@tcl/core — Conversation Truth & Risk Intelligence

TCL is not an “agent training score.” It is a Conversation Truth & Risk Intelligence engine for human calls, AI assistants, and hybrid workflows: lies, misleading claims, hallucinations, unsupported assertions, compliance risk, drift, disclosure gaps, attribution problems, and latent business signals (objections, churn, confusion, KB gaps).

Product promise

TCL turns conversations into defensible truth, compliance, drift, hallucination, and business-value intelligence.

It helps teams answer:

1. What was said?
2. Who said it (person vs AI vs system)?
3. Was it true and appropriately supported?
4. Was it compliant with policy and consistent across the call?
5. What should we do next (compliance, AI policy, KB, product, not “coaching only”)?

Defaults (multi-industry)

• Default domain pack: general_conversation_integrity runs when no template / pack ids are supplied.
• ProtectQA final expense: pass options.template: "final_expense" (or insurance / protectqa) or include protectqa_final_expense in options.domainPackIds.
• scores.tcl / scores.overall: primary Conversation Truth & Risk composite.
• scores.truth: factual/support/safety (not transcript presence). transcriptGrounding is tracked separately.

options.skipProtectqaDefault is no longer required for a non-insurance default; the final-expense pack is not prepended unless you select it or pass explicit pack ids.

Key API additions

• scores: tcl, transcriptGrounding, compliance, hallucination, drift, evidenceSupport, speakerConfidence, businessValue.
• risk: primaryRisk, recommendedAction, businessImpact.
• dashboardSummary: pre-structured sections for trust score, risks, unsupported claims, drift, insights, next actions.
• claimsAnalysis + evidenceDependencyGraph: claim ↔ evidence expectations.
• businessInsights: objections, churn, confusion, ProtectQA-specific signals.
• issuesBySeverity: critical | high | medium | low.

Docs & examples

• Positioning & framework: docs/PRODUCT_POSITIONING.md
• Generic use-case notes: docs/use-cases/README.md
• Sample transcripts: examples/fixtures/

Build

npm install
npm run build
npm test

Import

import { validate } from "@tcl/core";
// or from dist after build

This package powers ProtectQA and any other product surface that needs auditable conversation intelligence—not vanity scores.