YanAn

Three locks, all of them working, all of them with a blind spot

Checks today sit in three places: screen the question, match the text against a word list, review the finished answer. Each is effective, and each has a blind spot you can name in advance.

The check keeps pace with generation, and stops mid-sentence

YanAn produces a risk judgment at every step of generation. When the score crosses the line you set, output stops immediately, the remainder stays on the server, and the user gets a clear message. Every step carries a chance to check, and content reaches the screen only after it passes.

After a cutoff, the trigger, the score at that moment and the associated cues are written to the record together, so the safety team can review what happened. Alerting, pausing and handoff to a person are configured to your requirements.

RedShield scores, TingLan reads the cues

Both engines read the same running signals in parallel. RedShield returns a comparable risk score; TingLan returns readable word-level cues. Either one crossing the line you set triggers a cutoff.

RedShield 红盾 · the scoring engine

A comparable risk score at every generation step, computed from signals the main model has already produced. The model is not asked to write a review of its own answer, and no second generative reviewer of comparable size has to be stood up.

When the score crosses the threshold, output stops, the unreleased remainder is not sent, and the trigger is recorded. Ordinary answers continue at full speed.

TingLan 听澜 · the cue engine

An answer can be perfectly well phrased and still leave cues worth a look inside. In a government service exchange, the printed answer may never say "address", "phone number" or "ID number", while the internal cues still flag whether the reply touches another person's private data.

Word-level cues have to be read in context; surfacing a cue is not proof of intent. When the cues cross the line you set, TingLan triggers the same cutoff — the risky content is stopped before it is ever spoken. The cues, with their context, then go to a human reviewer who decides whether the answer actually crossed a line.

Each step adds one check of fixed cost

Have a second model re-read the whole prefix after every token and the duplicated work piles up as the answer grows. YanAn reuses the signals the main model has already computed and performs one incremental check per step — so as the answer gets longer, the added cost of each check stays constant.

Adaptation training costs roughly two orders of magnitude less, so when an industry's requirements change, rules and configuration are updated as needed rather than rebuilt. Acceptance rests on three numbers: added cost per thousand requests, added waiting time on the answer, and the request volume your existing resources can carry.

Four places where YanAn is worth evaluating

Two delivery forms, one core

Both forms ship the same YanAn core. The internal dual engine requires the integration conditions to be met; for a third-party closed-source API, integration normally takes the input/output checking form.

Four steps to production, any of them abortable

Start on a narrow slice, then open up. Every step has a defined acceptance artifact, every step can be stopped, and every step is judged on the same three numbers: how much risk was missed, how many ordinary questions were intercepted, and how much extra time and resource it cost.

Deployment boundaries: where it sits, what it reads, who decides

Control over the model, the data and the rules stays with you. Deployment, logging scope and interception behaviour are written into the delivery plan.

Nine things your technical team should confirm before deployment

They fall on three lines: integration, running cost, and how well the checking works; the implementation itself is packaged inside the product. The acceptance principle is one comparison — run the same batch of live questions through your current approach and through YanAn, confirm the checking result first, then compare speed, resource use and cost.