The words spoken in the Unit 16 video, scene by scene, as its captions show them, with what the animations show marked [On screen: ...] where they play. Each speaker is named where the speaker changes. The Unit 16 reading covers the same material.
The narration is spoken by synthetic voices made for this course, each shown on screen as a drawn robot. No person is speaking.
Unit 16 — The historical bug and historical honesty
Narrator: This unit is part of the complete route, and required there. The other routes leave it out, and the claim-status lesson it works through is also recorded in the course's list of misconceptions. It is here because sorting claims once, on a real document, teaches more than reading about it.
What the source is
You now have two things you lacked at the start of this course: a Bernoulli value you worked out yourself, and an oracle, a trusted reference program, that you can run whenever you want. Here you point both at something much older than either: a table printed in 1843. The point is not that the table is interesting, although it is. It is what happens to a number once it has been printed in a respected place. Here is what this course treats as established. In 1843 an English translation of Menabrea's account of Babbage's Analytical Engine appeared in volume three of Taylor's Scientific Memoirs, translated by Ada Lovelace, who appended a series of her own notes, lettered A to G. Note G is the last of them, and contains a table setting out, step by step, how the engine would compute a Bernoulli number.
Which number the table computes
The number it works out is the one Lovelace labels B-seven.
[On screen: Headed 'Same symbol, two different numbers', two number lines run across the frame, one above the other. The upper line is modern Bernoulli indexing: B0 = 1, B1 = -1/2, B2 = 1/6, B3 = 0, B4 has a value to derive, B5 = 0, B6 = 1/42, B7 = 0, and B8 is withheld. The lower line begins as an exact copy, aligned index for index. Then the lower line changes. The first two terms, 1 and -1/2, fall away, marked 'not numbered in Note G', because Note G's numbering begins at 1/6. Then the three zero terms at B3, B5 and B7 fall away, each marked with the word 'zero', because Note G numbers only the non-zero terms. Four terms remain. They are renumbered as Ada's B1, B3, B5 and B7, and the labels visibly slide one place to the right as they arrive, so the two lines are no longer in step. The animation ends with Ada's B7 highlighted directly beneath the modern B8, and the statement that Note G computes Ada's B7, which is the modern B8: the same symbol, two different terms, because the modern B7 is 0. A last line warns: ask for B7 without saying which indexing, and the answer you get back may be the wrong term.]
Her labels start at what we call B-two, and from there she numbers only the values that are actually there, the non-zero ones, using the odd numbers. So each of her labels is one less than the modern index: her B-seven is the modern B-eight. That is the mapping you met in Unit 6.
The one claim this unit asserts
The published table contains at least one error. That is established. Notice how little that sentence says. Not where the error is, not what kind of error it is, not whose it was. Each of those is a separate claim, and each has to earn its status on its own. Listen to what comes next: the shape of it matters more than the content.
One claim checked, one still refused
You will meet one description of this error everywhere. It says that one division in the table has its two numbers the wrong way round: the number being divided and the number it is divided by are swapped. For a long time this course treated that as contested, on honest grounds: it had checked the description only against others repeating it, not against the printing. And repetition is not evidence.
Questioner: Has anyone checked it against the printed page?
[On screen: The 1843 Note G diagram. One row, operation 4, is lifted out and enlarged, showing that it acts on V5 and V4 in that order while stating a result that requires dividing V4 by V5. A panel headed 'Established' says the operands are inverted, checked against the printed diagram, not against other accounts. A label says what the copy shows: in this copy, a processed image of the printed plate, the division sign is legible too. A separate panel headed 'Contested' asks whose slip it was -- compositor, Lovelace, Babbage -- and marks nothing on the table, because nothing on the page can settle it.]
Checker: Then the check got done. A facsimile, an exact image of the printed 1843 diagram, was examined. An error is in operation four. It divides V-five by V-four, but its stated result needs V-four divided by V-five. The two numbers are the wrong way round, exactly as the common account says. The copy used is a processed image of the printed plate. In it, the location, the operand order and the division sign itself are all legible.
[On screen: Two claims sit side by side, both labelled widely repeated: that operation 4 has its divisor and dividend the wrong way round, and that the slip was the compositor's. Three buckets appear below, established, contested and unsourced. The first claim is compared against the printed diagram, its label changes to checked against the source, and it moves down into established. The second is examined: who introduced the slip is not something the printed page can settle, no check is available from it, and the claim moves into contested still labelled widely repeated. The closing line reads: neither claim moved because it was repeated, one moved because it was checked.]
Narrator: So that claim moved into established. Not because it is popular, but because the printed page was finally checked.
Checker: What has not moved is who put the error there. Lovelace, Babbage, the printer: three candidates get named, and no facsimile of the printed page can settle between them, because the page shows the result, not the process. The attribution stays contested. And the common account that it was a typesetting error is an inference, not a finding, so it stays contested too.
Narrator: Notice the difference: the check that was available got done and one claim moved; the check that is not available has not, and that claim stays put. That is the whole lesson.
Why it could survive
An error in a private notebook is caught or not, and either way it stays small. This one was different. It was in print, in a respected series, attached to two well-regarded names, and it concerned a computation few readers would redo by hand. Each of those properties would plausibly make a table less likely to be checked, not more. That is a possible explanation, not a record. This course has no source on who read the table, or how closely. If readers did pass over it, they need not have been careless. Readers have limited time. It is reasonable to spend less of it on a source that others have already approved.
Checker: How long the error went unremarked, and who first published a correction, remain unsourced in this course. They are open questions, not facts to repeat.
This course's own case
Narrator: This is not only a nineteenth-century problem, and the clearest example is this course's own. Across the web, the table as printed is said to compute minus twenty-five thousand six hundred and twenty-one over six hundred and thirty. This course recomputed it, got one hundred and thirty-nine over six hundred and thirty, and published that as a refutation. Both figures are reproducible; neither claim was complete.
[On screen: A card headed 'Note G table, as printed' lists more than one printed fault: operation 4, as printed, and operation 24, as printed, on Glaschick's reading, citation S7. Below them: earlier B values, supplied or generated? Operation 24 changes to repaired, marked 'not said'. Two arrows fork from the table. The upper run, with operation 4 as printed, operation 24 repaired and the earlier values supplied, gives 139/630. The lower run, the same routine with operation 24 still repaired and the earlier values generated by the routine itself, a dashed arrow feeding its output back into the table, gives -25621/630. Then the conditions in both boxes are struck through and dimmed: each figure was reported as what 'the table as printed' computes, the upper by this course, the lower across the web, and neither said which object was being run. The bare figure -25621/630 is copied onto six web pages, one after another. A box headed 'retrieval can amplify it', citation D4, and labelled illustrative, shows a cited answer: 'The table as printed computes -25621/630', followed by six citation numbers that light up on the six pages in turn. The last line reads: its repetition looking like corroboration.]
The table carries more than one printed fault. Operation four divides the wrong way round. And on Glaschick's reading, operation twenty-four, as printed, leaves out a change of sign that the computation needs. And the earlier Bernoulli numbers it uses can be supplied, or generated by the routine itself. Supply them, and you get one hundred and thirty-nine over six hundred and thirty, but only having repaired operation twenty-four without saying so. The larger figure is the same routine, operation twenty-four still repaired, feeding its own output back in. Neither account said which object was being run: which version of the table, under which conditions. Retrieval can amplify a figure like this rather than correct it. One major assistant, as of November twenty twenty-six, may search the web without being asked. A search returns what pages say, so a figure repeated across the web can come back as a cited answer, its repetition looking like corroboration. The fix is not more care. It is stating the interpretation, then recomputing under it.
The same shape, today
Now change one variable and leave the rest. Fluent generated text can have the same surface properties. An unsupported citation can be as polished as a correct one, so style and confidence are insufficient evidence. Some systems expose sources or uncertainty signals, but those still need checking for the task at hand. The historical analogy is limited: modern systems differ in scale, speed, variability, and attached tools. The shared problem is narrower. A polished claim does not establish its source. And notice who made this error. Every candidate is a person: in 1843 there was nobody else. The problem is older than AI. What generation adds is mainly speed and volume, a plausible table in seconds, while checking one still takes the minutes it took then. What settled the operation-4 claim here was the facsimile: a source anyone can open and read for themselves.
Sort into three buckets
The habit worth taking away is not scepticism. Blanket scepticism is as useless as blanket trust and more tiring. It is sorting. Every claim in front of you goes into one of three buckets.
[On screen: Three buckets with their rules. Established: you can name the source, and someone accountable looked at it. Contested: sources disagree, or one account copied forward with no check. Unsourced: no source you can name yet; no weight, not the same as false. A claim stating "Note G has an error" drops into Established. A claim stating "Ada made the error" drops into Contested. The closing line reads: repetition is not corroboration.]
First, established: a primary source, or consistent secondary sources with no serious dispute, and one you or someone accountable to you has actually looked at. Second, contested: sources disagree, or a single account has been copied forward with no independent check behind it. Repetition is not corroboration. Third, unsourced: you cannot currently say where it came from at all. That is not the same as false; it is a claim with no weight yet. Most claims that go wrong do so because bucket two and bucket three were quietly filed as bucket one. And for any claim, say what would move it: a specific thing a person could go and consult.
Your turn
The activity has two parts. First you compare the historical table against a value you can now produce yourself, and say what the comparison actually catches. Then you take a repeated claim about Note G and sort it into one of the three buckets. Then, harder, say what evidence would move it. Then run checkpoint cp16-note-g.
Closing, 20 seconds, no speech. A candidate record on the left, a check against an independent reference in the middle, and a bracketed claim on the right fade in side by side. Below them: Continue with the course material, ai.qyber.dev/edu-ai-literacy-ada-verification/, Course playlist: youtube.com/playlist?list=PLPsBWTfgvakM, All links are also in the description, and Keep your record. Check the next claim.