Chapter 11 — How Do You Confirm AI's Output?

Separating what an AI said from the external facts

So far, we've looked at how AI generates output, how it uses external information and Tools, and what limitations it has. From here, we consider how to confirm AI's output when using it in actual development work.

The first thing to keep in mind is the distinction that what an AI said and what could be confirmed as an external fact are different:

the AI said so
        !=
confirmed as an external fact

Suppose an AI answers, "I ran the tests. All PASS." What you can directly confirm from that sentence is:

that the sentence was generated.

Whether the tests were actually run, and whether the result was actually PASS, need to be confirmed separately.

Claim and Evidence

When confirming, rather than reading an AI's text as one lump, look at what Claims (assertions) are contained within it. For example, a report saying:

I changed the configuration.
I also ran the tests.
All 43 returned PASS.

contains at least the separate Claims Claim 1: the configuration was changed, Claim 2: the tests were run, and Claim 3: 43 items returned PASS. Each is not necessarily confirmable with the same Evidence.

Doesn't a generated report count as Evidence at all?

It's not that an AI's report is Evidence for nothing at all. If an AI outputs "43 items returned PASS," the fact that the AI output that is Evidence — but for the fact that it produced that output. That is separate from independent Evidence that 43 tests actually returned PASS. In other words, evidence that the report was made and evidence of the external fact are different:

Evidence of the report
        !=
Evidence of the external fact

Independent Evidence

We treat as independent Evidence anything that lets you confirm the object separately from the AI's own self-report. Depending on the situation, things such as:

• an actual file in the repository
• a Git diff
• commit history
• a test execution log
• command output
• runtime observation
• an authoritative document

can serve as confirmation material. However, no single one of these is always a universal Evidence source. What Evidence is needed depends on "what you want to confirm."

Strong wording does not add more Evidence

Even if an AI writes strongly — "I'm certain," "no mistake," "fully confirmed" — the words themselves don't increase external Evidence:

confidence wording
        !=
additional Evidence

This doesn't mean "you should never use confident-sounding language." It means keeping the strength of the wording separate from the strength of the Evidence.

Multiple AIs agreeing doesn't necessarily mean external Evidence

Suppose you ask the same question to two or three AIs, and they all return the same answer. That can serve as a useful review signal. But multiple AIs agreeing and an external fact being independently confirmed are not the same thing:

multiple AIs agreed
        !=
the external fact was independently confirmed

If multiple AIs point out the same issue, that can serve as material for judging "this part is worth checking more carefully." But it alone does not settle the external fact.

VERIFIED / OBSERVED

This book divides confirmation status through Evidence into three. The first is VERIFIED / OBSERVED. This represents something directly confirmed through an authoritative, confirmable source, or a reproducible observation — for example:

actually read the repository
actually observed command output
confirmed the relevant part of an authoritative document

That said, what counts as sufficient confirmation depends on the object.

INFERENCE

INFERENCE is a conclusion drawn from Evidence. For example:

observation A
observation B
   ↓
C is likely

In this case, even if A and B themselves are OBSERVED, if C is not directly observed, C is treated as INFERENCE.

What matters here is that INFERENCE != baseless speculation. An INFERENCE has supporting Evidence — but the conclusion itself was not directly confirmed.

UNKNOWN / UNDISCLOSED

When something cannot be confirmed from the available Evidence, we treat it as UNKNOWN / UNDISCLOSED. There are different reasons something can become UNKNOWN — for example:

• it hasn't been checked yet
• it was investigated but the Evidence is insufficient
• the information is not public
• there is no way to confirm it under current conditions

There's no need to fill all of these in with "it's probably such-and-such."

UNKNOWN is not FAIL

This is very important: UNKNOWN != FAIL. Being unable to confirm something does not let you say it does not exist, was not executed, or is wrong. Conversely, UNKNOWN != PASS as well. It matters not to convert an unknown state into either success or failure at will.

Read PASS with its scope attached

A test result's PASS is not an unconditional "everything is correct" either. PASS is interpreted together with at least:

• target
• version
• test set
• conditions
• observation point

For example, it means:

for this target,
for this version,
against this test set,
under these conditions,
observed at this point in time, it returned PASS

Therefore, PASS does not mean the system as a whole is complete, safe, production-ready, or that UNKNOWN has been reduced to zero:

PASS != the system as a whole is complete
PASS != safe
PASS != production-ready
PASS != zero UNKNOWNs

PROJECT-METHODOLOGY is not a fourth class

This book uses three Evidence classifications:

VERIFIED / OBSERVED
INFERENCE
UNKNOWN / UNDISCLOSED

Meanwhile, a rule such as an AI's self-report alone does not settle an external fact is a Verification approach this project has adopted. We treat this as PROJECT-METHODOLOGY.

PROJECT-METHODOLOGY is not a fourth class alongside these:

VERIFIED
INFERENCE
UNKNOWN
PROJECT-METHODOLOGY

The kind of question is different. Evidence classification represents how much has been confirmed. PROJECT-METHODOLOGY represents how this project chooses to handle Evidence — a rule.

Chapter summary

AI report != external fact
confidence wording != additional Evidence
multiple-model agreement != independent external Evidence

Evidence classification:
VERIFIED / OBSERVED
INFERENCE
UNKNOWN / UNDISCLOSED

• separate AI's output from the external fact
• consider, per Claim, what needs to be confirmed
• distinguish Evidence of the report itself from Evidence of the external fact
• strong wording alone does not add Evidence
• multiple AIs agreeing can serve as a review signal, but it is not independent external Evidence itself
• an INFERENCE has Evidence, but the conclusion itself was not directly confirmed
• UNKNOWN is neither FAIL nor PASS
• read PASS as limited to target, version, test set, conditions, and observation point
• PROJECT-METHODOLOGY is not a fourth Evidence class

The next chapter looks at what should be treated as the Source of Truth when confirming a Claim.