OPEN METHODOLOGY · APACHE-2.0 · CC-BY-4.0 · US-ONLY ZERO TRUST

US Citizen AI Commons

Bring a LoRA adapter, a distilled prompt-response corpus, or an evaluation dataset. If it passes review, it goes onto the public append-only ledger and into the next federated re-training run. Long-horizon hypothesis: a peer-trained U.S.-citizen open-weight commons can collectively reach about 10% reflection capability — a sovereign AI commons that is not gate-kept by any single foreign-controlled foundation lab.

This is the open-source half of the Three-Pillar deep-tech house: pillar 3 (SilverDuck) shipped its training pipeline under Apache-2.0 + CC-BY-4.0 so anyone can reproduce, extend, and contribute back. The Commons is the public on-ramp.

Submit a contribution View ledger Schema (JSON)

How it works

1
Prepare your artifact under a permissive open license.
Apache-2.0, CC-BY-4.0, or MIT only. Publish to Hugging Face, GitHub releases, or IPFS — wherever it has a stable HTTPS URL.
2
Submit through the form below or the JSON API.
Include the artifact URL, SHA-256 checksum, base model (US-only Tier 1: Llama 3.1 or Phi-3.5), and a U.S.-person attestation. All fields are checked against commons-schema.json.
3
Human review within 5 business days.
The curator (Artem Shakin) downloads the artifact, verifies the checksum, sanity-loads it, and checks for license / origin / hygiene compliance. Decision is logged either way.
4
Accepted contributions enter the federated training run.
LoRA adapters and corpora are merged into the next nightly fine-tune. The merged release is published under the same license family (re-share-alike) on the Nightbox Hugging Face org with full attribution to your handle in the ledger.

What we accept

Kind	Description	Typical size
lora_adapter	LoRA / QLoRA adapter for a Tier 1 US-origin base model.	10 MB – 2 GB
distilled_corpus	JSONL of prompt-response pairs (with provenance) for SFT or DPO.	1 MB – 200 MB
eval_dataset	Held-out evaluation set with reference answers.	100 KB – 50 MB
prompt_pack	Curated prompt library (system prompts, few-shot exemplars, jailbreak hardening).	10 KB – 5 MB
tokenizer_extension	Domain-specific tokenizer additions (BPE merges, special tokens).	10 KB – 2 MB

Posture & legal

License whitelist
Apache-2.0, CC-BY-4.0, or MIT only.
U.S.-person attestation
Submitter must self-attest U.S. citizen or U.S. permanent resident, per the Absolute Zero Trust posture (zero-trust.json).
Tier 1 base models only
Llama 3.1 (Meta, US) or Phi-3.5 (Microsoft, US). For corpus-only or eval-only contributions, base model is N/A.
Append-only ledger
Every accepted contribution lives forever in data/commons-ledger.jsonl, attributed to your handle, auditable in git history.
Re-share-alike
Merged training releases ship under the same license family (or stricter share-alike) so the commons stays open downstream.
No PII, no secrets
Corpora must not contain personal identifiers or copyright-leaking source material. Reviewer rejects on suspicion.

Submit a contribution

Fill out the form below. The form posts to /api/commons-submit. You will get a submission ID back; the curator will review within 5 business days.

Kind License

Base model (Tier 1 US-origin only) Description (30–2000 chars)

Submitter handle Email (optional, for review correspondence)

Artifact URL (HTTPS — Hugging Face, GitHub release, IPFS)

SHA-256 (64 hex chars) Size (bytes)

Tags (comma-separated, optional) Training recipe URL (optional) U.S.-person attestation — I attest that I am a U.S. citizen or U.S. permanent resident. Required by the Absolute Zero Trust posture. False attestation may result in rejection of the contribution and removal of any prior accepted entries from this handle.

Public ledger

Accepted contributions, newest first. Source: /api/commons-ledger · raw: /data/commons-ledger.jsonl

Loading ledger…

Maintainer: @ArtemShkin · curator email artem@nightboxllc.com · review SLA 5 business days · all decisions logged. Posture authority: zero-trust.json#ai_brain_origin_posture.