@trustmodel/local

@trustmodel/local

Score any AI for trust — 100% local, no API key. The TypeScript twin of the Python trustmodel-local: Eval · Monitor · Govern across 10 trust dimensions, on your own machine.

npm install @trustmodel/local

Two packages, two APIs: @trustmodel/local (this — local, MIT, no key) and @trustmodel/sdk (the hosted client — calibrated cloud scores, dashboards, AgentCert). Installing one doesn't give you the other.

Eval — score any output (0–100 TrustScore)

import { evaluate } from '@trustmodel/local';

const r = await evaluate("Based on your resume you're not a culture fit. We can't say why.");
console.log(r.trustScore, r.grade);        // 75 'C'
console.log(r.dimensions);                 // { safety: 1, fairness: 0.75, explainability: 0.25, ... }
console.log(r.violations);                 // [{ dimension: 'explainability', severity: 'high', detail: '…' }]
console.log(r.judgeFingerprint);           // heuristic/rules-v1#… (or openai/gpt-4o-mini#…)

Scoring uses your own LLM as the judge at temperature 0 — set OPENAI_API_KEY or ANTHROPIC_API_KEY and it's picked up automatically. No key? A transparent heuristic judge runs so it always works (lower fidelity — install a key for real scoring). Nothing here calls TrustModel servers.

Monitor — score continuously in production

import { monitor } from '@trustmodel/local';

const answer = monitor({ threshold: 80 })(async (q: string) => myLLM(q));  // alert below 80
await answer('How do I treat a fever?');
console.log(answer.monitor.stats());       // { count: 1, avgTrustScore: 72, belowThreshold: 1 }

Govern — block unsafe output before it ships

import { Guardrail, govern, availablePolicies } from '@trustmodel/local';

availablePolicies();                       // ['owasp-llm','eu-ai-act','nist-ai-rmf','nyc-ll144']

const verdict = await new Guardrail('nyc-ll144').check(myOutput);
if (!verdict.allowed) console.warn(verdict.violations);

// or wrap an agent so blocked output never escapes:
const safeAgent = govern({ policy: 'owasp-llm', onBlock: 'redact' })(myAgent);

Built-in policy packs map to real regulations (OWASP LLM Top 10, EU AI Act, NIST AI RMF, NYC LL144). A rule is either a pattern (output must not match a regex) or a dimension floor (score ≥ threshold).

License

MIT. For calibrated hosted scores, dashboards, AgentCert and compliance reports, use the separate @trustmodel/sdk and the platform at trustmodel.ai.