Skip to content

README.md

Latest commit

聽

History

History
160 lines (102 loc) 路 3.53 KB

README.md

File metadata and controls

160 lines (102 loc) 路 3.53 KB
Raw

Module 0: Prerequisites - Trust-Boundary Basics

Duration: 20 minutes
For: Developers new to LLMs, verification, or trust-critical AI systems
Goal: Build the right mental model before touching any QWED API

What You'll Learn

  • Proof vs. confidence
  • What an LLM actually does
  • Why hallucinations are inevitable
  • Probabilistic vs. deterministic systems
  • Why fail-closed verification matters

Start Here First

Before anything else, read:

These two lessons establish the distinction that drives the entire QWED ecosystem:

  • parsing is not proof
  • simplification is not verification
  • confidence is not evidence
  • unsupported is not approved

1. What Is a Large Language Model?

An LLM is a text prediction system trained on massive corpora.

It does not "know" facts the way a deterministic engine proves facts. It predicts likely continuations.

Example

Prompt: "2 + 2 ="
LLM predicts: "4"

Prompt: "2843 + 7291 ="
LLM predicts a likely continuation, which may still be wrong

What LLMs Are Good At

  • drafting language
  • summarization
  • translation
  • extracting structured information
  • conversational interaction

What LLMs Are Bad At

  • exact arithmetic
  • formal logic
  • safety-critical execution decisions
  • distinguishing a plausible answer from a proved answer

2. What Are Hallucinations?

A hallucination is a plausible-sounding output that is still wrong.

This happens because LLMs optimize for pattern completion, not truth.

Examples of Hallucinations

  • wrong calculations
  • fake legal citations
  • invalid medical dosages
  • invented policies, studies, or references

3. Probabilistic vs. Deterministic Systems

Probabilistic Systems

LLMs are probabilistic:

  • they generate likely outputs
  • they may vary across runs
  • they can be helpful without being provable

Deterministic Systems

Deterministic engines:

  • follow exact rules
  • produce the same result for the same input
  • can prove or reject claims within supported domains

Why QWED Uses Both

QWED uses the LLM as an untrusted translator and the deterministic engine as the trust decision layer.

User query -> LLM translation -> deterministic verification -> verified / invalid / unverifiable

4. Why Verification Is Critical

Without verification:

  • bugs ship to production
  • users inherit silent trust failures
  • "helpful" outputs can cause financial, legal, or safety harm

With verification:

  • supported claims can be checked deterministically
  • invalid claims can be blocked
  • unsupported claims can be surfaced honestly as UNVERIFIABLE

5. What Fail-Closed Means

When QWED cannot establish proof, the answer should not silently degrade into:

  • a fallback guess
  • a lower confidence answer
  • a default value that looks safe

The right outcomes are:

  • BLOCKED
  • UNVERIFIABLE
  • QUARANTINED
  • HUMAN_REVIEW_REQUIRED

This is what makes QWED a trust-boundary system rather than just another AI helper.

Quick Check

  1. What is the difference between a useful answer and a verified answer?
  2. Why is confidence not the same thing as proof?
  3. What should happen when a claim is unsupported?
  4. Why is "safe default" often the wrong pattern for trust-critical AI?

If you can answer those clearly, you are ready for the rest of the course.

Next