SZL Constellation - the living archive. 49 estates unarchived as a WebGL hologram, 8 verticals wired to their real repos/kernels/models/datasets, FastAPI JSON backend with recomputing receipts. Source mirror of HF Space SZLHOLDINGS/szl-constellation. Doctrine v11 - nothing glows that didn't earn it. Apache-2.0.
Founder and CEO, SZL Holdings - governed AI infrastructure with inspectable source, runtime state, receipts, and proof boundaries.
ai-governancecounter-uascryptographic-proofdecision-intelligencedefensefounder
A request arrives with its evidence; policy can stop it; the decision keeps its reasons; and a governed write can carry a hash-chained receipt that anyone can check offline. Deny by default. Runtime readback (2026-10-19): service health ok, operational readiness NOT_READY (1 of 8 frontier endpoints live). Of 138 registered surfaces, the runtime labels 4 LIVE, 9 MEASURED and 110 MODELED. Source: github.com/szl-holdings/a11oy (Apache-2.0).
governed-aiai-governancedecision-receiptsfail-closedprovenance
The evidence origin for a11oy, kept separate from the product: a RECORD index, a Hub atlas, a diligence room and a ROADMAP. It is served as static pages built from github.com/szl-holdings/a11oy-net, and the served source witness matches main exactly (e7479e87). It is not a runtime and not a receipt store, and the page says so.
evidenceprovenancetransparencygoverned-ai
129 public repositories: 100 active, plus 29 archived and kept read-only for provenance. They include receipt formats and offline verifiers (governed-receipt-spec, szl-receipt, szl-evidence-core), Lean 4 proofs (lutar-lean), benchmark harnesses (frontier-bench, retrieval-bench, quant-curve), Triton attention and KV-cache kernels, and the a11oy and killinchu products. Mostly Apache-2.0.
open-sourceai-governancereceiptsbenchmarkskernels
48 models, 37 datasets, 38 Spaces and 14 kernels, all public. Each model is labeled by evidence class instead of being presented as a finished model. 18 hold trained weights: small fine-tunes and LoRA adapters on Qwen2.5 and Qwen3.5 bases from 0.5B to 1.5B. Of the rest, 16 are kernel software, 7 are code or scripts, 5 are nano synthetic, and 2 are empty roadmap cards (a11oy.net record, 2026-10-13). None is claimed to be operational or fully benchmarked.
modelsdatasetslorakernelsprovenance
@stephenlutar2-hash3
