLHOLDINGS">SZL Holdings</a> ·
269
+ <a href="https://a-11-oy.com">a-11-oy.com</a> ·
270
+ <a href="https://huggingface.co/SZLHOLDINGS/a11oy-v19-substrate">a11oy-v19-substrate</a>
271
+ </p>
272
+
273
+ <p align="center"><sub>SLSA: L1 honest · L2 attested · L3 roadmap. Λ = Conjecture 1. Trust ceiling 0.97.</sub></p>
build/torch-cpu/metadata.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "szl-kernels",
3
+ "id": "_szl_kernels_cpu_0d1a832737",
4
+ "version": 1,
5
+ "license": "Apache-2.0",
6
+ "python-depends": [],
7
+ "backend": {
8
+ "type": "cpu"
9
+ }
10
+ }
build/torch-cpu/szl_kernels/__init__.py ADDED
@@ -0,0 +1,257 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: Apache-2.0
2
+ # © 2026 SZL Holdings · Stephen P. Lutar · ORCID 0009-0001-0110-4173
3
+ """szl_kernels — SZL Holdings' unified governed-kernel SUITE.
4
+
5
+ ONE get_kernel-discoverable package that ties the three governed kernels plus
6
+ the energy meter into a single suite with a SHARED, op-agnostic
7
+ ``UnifiedReceiptChain`` — cross-kernel provenance in one auditable forward pass.
8
+ This is the gap the Kernel Hub leaders leave open: they optimize FLOPs per op;
9
+ SZL governs provenance ACROSS ops.
10
+
11
+ Suite members (each independently published as its own Hub kernel):
12
+ * governed_norm -> SZLHOLDINGS/szl-governed-norm (RMSNorm/LayerNorm + receipts)
13
+ * lambda_gate -> SZLHOLDINGS/szl-lambda-gate (advisory Λ gate; Conjecture 1)
14
+ * energy_core -> szl_energy_core / governed-inference-meter (MEASURED joules)
15
+ * govsign -> SZLHOLDINGS/szl-govsign (signed governance attestation; DSSE / in-toto)
16
+ * blocked -> SZLHOLDINGS/szl-blocked (honest-BLOCKED first-class state + EU AI Act Annex IV)
17
+ * provctl -> SZLHOLDINGS/szl-provctl (provenance-DAG verify + in-toto v1 / SLSA v1 interop + per-kernel MEASURED energy)
18
+
19
+ The first three are the NUMERIC core exercised by the shared forward pass.
20
+ govsign + blocked are GOVERNANCE-LAYER series companions: govsign makes a
21
+ verdict third-party-verifiable (sign/verify offline), and blocked makes a
22
+ refusal a first-class, provenanced state (the op never runs — no fake-green)
23
+ and derives an EU AI Act Annex IV DRAFT skeleton from the same chain.
24
+
25
+ Discover + use the suite:
26
+
27
+ from kernels import get_kernel
28
+ suite = get_kernel("SZLHOLDINGS/szl-kernels", revision="main", trust_remote_code=True)
29
+ print(suite.list_kernels()) # {'governed_norm':..., 'lambda_gate':..., 'energy_core':...}
30
+ print(suite.list_series()) # {'govsign':..., 'blocked':..., 'provctl':...}
31
+ print(suite.selfcheck()) # one-shot suite health (CPU only)
32
+
33
+ # ONE shared chain spanning multiple ops:
34
+ chain = suite.UnifiedReceiptChain()
35
+ import torch
36
+ x = torch.randn(4, 64)
37
+ y = suite.governed_rms_norm(chain, x, eps=1e-6)
38
+ gate = suite.governed_lambda_gate(chain, torch.tensor([0.9, 0.8, 0.95]), threshold=0.5)
39
+ e = suite.governed_measure_energy(chain) # None joules when no NVML — never faked
40
+ ok, depth, brk = chain.verify() # the WHOLE pass verifies as one chain
41
+ print(chain.kernels_touched()) # ['governed_norm','lambda_gate','energy_core']
42
+
43
+ # Flagship composition — a governed transformer sub-block:
44
+ blk = suite.GovernedBlock()
45
+ res = blk.forward(x, gov_axes=torch.tensor([0.95, 0.9, 0.92]))
46
+ print(res["chain_ok"], res["chain_depth"], res["kernels_touched"])
47
+
48
+ HONESTY (SZL doctrine v11):
49
+ * Λ is the weighted-geometric-mean ADVISORY aggregator; its uniqueness is
50
+ Conjecture 1 (OPEN). A recorded gate "pass" is advisory, NEVER proven trust.
51
+ * Energy joules are MEASURED-only (real NVML delta). No GPU/NVML => None /
52
+ UNAVAILABLE_NO_NVML, NEVER a fabricated joule.
53
+ * Receipt digests are SHA3-256 integrity fingerprints — tamper-evidence +
54
+ ordering, NOT signatures and NOT proof of authorship.
55
+ * This is a UNIVERSAL (pure-Python) suite: a correctness + provenance
56
+ reference, not a hand-tuned CUDA speed record. No fabricated benchmarks.
57
+ """
58
+ from typing import Any, Dict
59
+
60
+ import torch
61
+
62
+ from ._chain import GENESIS, UnifiedReceiptChain, tensor_digest
63
+ from ._ops import (
64
+ GovernedBlock,
65
+ governed_layer_norm,
66
+ governed_lambda_gate,
67
+ governed_measure_energy,
68
+ governed_rms_norm,
69
+ )
70
+
71
+ __all__ = [
72
+ "UnifiedReceiptChain",
73
+ "tensor_digest",
74
+ "GENESIS",
75
+ "governed_rms_norm",
76
+ "governed_layer_norm",
77
+ "governed_lambda_gate",
78
+ "governed_measure_energy",
79
+ "GovernedBlock",
80
+ "list_kernels",
81
+ "list_series",
82
+ "get_member",
83
+ "selfcheck",
84
+ "DOCTRINE_FOOTER",
85
+ "PROVENANCE",
86
+ "__version__",
87
+ ]
88
+
89
+ __version__ = "0.1.0"
90
+
91
+ DOCTRINE_FOOTER = (
92
+ "SZL Holdings · unified governed-kernel suite · cross-kernel provenance · "
93
+ "Λ = Conjecture 1 (advisory) · energy MEASURED-only · honesty over checklist"
94
+ )
95
+
96
+ # Suite registry: maps a suite-member key to its Hub coordinates + role. In
97
+ # production, get_member() would resolve these via kernels.get_kernel(); the
98
+ # reference package ships byte-faithful numerics so it runs standalone.
99
+ _REGISTRY: Dict[str, Dict[str, str]] = {
100
+ "governed_norm": {
101
+ "hub_id": "SZLHOLDINGS/szl-governed-norm",
102
+ "role": "RMSNorm/LayerNorm + SHA3-256 hash-chained receipts",
103
+ "honesty": "integrity fingerprint, not a signature",
104
+ },
105
+ "lambda_gate": {
106
+ "hub_id": "SZLHOLDINGS/szl-lambda-gate",
107
+ "role": "advisory Λ weighted-geometric-mean gate",
108
+ "honesty": "Λ uniqueness = Conjecture 1 (open); advisory, NOT proven trust",
109
+ },
110
+ "energy_core": {
111
+ "hub_id": "SZLHOLDINGS/governed-inference-meter",
112
+ "role": "MEASURED-joule energy accounting (real NVML delta)",
113
+ "honesty": "no GPU/NVML => UNAVAILABLE_NO_NVML, joules None — never fabricated",
114
+ },
115
+ }
116
+
117
+ # Governance-layer series companions (not part of the numeric forward pass).
118
+ # Each is independently published and operates on the SAME UnifiedReceiptChain
119
+ # provenance, extending the suite from "govern provenance ACROSS ops" to
120
+ # "sign that provenance" and "refuse-and-record honestly."
121
+ _SERIES: Dict[str, Dict[str, str]] = {
122
+ "govsign": {
123
+ "hub_id": "SZLHOLDINGS/szl-govsign",
124
+ "role": "signed governance attestation (DSSE / in-toto, ECDSA P-256)",
125
+ "honesty": "signature proves authorship+integrity, NOT Λ uniqueness; proven_trust locked False; a BLOCKED verdict is signed as BLOCKED",
126
+ },
127
+ "blocked": {
128
+ "hub_id": "SZLHOLDINGS/szl-blocked",
129
+ "role": "honest-BLOCKED first-class state + EU AI Act Annex IV DRAFT derivation",
130
+ "honesty": "a BLOCKED op never executes (no fake-green); Annex IV output is a DRAFT skeleton, NOT legal advice",
131
+ },
132
+ "provctl": {
133
+ "hub_id": "SZLHOLDINGS/szl-provctl",
134
+ "role": "provenance-DAG verify + in-toto v1 / SLSA v1 interop + per-kernel MEASURED energy",
135
+ "honesty": "in-toto/SLSA field names spec-exact; proven_trust locked False; energy MEASURED-only (never fabricated); honest-BLOCKED DAG nodes surfaced, never dropped",
136
+ },
137
+ }
138
+
139
+ PROVENANCE = {
140
+ "suite": "szl_kernels",
141
+ "members": _REGISTRY,
142
+ "series_companions": _SERIES,
143
+ "lean_repo": "szl-holdings/lutar-lean",
144
+ "doi_lutar_lean": "10.5281/zenodo.20434308",
145
+ "lambda_status": "Conjecture 1 (open) — uniqueness unproven; advisory only",
146
+ "shared_chain": "UnifiedReceiptChain — op-agnostic SHA3-256 cross-kernel provenance",
147
+ }
148
+
149
+
150
+ def list_kernels() -> Dict[str, Dict[str, str]]:
151
+ """Return the suite registry: member key -> {hub_id, role, honesty}."""
152
+ return {k: dict(v) for k, v in _REGISTRY.items()}
153
+
154
+
155
+ def list_series() -> Dict[str, Dict[str, str]]:
156
+ """Return the governance-layer series companions (govsign, blocked, provctl).
157
+
158
+ These extend the numeric suite with signed attestation (govsign),
159
+ honest-BLOCKED + EU AI Act Annex IV derivation (blocked), and provenance-DAG
160
+ verification + in-toto/SLSA interop + per-kernel MEASURED energy (provctl).
161
+ They operate on the same UnifiedReceiptChain but are not part of the numeric
162
+ forward pass.
163
+ """
164
+ return {k: dict(v) for k, v in _SERIES.items()}
165
+
166
+
167
+ def get_member(key: str) -> Dict[str, str]:
168
+ """Resolve one suite member's Hub coordinates + honest role string."""
169
+ if key not in _REGISTRY:
170
+ raise KeyError(f"unknown suite member {key!r}; have {list(_REGISTRY)}")
171
+ return dict(_REGISTRY[key])
172
+
173
+
174
+ def selfcheck() -> Dict[str, Any]:
175
+ """One-shot CPU-only suite health check; never raises.
176
+
177
+ Proves the suite's CENTRAL claim end-to-end: a single shared chain spans
178
+ governed_norm + lambda_gate + energy_core in ONE forward pass, verifies as
179
+ one tamper-evident sequence, AND a deliberately mutated copy fails
180
+ verification (so the tamper-evidence is real, not decorative).
181
+
182
+ Returns {ok, version, checks:{...}, chain_head, kernels_touched,
183
+ tamper_detected, error}.
184
+ """
185
+ checks: Dict[str, bool] = {}
186
+ error = None
187
+ chain_head = GENESIS
188
+ kernels_touched = []
189
+ tamper_detected = False
190
+ try:
191
+ torch.manual_seed(0)
192
+ x = torch.randn(4, 64, dtype=torch.float32)
193
+ w = torch.randn(64, dtype=torch.float32)
194
+
195
+ chain = UnifiedReceiptChain()
196
+
197
+ # norm matches a fp32 RMSNorm reference
198
+ y = governed_rms_norm(chain, x, weight=w, eps=1e-6)
199
+ ref = (x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + 1e-6)) * w
200
+ checks["norm_correct"] = bool(torch.allclose(y, ref, rtol=1e-5, atol=1e-5))
201
+
202
+ # advisory Λ gate runs and is recorded as advisory
203
+ gate = governed_lambda_gate(
204
+ chain, torch.tensor([0.9, 0.8, 0.95]), threshold=0.5
205
+ )
206
+ checks["lambda_advisory"] = bool(gate["advisory"] is True)
207
+
208
+ # energy is MEASURED-only: no NVML here => joules None, honest label
209
+ e = governed_measure_energy(chain)
210
+ checks["energy_honest"] = bool(
211
+ e["joules"] is None and e["label"] == "UNAVAILABLE_NO_NVML"
212
+ )
213
+
214
+ # the WHOLE multi-op pass verifies as ONE chain
215
+ ok, depth, brk = chain.verify()
216
+ checks["cross_kernel_verify"] = bool(ok and depth == 3 and brk == -1)
217
+ chain_head = chain.head()
218
+ kernels_touched = chain.kernels_touched()
219
+ checks["spans_three_kernels"] = bool(
220
+ kernels_touched == ["governed_norm", "lambda_gate", "energy_core"]
221
+ )
222
+
223
+ # offline re-verification of the exported chain
224
+ ok2, _, _ = UnifiedReceiptChain.verify_json(chain.to_json())
225
+ checks["offline_reverify"] = bool(ok2)
226
+
227
+ # tamper-evidence is REAL: mutate one recorded attr and confirm break
228
+ blob = chain.to_json()
229
+ import json as _json
230
+ recs = _json.loads(blob)
231
+ recs[0]["attrs"]["eps"] = 9.99e-3 # flip a recorded value
232
+ ok3, _, brk3 = UnifiedReceiptChain.verify_json(_json.dumps(recs))
233
+ tamper_detected = bool((not ok3) and brk3 == 0)
234
+ checks["tamper_detected"] = tamper_detected
235
+
236
+ # GovernedBlock composition runs and self-verifies
237
+ blk = GovernedBlock()
238
+ res = blk.forward(x, weight=w, gov_axes=torch.tensor([0.95, 0.9, 0.92]))
239
+ checks["block_forward"] = bool(
240
+ res["chain_ok"]
241
+ and res["chain_depth"] == 4
242
+ and res["kernels_touched"]
243
+ == ["governed_norm", "lambda_gate", "energy_core", "governed_block"]
244
+ )
245
+ except Exception as exc: # never raise from a health probe
246
+ error = f"{type(exc).__name__}: {exc}"
247
+
248
+ ok = bool(checks) and all(checks.values()) and error is None
249
+ return {
250
+ "ok": ok,
251
+ "version": __version__,
252
+ "checks": checks,
253
+ "chain_head": chain_head,
254
+ "kernels_touched": kernels_touched,
255
+ "tamper_detected": tamper_detected,
256
+ "error": error,
257
+ }
build/torch-cpu/szl_kernels/_chain.py ADDED
@@ -0,0 +1,265 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: Apache-2.0
2
+ # © 2026 SZL Holdings · Stephen P. Lutar · ORCID 0009-0001-0110-4173
3
+ """Unified, op-agnostic, SHA3-256 hash-chained provenance for the SZL kernel suite.
4
+
5
+ THE FRONTIER GAP THIS CLOSES
6
+ ----------------------------
7
+ Every governed SZL kernel today owns its OWN receipt log:
8
+ * szl_governed_norm -> ReceiptChain over normalization calls
9
+ * szl_energy_core -> CheapestWattLedger over placement decisions
10
+ * szl_lambda_gate -> (no chain; advisory self-checks only)
11
+ So a real forward pass that touches norm AND an advisory Λ gate AND the energy
12
+ meter produces THREE disconnected logs. There is no single artifact a third
13
+ party can re-walk to prove "these ops happened, in this order, on these
14
+ tensors, and nothing was inserted or reordered between them". That cross-op
15
+ provenance is the genuine gap the Kernel Hub leaders leave open: they optimize
16
+ FLOPs per op, not auditable provenance ACROSS ops.
17
+
18
+ ``UnifiedReceiptChain`` is that missing artifact: one append-only,
19
+ SHA3-256 hash-chained log whose entries are op-agnostic. A norm call, an
20
+ advisory Λ-gate call, and an energy reading all hash-chain into the SAME chain,
21
+ in call order, so the whole forward pass verifies as one tamper-evident
22
+ sequence.
23
+
24
+ HONESTY (SZL doctrine v11)
25
+ --------------------------
26
+ * The digest is a real SHA3-256 over a canonical JSON body. It is an INTEGRITY
27
+ fingerprint (tamper-evidence + ordering), NOT a cryptographic signature and
28
+ NOT a proof of authorship. DSSE / sigstore signing is a separate, out-of-band
29
+ concern and is explicitly NOT claimed here.
30
+ * Tensor fingerprints round to a fixed decimal precision so they reproduce
31
+ across devices/dtypes for the same logical values. This is the same scheme
32
+ szl_governed_norm already ships — we reuse it verbatim so a unified-chain
33
+ tensor digest equals the per-kernel chain's digest for the same tensor.
34
+ * Any Λ entry is recorded as ADVISORY metadata only. A recorded "passed=True"
35
+ is a non-compensatory advisory signal, NEVER "proven trust" (Λ uniqueness =
36
+ Conjecture 1, OPEN).
37
+ * Any energy entry carries the kernel's honest label verbatim (MEASURED /
38
+ SAMPLE / UNAVAILABLE_NO_NVML / ESTIMATE / UNKNOWN). The chain NEVER upgrades
39
+ an UNAVAILABLE reading to a number — if joules is None it stays None.
40
+ * Stdlib + torch only (Kernel Hub universal-kernel requirement). Nothing is
41
+ written to disk or the network from inside the chain.
42
+ """
43
+ from __future__ import annotations
44
+
45
+ import hashlib
46
+ import json
47
+ import threading
48
+ import time
49
+ from typing import Any, Dict, List, Optional
50
+
51
+ try: # torch is optional at import time so the chain stays inspectable headless
52
+ import torch # noqa: F401
53
+ _HAS_TORCH = True
54
+ except Exception: # pragma: no cover - exercised only in torch-less envs
55
+ _HAS_TORCH = False
56
+
57
+ # Genesis link: 64 hex zeros, identical to szl_governed_norm and szl_energy_core
58
+ # so the unified chain's genesis is byte-compatible with the per-kernel chains.
59
+ GENESIS = "0" * 64
60
+
61
+ # Default rounding precision for tensor fingerprints. MUST match
62
+ # szl_governed_norm._receipt._tensor_digest so a unified-chain norm receipt's
63
+ # out_digest is bit-identical to the standalone kernel's.
64
+ _DECIMALS = 6
65
+
66
+
67
+ def tensor_digest(t: "Any", decimals: int = _DECIMALS) -> str:
68
+ """Deterministic SHA3-256 over a tensor's rounded float32 contents.
69
+
70
+ Identical scheme to szl_governed_norm._receipt._tensor_digest: round to a
71
+ fixed number of decimals, integerize, hash the raw little-endian bytes. This
72
+ makes the digest stable across devices/dtypes for the same logical values
73
+ (tiny FP noise won't change it). Integrity fingerprint, NOT a signature.
74
+
75
+ Falls back to hashing the repr only if torch is unavailable (headless
76
+ inspection); in a real kernel run torch is always present.
77
+ """
78
+ if _HAS_TORCH and hasattr(t, "detach"):
79
+ flat = t.detach().to(torch.float32).reshape(-1)
80
+ scaled = (
81
+ torch.round(flat * (10 ** decimals)).to(torch.int64).cpu().numpy().tobytes()
82
+ )
83
+ return hashlib.sha3_256(scaled).hexdigest()
84
+ return hashlib.sha3_256(repr(t).encode("utf-8")).hexdigest()
85
+
86
+
87
+ class UnifiedReceiptChain:
88
+ """Append-only, SHA3-256 hash-chained log spanning MANY kernel ops.
89
+
90
+ Each receipt body is canonical JSON with a fixed, sorted schema:
91
+ {seq, kernel, op, attrs, prev}
92
+ where:
93
+ seq -- 0-based position in the chain
94
+ kernel -- which suite member emitted it ('governed_norm', 'lambda_gate',
95
+ 'energy_core', or a composed block name)
96
+ op -- the operation ('rms_norm', 'lambda_gate', 'measure_energy', ...)
97
+ attrs -- an op-agnostic, JSON-able dict of honest, reproducible
98
+ attributes (shapes, eps, out_digest, advisory Λ fields,
99
+ energy label/joules, ...). NEVER fabricated.
100
+ prev -- digest of the previous receipt (GENESIS for seq 0)
101
+
102
+ digest = SHA3-256 over the canonical JSON of {seq,kernel,op,attrs,prev}
103
+ (ts is excluded so the digest is reproducible offline).
104
+
105
+ verify() re-walks the chain and returns (ok, depth, first_break_seq):
106
+ it recomputes every digest and checks every prev-link, so ANY insertion,
107
+ deletion, reordering, or mutation of a recorded attribute is detected.
108
+ """
109
+
110
+ def __init__(self) -> None:
111
+ self._lock = threading.RLock()
112
+ self._records: List[Dict[str, Any]] = []
113
+
114
+ # -- canonical hashing (sorted keys, tight separators, strict no-NaN) --
115
+ @staticmethod
116
+ def _digest_body(body: Dict[str, Any]) -> str:
117
+ raw = json.dumps(
118
+ body, sort_keys=True, separators=(",", ":"), allow_nan=False
119
+ ).encode("utf-8")
120
+ return hashlib.sha3_256(raw).hexdigest()
121
+
122
+ def emit(self, kernel: str, op: str, attrs: Dict[str, Any]) -> Dict[str, Any]:
123
+ """Append one op-agnostic receipt and return it (with digest + ts).
124
+
125
+ ``attrs`` must be JSON-able and finite (no NaN/Inf) — a non-finite value
126
+ would make the receipt un-attestable by a strict third-party verifier,
127
+ so it is rejected at hash time by allow_nan=False. Callers screen
128
+ non-finite numbers to None/UNKNOWN upstream (the energy kernel already
129
+ does this).
130
+ """
131
+ with self._lock:
132
+ prev = self._records[-1]["digest"] if self._records else GENESIS
133
+ seq = len(self._records)
134
+ body = {
135
+ "seq": seq,
136
+ "kernel": str(kernel),
137
+ "op": str(op),
138
+ "attrs": attrs,
139
+ "prev": prev,
140
+ }
141
+ digest = self._digest_body(body)
142
+ rec = dict(body, digest=digest, ts=time.time())
143
+ self._records.append(rec)
144
+ return rec
145
+
146
+ # -- convenience emitters that enforce the honesty schema per kernel ----
147
+ def emit_norm(self, op: str, x: "Any", out: "Any", eps: float) -> Dict[str, Any]:
148
+ """Record a governed-norm op. out_digest matches szl_governed_norm."""
149
+ return self.emit(
150
+ "governed_norm",
151
+ op,
152
+ {
153
+ "in_shape": list(getattr(x, "shape", [])),
154
+ "in_dtype": str(getattr(x, "dtype", "")).replace("torch.", ""),
155
+ "eps": float(eps),
156
+ "out_digest": tensor_digest(out),
157
+ },
158
+ )
159
+
160
+ def emit_lambda(
161
+ self,
162
+ score: float,
163
+ threshold: float,
164
+ passed: bool,
165
+ k: int,
166
+ ) -> Dict[str, Any]:
167
+ """Record an ADVISORY Λ-gate evaluation.
168
+
169
+ ``advisory`` is hard-coded True and ``lambda_status`` is stamped on every
170
+ entry so the chain itself is self-documenting: a recorded pass is NEVER
171
+ proven trust (Λ uniqueness = Conjecture 1, open).
172
+ """
173
+ return self.emit(
174
+ "lambda_gate",
175
+ "lambda_gate",
176
+ {
177
+ "score": float(score),
178
+ "threshold": float(threshold),
179
+ "passed": bool(passed),
180
+ "k": int(k),
181
+ "advisory": True,
182
+ "lambda_status": "Conjecture 1 (open) — advisory only, NOT proven trust",
183
+ },
184
+ )
185
+
186
+ def emit_energy(self, measurement: Dict[str, Any]) -> Dict[str, Any]:
187
+ """Record an energy reading VERBATIM (label + joules as-given).
188
+
189
+ ``joules`` may be None (UNAVAILABLE/UNKNOWN) — it is recorded as None,
190
+ never upgraded to a fabricated number. ``label`` is the kernel's honest
191
+ MEASURED/SAMPLE/UNAVAILABLE_NO_NVML/ESTIMATE/UNKNOWN string.
192
+ """
193
+ joules = measurement.get("joules", None)
194
+ return self.emit(
195
+ "energy_core",
196
+ "measure_energy",
197
+ {
198
+ "label": str(measurement.get("label", "UNKNOWN")),
199
+ "joules": (None if joules is None else float(joules)),
200
+ "source": str(measurement.get("source", "")),
201
+ },
202
+ )
203
+
204
+ # -- read surface --------------------------------------------------------
205
+ def head(self) -> str:
206
+ with self._lock:
207
+ return self._records[-1]["digest"] if self._records else GENESIS
208
+
209
+ def count(self) -> int:
210
+ with self._lock:
211
+ return len(self._records)
212
+
213
+ def tail(self, n: int = 10) -> List[Dict[str, Any]]:
214
+ with self._lock:
215
+ return list(self._records[-n:])
216
+
217
+ def kernels_touched(self) -> List[str]:
218
+ """Distinct kernels that appear in the chain, in first-seen order.
219
+
220
+ The cross-kernel provenance summary: proof that a single auditable chain
221
+ actually spanned multiple suite members in one run.
222
+ """
223
+ with self._lock:
224
+ seen: List[str] = []
225
+ for r in self._records:
226
+ if r["kernel"] not in seen:
227
+ seen.append(r["kernel"])
228
+ return seen
229
+
230
+ def verify(self):
231
+ """Re-walk the chain. Returns (ok: bool, depth: int, first_break: int).
232
+
233
+ Recomputes every digest from {seq,kernel,op,attrs,prev} and checks each
234
+ prev-link. Detects insertion, deletion, reorder, or any attribute
235
+ mutation. first_break is the seq of the first bad record, or -1 if clean.
236
+ """
237
+ with self._lock:
238
+ prev = GENESIS
239
+ for i, rec in enumerate(self._records):
240
+ body = {k: rec[k] for k in ("seq", "kernel", "op", "attrs", "prev")}
241
+ if rec["prev"] != prev or rec["digest"] != self._digest_body(body):
242
+ return (False, len(self._records), i)
243
+ prev = rec["digest"]
244
+ return (True, len(self._records), -1)
245
+
246
+ def to_json(self) -> str:
247
+ """Export the full chain as canonical JSON for OFFLINE re-verification.
248
+
249
+ A third party can load this, recompute each digest, and confirm the
250
+ chain independently — no trust in the emitting process required.
251
+ """
252
+ with self._lock:
253
+ return json.dumps(self._records, sort_keys=True, separators=(",", ":"))
254
+
255
+ @staticmethod
256
+ def verify_json(blob: str):
257
+ """Verify an exported chain offline. Returns (ok, depth, first_break)."""
258
+ records = json.loads(blob)
259
+ prev = GENESIS
260
+ for i, rec in enumerate(records):
261
+ body = {k: rec[k] for k in ("seq", "kernel", "op", "attrs", "prev")}
262
+ if rec["prev"] != prev or rec["digest"] != UnifiedReceiptChain._digest_body(body):
263
+ return (False, len(records), i)
264
+ prev = rec["digest"]
265
+ return (True, len(records), -1)
build/torch-cpu/szl_kernels/_ops.py ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: Apache-2.0
2
+ # © 2026 SZL Holdings · Stephen P. Lutar · ORCID 0009-0001-0110-4173
3
+ """Governed ops + a composed GovernedBlock, all writing ONE unified chain.
4
+
5
+ This module is the thin "suite" layer. It does NOT reimplement the kernels'
6
+ numerics — the canonical math lives in szl_governed_norm / szl_lambda_gate /
7
+ szl_energy_core. Here we provide:
8
+
9
+ * self-contained REFERENCE numerics (rms_norm, layer_norm, Λ aggregate) that
10
+ are byte-faithful ports of the published kernels, so this reference package
11
+ RUNS standalone with only torch installed (no Hub fetch needed for the
12
+ self-test). In production the suite delegates to the installed kernels via
13
+ get_kernel(); see NEXT_ARTIFACT_SPEC.md "Delegation".
14
+ * governed wrappers that emit op-agnostic receipts into a shared
15
+ UnifiedReceiptChain (cross-kernel provenance).
16
+ * GovernedBlock: a pre-norm transformer sub-block that composes governed-norm
17
+ + an ADVISORY Λ gate + an energy measurement into ONE auditable forward
18
+ pass with an end-to-end receipt.
19
+
20
+ HONESTY: Λ stays advisory (Conjecture 1, open). Energy stays MEASURED-only
21
+ (None when no NVML — never fabricated). No benchmarks are claimed.
22
+ """
23
+ from __future__ import annotations
24
+
25
+ import time
26
+ from typing import Any, Dict, Optional, Tuple
27
+
28
+ import torch
29
+
30
+ from ._chain import UnifiedReceiptChain
31
+
32
+ # ---------------------------------------------------------------------------
33
+ # Reference numerics — faithful ports of the published kernel math.
34
+ # (rms_norm / layer_norm mirror szl_governed_norm._norm; lambda_aggregate
35
+ # mirrors szl_lambda_gate._lambda. Kept minimal but numerically identical.)
36
+ # ---------------------------------------------------------------------------
37
+ def _rms_norm(x: torch.Tensor, weight: Optional[torch.Tensor], eps: float) -> torch.Tensor:
38
+ xf = x.to(torch.float32)
39
+ out = xf * torch.rsqrt(xf.pow(2).mean(-1, keepdim=True) + eps)
40
+ if weight is not None:
41
+ out = out * weight.to(torch.float32)
42
+ return out.to(x.dtype)
43
+
44
+
45
+ def _layer_norm(
46
+ x: torch.Tensor,
47
+ weight: Optional[torch.Tensor],
48
+ bias: Optional[torch.Tensor],
49
+ eps: float,
50
+ ) -> torch.Tensor:
51
+ return torch.nn.functional.layer_norm(
52
+ x, (x.shape[-1],), weight=weight, bias=bias, eps=eps
53
+ )
54
+
55
+
56
+ def _lambda_aggregate(axes: torch.Tensor, weights: Optional[torch.Tensor]) -> torch.Tensor:
57
+ """Weighted geometric mean Λ(x)=∏ xᵢ^{wᵢ} with non-compensatory zero-routing.
58
+
59
+ Faithful to szl_lambda_gate: any zero or non-finite axis drives Λ to 0.
60
+ """
61
+ cdt = torch.float32 if axes.dtype in (torch.float16, torch.bfloat16) else axes.dtype
62
+ xf = axes.to(cdt)
63
+ k = xf.shape[-1]
64
+ if weights is None:
65
+ w = torch.full((k,), 1.0 / k, dtype=cdt, device=xf.device)
66
+ else:
67
+ w = weights.to(cdt)
68
+ w = w / w.sum()
69
+ finite = torch.isfinite(xf)
70
+ xc = xf.clamp(0.0, 1.0)
71
+ bad = (~finite) | (xc <= 0.0)
72
+ any_bad = torch.any(bad, dim=-1)
73
+ safe = torch.where(bad, torch.ones_like(xc), xc)
74
+ val = torch.exp((torch.log(safe) * w).sum(dim=-1))
75
+ out = torch.where(any_bad, torch.zeros_like(val), val).clamp(0.0, 1.0)
76
+ return out.to(axes.dtype)
77
+
78
+
79
+ # ---------------------------------------------------------------------------
80
+ # Governed wrappers — numerics + a receipt into the SHARED unified chain.
81
+ # ---------------------------------------------------------------------------
82
+ def governed_rms_norm(
83
+ chain: UnifiedReceiptChain,
84
+ x: torch.Tensor,
85
+ weight: Optional[torch.Tensor] = None,
86
+ eps: float = 1e-6,
87
+ ) -> torch.Tensor:
88
+ out = _rms_norm(x, weight, eps)
89
+ chain.emit_norm("rms_norm", x, out, eps)
90
+ return out
91
+
92
+
93
+ def governed_layer_norm(
94
+ chain: UnifiedReceiptChain,
95
+ x: torch.Tensor,
96
+ weight: Optional[torch.Tensor] = None,
97
+ bias: Optional[torch.Tensor] = None,
98
+ eps: float = 1e-5,
99
+ ) -> torch.Tensor:
100
+ out = _layer_norm(x, weight, bias, eps)
101
+ chain.emit_norm("layer_norm", x, out, eps)
102
+ return out
103
+
104
+
105
+ def governed_lambda_gate(
106
+ chain: UnifiedReceiptChain,
107
+ axes: torch.Tensor,
108
+ weights: Optional[torch.Tensor] = None,
109
+ threshold: float = 0.5,
110
+ ) -> Dict[str, Any]:
111
+ """ADVISORY Λ gate that records an advisory receipt. NOT proven trust.
112
+
113
+ Returns a dict {score, passed, threshold, advisory=True}. ``passed`` is an
114
+ advisory, non-compensatory signal only (Λ uniqueness = Conjecture 1, open).
115
+ """
116
+ score = _lambda_aggregate(axes, weights)
117
+ s = float(score.reshape(-1)[0]) if score.dim() else float(score)
118
+ passed = s >= float(threshold)
119
+ chain.emit_lambda(s, float(threshold), passed, int(axes.shape[-1]))
120
+ return {"score": s, "passed": passed, "threshold": float(threshold), "advisory": True}
121
+
122
+
123
+ def governed_measure_energy(
124
+ chain: UnifiedReceiptChain,
125
+ measurement: Optional[Dict[str, Any]] = None,
126
+ ) -> Dict[str, Any]:
127
+ """Record an energy reading. MEASURED-only honesty.
128
+
129
+ If ``measurement`` is None we DO NOT fabricate joules: we record the honest
130
+ UNAVAILABLE_NO_NVML label with joules=None. (In production this delegates to
131
+ szl_energy_core.measure_energy(), which reads a real NVML cumulative-energy
132
+ delta when a GPU is present.)
133
+ """
134
+ if measurement is None:
135
+ measurement = {
136
+ "joules": None,
137
+ "label": "UNAVAILABLE_NO_NVML",
138
+ "source": "no GPU/NVML in this reference run — joules NOT fabricated",
139
+ }
140
+ chain.emit_energy(measurement)
141
+ return measurement
142
+
143
+
144
+ # ---------------------------------------------------------------------------
145
+ # GovernedBlock — the flagship composition (candidate (c)) on the meta-package.
146
+ # ---------------------------------------------------------------------------
147
+ class GovernedBlock:
148
+ """A pre-norm transformer sub-block with end-to-end cross-kernel provenance.
149
+
150
+ Forward pass, all into ONE shared chain, in order:
151
+ 1. governed RMSNorm of the hidden state (governed_norm)
152
+ 2. an ADVISORY Λ gate over caller-supplied governance axis scores
153
+ (lambda_gate) — recorded as advisory, never gating the math
154
+ 3. an energy measurement of the step (energy_core)
155
+ 4. a final 'block_forward' receipt binding the block's input/output digests
156
+
157
+ The result is a single hash-chained sequence a third party can re-walk to
158
+ confirm: norm ran, an advisory gate was evaluated, energy was accounted
159
+ (honestly labeled), and the block output is the one that was produced — with
160
+ nothing inserted or reordered between steps.
161
+
162
+ HONESTY: the Λ gate is ADVISORY; ``forward`` does NOT block, mask, or alter
163
+ the numerics based on the gate. The gate result is recorded for audit only.
164
+ Energy is MEASURED-only. No speed claims.
165
+ """
166
+
167
+ def __init__(self, chain: Optional[UnifiedReceiptChain] = None) -> None:
168
+ self.chain = chain if chain is not None else UnifiedReceiptChain()
169
+
170
+ def forward(
171
+ self,
172
+ hidden: torch.Tensor,
173
+ weight: Optional[torch.Tensor] = None,
174
+ eps: float = 1e-6,
175
+ gov_axes: Optional[torch.Tensor] = None,
176
+ gov_weights: Optional[torch.Tensor] = None,
177
+ threshold: float = 0.5,
178
+ energy_measurement: Optional[Dict[str, Any]] = None,
179
+ ) -> Dict[str, Any]:
180
+ c = self.chain
181
+ in_digest = __import__("hashlib").sha3_256(
182
+ __import__("json").dumps(list(hidden.shape)).encode()
183
+ ).hexdigest()
184
+
185
+ # 1. governed normalization (numerics + receipt)
186
+ normed = governed_rms_norm(c, hidden, weight=weight, eps=eps)
187
+
188
+ # 2. ADVISORY Λ gate (recorded, does NOT alter numerics)
189
+ if gov_axes is None:
190
+ # default: neutral advisory axes; still recorded honestly as advisory
191
+ gov_axes = torch.tensor([0.9, 0.9, 0.9], dtype=torch.float32)
192
+ gate = governed_lambda_gate(c, gov_axes, weights=gov_weights, threshold=threshold)
193
+
194
+ # 3. energy measurement (MEASURED-only; None when no NVML)
195
+ energy = governed_measure_energy(c, energy_measurement)
196
+
197
+ # 4. bind the block I/O into a final receipt
198
+ from ._chain import tensor_digest
199
+ c.emit(
200
+ "governed_block",
201
+ "block_forward",
202
+ {
203
+ "in_shape_digest": in_digest,
204
+ "out_digest": tensor_digest(normed),
205
+ "gate_passed_advisory": bool(gate["passed"]),
206
+ "energy_label": str(energy.get("label", "UNKNOWN")),
207
+ },
208
+ )
209
+
210
+ ok, depth, brk = c.verify()
211
+ return {
212
+ "output": normed,
213
+ "gate": gate, # advisory only
214
+ "energy": energy, # MEASURED-only label/joules
215
+ "chain_head": c.head(),
216
+ "chain_depth": depth,
217
+ "chain_ok": ok,
218
+ "chain_first_break": brk,
219
+ "kernels_touched": c.kernels_touched(),
220
+ }