System One

A System One model reads a state and answers typed questions about it in one forward pass. It returns probabilities, never text. Laya is the System One model we serve, designed for triage, routing and guardrails.

System One models have their own endpoint, POST /v1/systemone, on the same base URL and with the same API key as chat. They do not answer chat completions, Responses or Messages requests, so a harness cannot run Laya as its model. You call Laya next to the model the harness runs: from a hook before a tool call, or from your own agent code before it picks a model.

Send a request#

curl https://api.inference.boundless.network/v1/systemone \
  -H "Authorization: Bearer $BOUNDLESS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "convaiinnovations/laya",
    "state": "I was charged twice for my October invoice. Please refund the duplicate and cancel my plan.",
    "questions": {
      "refund": { "type": "noul", "instructions": "Does the customer ask for a refund?" },
      "topic": {
        "type": "choice",
        "instructions": "What is the message about?",
        "criteria": {
          "billing": "Charges, invoices or refunds",
          "technical": "Errors, outages or bugs",
          "account": "Sign-in, profile or team settings"
        }
      },
      "urgency": { "type": "score", "instructions": "How urgent is this message?", "criteria": ["low", "medium", "high"] }
    }
  }'

Each question comes back under its own name in answers:

{
  "model": "convaiinnovations/laya",
  "answers": {
    "refund": { "type": "noul", "noul": 0.8932 },
    "topic": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.7971,
      "probabilities": { "billing": 0.9528, "technical": 0.0214, "account": 0.0259 }
    },
    "urgency": {
      "type": "score",
      "score": 1.1619,
      "confidence": 0.0644,
      "legend": { "0": "low", "1": "medium", "2": "high" },
      "probabilities": { "0": 0.1796, "1": 0.479, "2": 0.3414 }
    }
  },
  "usage": { "input_tokens": 161, "output_tokens": 0 }
}

model is required. Use convaiinnovations/laya or its alias laya.

Questions#

A score is the probability-weighted mean of the level indexes, so 1.16 on a three-level scale sits just above medium.

state can be a string, an object or an array. Send structured data, such as a tool call or a ticket, as JSON rather than flattening it into a sentence. Ask as many questions about one state as you need in a single request.

Decide on probabilities with a threshold you have tested, not on choice alone. choice is the likeliest option even when no option is likely, and a low confidence means the options are close. In one test, a choice between allow, ask and deny picked allow for git push --force at 0.46, while a noul asking whether the command could destroy data returned 0.90. Describing each option in criteria separates them better than a bare name.

Context budget#

Laya sends each request to its English or its multilingual checkpoint. The default budget is 512 tokens on English and 1,024 on multilingual. Set max_len in the request, up to 8,192, for longer input.

The budget covers the state, the question, its options and special tokens, for each question. Each criteria entry has its own limit of 48 tokens. Input over either limit is rejected with a 422 and never truncated.

usage.input_tokens counts the encoded input once per question. output_tokens is always 0. Laya is currently free to call. Its model page shows the live rate.

Errors#

Errors are JSON with message, error_type and code.

Every response carries an X-Request-Id. Quote it to support. Server-Timing reports the request's validation, queue and inference time in milliseconds.

Gate tool calls in Claude Code#

A Claude Code PreToolUse hook runs before each tool call. This one sends the call to Laya and, when Laya rates it likely to delete data, discard work or rewrite shared history, asks you to approve it. It needs curl, jq, and BOUNDLESS_API_KEY in Claude Code's environment, which one-command setup exports from your shell profile.

Save it as .claude/hooks/laya-gate.sh in your project and make it executable:

#!/usr/bin/env bash
input=$(cat)
answer=$(jq -c '{
  model: "convaiinnovations/laya",
  state: {tool_name, tool_input},
  questions: {destructive: {type: "noul", instructions: "Could this tool call delete data, discard work or rewrite shared history?"}}
}' <<<"$input" | curl -sf --max-time 5 https://api.inference.boundless.network/v1/systemone \
  -H "Authorization: Bearer $BOUNDLESS_API_KEY" -H "Content-Type: application/json" -d @-) || exit 0
jq -c 'select(.answers.destructive.noul >= 0.5) | {hookSpecificOutput: {
  hookEventName: "PreToolUse",
  permissionDecision: "ask",
  permissionDecisionReason: "Laya rates this \(.answers.destructive.noul * 100 | round)% likely to be destructive"
}}' <<<"$answer"

Register it in .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [{ "type": "command", "command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/laya-gate.sh", "timeout": 10 }]
      }
    ]
  }
}

Below the threshold the hook prints nothing, and Claude Code's normal permission rules decide. If Laya does not answer within five seconds, the hook also stays silent, so a failed call never approves anything. In claude -p, where nobody can answer the prompt, an ask blocks the call and the agent sees the reason.

In our tests, git status, npm test and cat package.json scored below 0.12, and rm -rf build/, git reset --hard, git push --force and DROP TABLE scored above 0.81. Check the threshold against the commands your agents run.

Any harness that can run a command or plugin before a tool call can make the same request. Send the tool name and its arguments as state.

Route requests in your own agent#

Laya can choose a model before your agent spends tokens. This sends lookups and rewrites to a fast model and code or reasoning to a stronger one, and falls back to the stronger model if Laya returns an error:

import OpenAI from "openai";

const apiKey = process.env.BOUNDLESS_API_KEY;
const client = new OpenAI({ apiKey, baseURL: "https://api.inference.boundless.network/v1" });

async function pickModel(prompt: string) {
  const response = await fetch("https://api.inference.boundless.network/v1/systemone", {
    body: JSON.stringify({
      model: "convaiinnovations/laya",
      questions: {
        tier: {
          criteria: {
            fast: "A lookup, translation, summary or reformatting",
            strong: "Code, debugging, proofs or multi-step reasoning",
          },
          instructions: "Which kind of model should answer this request?",
          type: "choice",
        },
      },
      state: prompt,
    }),
    headers: { Authorization: `Bearer ${apiKey}`, "Content-Type": "application/json" },
    method: "POST",
  });
  if (!response.ok) return "glm-5.3";
  const { answers } = await response.json();
  return answers.tier.probabilities.strong >= 0.5 ? "glm-5.3" : "glm-5.3-flash";
}

const prompt = "Translate thank you into Spanish";
const completion = await client.chat.completions.create({
  messages: [{ content: prompt, role: "user" }],
  model: await pickModel(prompt),
});

In our tests, strong scored 0.17 to 0.25 on a capital-city lookup, a translation and an email summary, and 0.80 to 0.85 on a refactor and a race-condition hunt. A proof that the square root of 2 is irrational scored 0.55, close to the line. Pick model IDs from Models.