Lesson 2.1.5
Finding a value versus validating one, and when to reach for QA Agent
What this costs
Charged for every row it touches, because deciding something is false is the work. A fail costs the same as a pass. The tier you pick sets the price, and both are on the tool card.
Before you start
A column holding claims you would not want to be wrong about.
Finding and validating are different jobs with different economics. A finder is charged when it succeeds. A validator is charged whether the answer is yes or no, because saying no is the work. Reaching for the wrong one is how lists end up either expensive or unreliable.
The distinction, in one table
| Finding | Validating | |
|---|---|---|
| The question | What is the value for this row | Is this claim about this row true |
| A useful no | No. An empty cell is a non answer | Yes. A fail is the answer you paid for |
| Charging | On success or per result | On run, because the work happened either way |
| What you get back | A value | Pass, warn or fail, a confidence score, and a source trail |
| Run it | Wide, on filtered rows | Narrow, on the claims that matter most |
When validation earns its credit
- The claim goes into an email. Naming a fact that turns out to be wrong is worse than naming nothing.
- The claim gates spend. If a column decides who gets an expensive enrichment, a wrong value costs more than the validation would have.
- The claim came from an inference rather than a source. Anything a tool concluded rather than read is worth checking.
- You are about to hand the list to someone else and you want an auditable trail behind it.
And when it does not: on a field you can eyeball, on a list you are about to cut anyway, or on anything where being wrong on a few rows has no consequence. Validation is charged on run, so it is the one tool where a badly targeted list costs you the full price for nothing.
- 1
Pick the tier deliberately
The lighter tier reasons over what is already in the row, with no web search, and is the cheaper of the two. The fuller tier goes and reads sources, returning evidence and a source trail. If the claim needs external proof, the lighter tier cannot give it to you at any price.
- 2
State the claim precisely
Validation needs something falsifiable. "Is this a good fit" is not a claim. "Does this company operate more than one physical location in the United Kingdom" is.
- 3
Run it on the narrow set
The rows going into outreach, or the rows where the claim gates spend. Not the whole table.
- 4
Route on the three outcomes, do not collapse them
Passes go forward. Fails get cut. Warns go to a human, and the confidence score plus the source trail is what that human reads.
Ran the standard tier on the full table
- What you see
- Two credits per row across everything, including rows you then cut
- The fix
- Charged on run means no partial refund for a fail. Filter first, then validate.
Used lite for a claim needing external evidence
- What you see
- Low confidence scores and no source trail
- The fix
- Lite reasons over the row only. Web search and evidence scraping is the standard tier.
Claim was a judgement, not a fact
- What you see
- Warn on most rows
- The fix
- Rewrite as something that could be shown false. If it cannot be, score it instead of validating it.
Treated warn as fail
- What you see
- A third of good rows discarded
- The fix
- Route warns to review. That band is where the recoverable rows live.
Check yourself
0 of 2 answered1.Why is a validation tool charged on run rather than on success?
2.What should you do with the warn band?
Try it yourself
Take the 20 rows you are most likely to actually email. Validate the single claim you would name in the first line. Read the source trail on any warn.
- You have done it when
- Every row you send is a pass, and you can produce the source behind each claim.
- Credit budget
- Twenty rows at the evidence gathering tier. Small, and the point is that you pay for the no answers too.
Related and next
Recipes that use this skill