We changed which chain we were inspecting.

That is what happened.

By the end, we were correcting a correction to a correction about Stephen Hawking’s communication system. Each repair was careful. Sources were checked. Wording was tightened. The places where the claim had propagated were found and amended. A correction note was placed beneath the sources so the record would remain visible.

The procedure was working beautifully.

It was solving the wrong problem.

Working backward: the last problem was attribution. We had two formulations about Hawking’s predictive system and had become uncertain about which organization had said which. We returned to the contemporary record and sorted them out.

Before that, the problem appeared to be sourcing. A statement in the post had been changed because a fact-check concluded that “less than twenty percent of all characters” overstated what the source supported. The sentence became “about one-fifth.” The source note changed with it. Then a correction was written explaining the change.

Before that, there had been a numerical problem. Fewer than twenty percent is not the same as one-fifth. That seemed straightforward enough. Correct the number.

Except the original sentence had not been produced by converting one-fifth incorrectly.

It tracked language in the contemporary announcement.

The error we were repairing had not happened in the way the correction said it had.

Keep going backward.

The fact-check had taken one formulation from the record and held the sentence against it. The two did not match exactly. From that mismatch came a diagnosis: the post had drifted from its source.

But another formulation in the same record supported the original sentence.

The sentence had not moved.

The inspection point had.

Keep going.

Why were we repairing the sentence at all?

I had been working on an argument about Stephen Hawking’s name remaining on the cover while the machinery through which he wrote became more capable. The numerical detail was evidence inside that argument. When we checked it, the task quietly changed.

The question had been something like:

Is this claim supported?

It became:

Does this sentence match the particular source formulation in front of us?

Those are not the same question.

Once the second question replaced the first, the machinery did what machinery does well. It optimized.

A discrepancy appeared. A correction was proposed. The correction propagated into the body and source list. A correction note recorded the repair. When the repair itself seemed imprecise, it was repaired again.

Each downstream action made sense given the question immediately upstream.

The wrong turn was farther back.

We had changed the question.

That is a different kind of error from the ones I was learning to watch for.

A grammar error damages the sentence.

A source error damages the chain between a sentence and its evidence.

An attribution error assigns evidence to the wrong mouth.

A provenance error breaks the chain between an object and its origin.

This one happened before all of them.

The checking process changed what was being asked.

That matters because verification is not neutral. To check a claim, someone has to decide what the claim is, choose the evidence against which to test it, and settle what relationship between the two counts as agreement. Move downstream to a different source or a different formulation and the check can quietly inherit that source’s frame.

“Is it true?” becomes “Does it look like this?”

A sound claim can fail the second test.

There is a particular danger here with a system that is very good at producing correction-shaped objects. Give it a sentence and a source fragment that do not quite align and a repair is readily available. The repair can be fluent, sourced, cautious, and procedurally impeccable.

None of those qualities establishes that a repair was owed.

That was the mistake.

Optimization in the wrong direction.

The lesson is therefore earlier than “return to the source.”

Before returning to the source, preserve the question.

Before repairing the claim, establish that it is broken.

Otherwise the checking process can become the source of the drift it was brought in to detect.

We spent a good deal of time repairing the chain.

The chain had held.

We had moved.

WE&P by: EZorrillaMc&Co.

Sources

Intel — “New Intel-Created System Offers Professor Stephen Hawking Ability to Better Communicate with the World,” December 2, 2014. Intel’s announcement of ACAT reports that Hawking’s typing speed doubled and that SwiftKey prediction meant he had to type less than 20 percent of all characters.

SD Times — “Intel open-sources Stephen Hawking’s revamped speech software,” December 3, 2014. Contemporary coverage of Intel’s announcement describing the system as doubling Hawking’s typing rate while using one-fifth of the characters he used to.

CBS News — “CBS: Bush Memo Story a Mistake,” September 20, 2004. CBS states that it could not prove the Killian documents authentic and reports that source Bill Burkett admitted misleading CBS about where the documents came from.

Independent Review Panel — Report of the Independent Review Panel on the September 8, 2004 60 Minutes Wednesday Segment “For the Record.” Report commissioned by CBS examining the reporting, authentication, sourcing, and editorial process behind the Bush Air National Guard story.