# Hongchen Lingjing — deterministic verification for AI-computed numbers
We build three tools around one idea: AI-computed numbers must be checkable, deterministically, before anyone acts on them. No confidence scores, no vibes — three verdicts only: verified, refuted, unverified (honest abstention).
## Lingshu Solver — deterministic equation engine (Krawczyk-certified real roots, ≤6 vars)
- Hosted MCP endpoint: `https://hongchenlingjing.com/mcp` (streamable HTTP)
- Web app:
- Free forever via web / `npx lingshu-solver`; the hosted MCP endpoint is free as well, no metering — see [support](https://hongchenlingjing.com/pay/en.html) (corporate account, any amount, nothing to do after paying)
## Lingjing — personal AI brain / knowledge library (bring your own key)
-
## LingNao — AI number verification API (verify the numbers your AI produces)
- Playground:
- API: `POST https://hongchenlingjing.com/verify` — body `{"caseLabel":"...","items":[{"id":"i1","kind":"algebraic","payload":{"equations":["x+2=5"],"claimed":{"x":3}}}]}`; three verdicts only: verified | refuted | unverified; English in, English out.
- Claim kinds (`items[].kind`): `algebraic` (equations + claimed), `membership` (claimed value ∈ certified solution set), `entail` (∀ certified solutions satisfy a boundary), `consistency` (≥2 statements mutually non-exclusive), `constraint`. `items[].depends:[id]` builds a DAG; the kernel topologically sorts it, refuses on cycles (422), and composes by `∃ refuted ⇒ refuted; else ∀ verified ⇒ verified; else 𝕌`. Optional `items[].tol` (relative, capped at 1e-3) and request-level `budget` (default 200 primitive evaluations; unaffordable items abstain as `abstainedBy:"budget"`).
- **Variable names.** Multi-character names work when declared in `variables` `["subtotal","tax","total"]`: the solver keeps them whole and echoes the resolved list in `varNames`. With `variables` empty the default is single-letter product parsing (`xy` → `x*y`); if that would exceed the 6-variable cap the engine switches to whole-identifier reading and flags `interpretation:"whole_identifier_fallback"`. Read `varNames` back before trusting `values`.
- Integration guide (JS / Python / LangChain / OpenAI tool schema / MCP snippets):
- MCP (stdio): `npx lingnao-mcp` — one gateway tool `lingnao`
## Scope and refusals (by design)
Medical, mental-health, education & training, investment or financial advice, and legal advice are refused (HTTP 422, `compliance:true`). A verification service must refuse what it cannot check.
## Company
广州市红尘灵境数字科技有限公司 (Guangzhou Hongchen Lingjing Digital Technology Co., Ltd.) · 粤ICP备2026031206