Verification

Definition

Verification is checking whether something satisfies the relevant constraints.

Simple form:

candidate -> check -> accept, reject, or revise

The candidate may be an answer, plan, calculation, tool result, or intermediate step.

Verification tests whether a result should be trusted.

---

Why Verification Matters

Reasoning models can produce plausible but wrong outputs.

Verification adds a checking step.

Examples:

does the answer follow the instructions?
does the calculation work?
does the code compile?
does the claim match the source?
does the plan satisfy constraints?

Verification reduces unchecked error.

---

Verification vs Generation

Generation creates a candidate.

Verification evaluates it.

generation:
produce possible output

verification:
check output against requirements

These are different roles.

A system can be better when it separates them.

Producing an answer and checking an answer are not the same operation.

---

Internal And External Verification

Verification can be internal.

Example:

model checks consistency using its own representations

Verification can be external.

Example:

run tests
use a calculator
check a database
compare against a source

External verification is often stronger when the tool is reliable.

Verification can happen inside the model or through external systems.

---

Verification Is Not Perfect

Verification can fail.

Examples:

wrong check
missing constraint
bad tool result
false confidence
incomplete evidence

The quality of verification depends on the quality of the checking process.

Verification improves reliability, but it is not automatic certainty.

---

Final Summary

• Verification checks candidates against constraints.
• It is different from generation.
• It can apply to answers, plans, tools, or intermediate steps.
• It can be internal or external.
• Good verification reduces errors but does not guarantee perfection.