Explore AI systems, models and research tools. Follow an output to its source, understand its limits, and build on work you can inspect.
Models · Kernels · Datasets · Spaces
Artifact type, evidence class, and runtime state remain separate.
The single-cell example summarizes public GEO GSE85241 data and checks integrity and repeatability. Its receipt is unsigned; it makes no biological or clinical claim. The retained record reports repeat output and two rejected changes; it does not authenticate authorship.
Explore the portfolio, use a demonstration, inspect source code or reproduce a result. Each path shows the evidence it can support.
Inspect policy, authority, decision state, and the receipt emitted by a bounded workflow.
Open a11oy Space -> 02 / BuildPin models, datasets, and governed software by immutable revision and disclosed lineage.
Browse models -> 03 / ResearchTrace primary artifacts, derived forms, evaluations, datasets, and demonstrations as one bounded collection.
Open collections -> 04 / VerifyFollow receipts and provenance outside the product UI without treating reachability as proof.
Open proof registry ->Every flagship publishes the same chain: GitHub source, Hugging Face runtime, product surface, proof and known bounds. Runtime truth is read from the running system, never implied from this card.
Governed-AI command center: deny-by-default policy, a signed receipt for every governed action, and a human binding every consequential step.
Signing status and persistent storage are reported live by the runtime; the README lists the current PARTIAL states.
Counter-UAS and maritime intelligence with governed decisions and DSSE Khipu receipt evidence. Sensing feeds may be live; public effectors remain SIMULATED with a human on the loop.
Self-reporting is not independent qualification. The OSINT corpus is never trained on.
The governed training and release line behind Khipu and ReceiptAgent: receipted fine-tunes on disclosed base models, held-out gates before publication, GGUF forms for CPU serving.
Every weight card states base model, lineage, and what its evaluation does not prove.
One OpenAI-compatible endpoint in front of sovereign, free-grid, and paid fallback tiers. Every completed response carries an inspectable receipt of model, tier, attempts, and routing rationale.
Runtime truth reads LIVE, CONFIGURED_UNVERIFIED, OFFLINE_UNTIL_KEYED, or UNAVAILABLE. Unavailable providers are surfaced, never faked.
Hugging Face repository type does not redefine the artifact. Every card must say what it is, where it came from, how to reproduce it, and what it does not prove.
Trained weights, adapters, recipes, surrogates, and source-bound kernels stay explicitly classified.
View every model ->Inspect the native package API, backend requirements and immutable loading contract before executing kernel code.
View every native kernel ->Corpora, admitted evidence, evaluations, receipts, and proofs expose source and privacy boundaries.
View every dataset ->Command, verification, research, and visualization interfaces expose source and demonstration limits.
View every Space ->Native packages and model-namespace card mirrors are separate repositories, even when their names match. Load only a verified native package revision; a mirror revision does not identify executable package bytes. Their cards must not imply neural training, benchmark coverage, or production authorization that their receipts do not establish.
Six stages, one contract. The first three stages control; the last three prove. A receipt that enters a governed action is the same receipt that leaves it.
receipts.in ≡ receipts.outEvidence class says how a claim was established. Runtime state says what is available now. Neither grants authorization. HTTP 200 is reachability, not a production certificate.
Every consequential statement carries one of these explicit scopes. Demonstrations may be SIMULATED. SZLHOLDINGS/SZLHOLDINGS is a HISTORICAL profile mirror only; do not infer current inventory, runtime status, proof counts, or canonical links from it.
A signature establishes scoped integrity and origin, not accuracy, safety, or fitness.
Inspect source-bound status before relying on a running interface.
The map below is a dated, source-bound snapshot of the public Hub inventory. It is registry evidence only — not availability, operational readiness, or publication policy.
Source-bound snapshot: 37 public Spaces, 48 model repositories, 14 native kernels, 37 datasets. Observed under hf-public-author-membership/v2. Native kernel IDs may also appear in the model namespace; these figures do not count unique projects or trained language models. Private assets, collections and buckets are outside this scope. This dated observation is not a live count.
Inventory binding · Exact source inventory. Membership alone does not establish runtime readiness, model quality, or publication permission. The governed KEEP list remains docs/CANONICAL_FLEET.md.
HISTORICAL estate-alignment v1: 16 portfolio Spaces (membership observed 2026-09-05T21:41:24Z) · 48 models (observed 2026-10-11T14:19:39Z) · 37 datasets (observed 2026-10-11T14:19:39Z); 1 inventory-only Space: Yarqa · governedKeep=false · disposition=FOLD. These figures describe separately dated contract observations, not current Hub membership or visibility.
A cached card, screenshot, badge, or copied count is not current state.
Use exact revisions, keep runtime and evidence state separate, and report security issues through the protected channel.
