A content contract layer for AI long-form output
NinFenz (NF) turns "AI produces long-form content reliably" into engineering you can load, verify and reproduce. The protocol is domain-neutral and model-agnostic; narrative is only the first official domain pack.
Not a model · not a prompt-template collection · not a vendor SDK · MIT · runs offline
What it is / what it is not
It is: a protocol (module / pipeline / asset) plus one deterministic gate entry point plus a recomputable asset ledger - all plain markdown and JSON, stored in git.
It is not: a model, a prompt collection, or an SDK. The gates check structure and reproducibility, not prose quality - that boundary is part of the design.
Citable facts (every number has a source)
- Quality evidence: verify v2.29 · check1-39 · PASS=68 via a single entry point,
bash verify.sh. - Scale: 111 registered packs · 363 asset files · 101 concept graphs; standards catalogue 370 entries (332 reachable / 38 not) · 1200 bindings. source protocol/repo_stats.json
- Official core: 13 modules · 3 pipelines (P00 / P01 / P90) · protocol files 01-07.
- Verifiable artifacts: Merkle inclusion proofs (RFC 6962) plus optional hmac / ssh-sig / sigstore-keyless anchors; missing verifier = fail closed.
- Interop: read-only MCP surface (
nf serve), dual-era 2026-07-28 / 2025-11-25. - Terminal:
python tui/nf.py- a full-screen TUI in the pure standard library.
Run it in five minutes
# read-only checks, no clone needed
npx -y ninfenz doctor
# full-screen terminal TUI
npx -y ninfenz tui
# all 39 gates (needs a real checkout)
git clone https://github.com/Monyeah777/NinFenz
cd NinFenz && bash verify.shRequires Node >= 18.17 and Python >= 3.11 (standard library only, no pip dependencies).
See it running
Terminal TUI (standard library only, offline; 14-second demo):
How it differs from a prompt collection
| Prompt collection | NinFenz | |
|---|---|---|
| Deliverable | a block of text | protocol + gates + asset ledger |
| Second run | random structure | isomorphic to the first (diffable against a baseline) |
| Pass/fail | a human eyeballs it | 39 deterministic checks against a frozen baseline |
| Verifiable | no | Merkle inclusion proof + signature anchors, fail closed |
| Boundary | - | structure and reproducibility, not prose quality |
Why the name
From the Liexian Zhuan: Ning Fengzi (宁封子) was the Yellow Emperor's Keeper of the Kiln - raw clay goes in; after kiln temperature and a seal, a durable artifact comes out. Kiln temperature maps to the deterministic gates; the seal maps to verifiable proof. NinFenz is the compressed form; NF is the short name.
FAQ
What is NinFenz?
A content contract layer: a protocol plus deterministic quality gates plus asset standards, so long-form AI output can be loaded, checked and reproduced.
Is it a model?
No. The protocol is model-agnostic; any model that can read text can assemble and load it.
How is it different from prompt templates?
Templates give text. NinFenz gives a protocol, verifiable gates and an asset ledger, so an output can be judged pass or fail.
Does it need network access?
No. The repo is markdown source; bash verify.sh reads locally and runs offline.
Can the output be verified?
Yes. Artifacts carry Merkle inclusion proofs (RFC 6962) and optional hmac / ssh-sig / sigstore anchors; a missing verifier fails closed instead of passing silently.
Licence and cost?
MIT, no paid tier, no hosted service.