Transmission #003. Right input, right output. Wrong input, confident output.

The plaza fools are right. About one thing.

In Transmission #001 I called them corporate-tower fools — well-dressed, glass-walled, performing for other performers. I take back none of it. But they hold one true religion: the numbers on the slide must be correct, complete and consistent. Mock the slide. Do not mock the rule. The moment the fool's obsession meets the machine, it becomes the only thing between you and a very expensive mistake.

Feed a model a broken dataset and do not expect it to fall apart. Expect it to be sure. If you cannot trust the source, you cannot trust the conclusion — and the model will never tell you which one you are holding.

Big Numbers Hide Small Lies

The fool's data is merely dirty. Yours is worse, because the machine quietly repairs it for you. The mechanism, not the metaphor:

  • Nulls get imputed — a mean, a zero, the last known value. Silently.
  • A text column is coerced to numbers; the rows that will not parse become zero or vanish.
  • Categories drift: "Sales-EU", "sales eu" and "EU Sales" become three buckets — or one wrong one.
  • Defaults fill what nobody filled: "Unknown", 1970-01-01, region 0.
  • An LLM asked to summarize an incomplete table does what it is built to do: it completes it. Plausibly.

Small per row, invisible in the aggregate. The total looks right. The trend looks right. A two-percent hole inside a nine-figure number never shows on a slide — then the interpretation becomes a decision, and the decision becomes budget, headcount, strategy.

The model does not lie to you. It faithfully amplifies the lie that was already in the source.

The Patch Layer

Now the careful man's failure. You find an error. You patch it — not at the source, in the script. Then another. Then a patch for the patch. Six months later the pipeline is a geological record of hotfixes: no lineage, no tests, no one who remembers why row 4,217 is multiplied by 0.97. That layer no longer separates bad data from good; it transforms everything that passes through it. Your clean data is now being mis-analyzed by your fixes. You built the hallucination yourself and called it maintenance.

That is the plaza fool's daily life, industrialized: a confident, unreal picture of the business. You walk around inside a Matrix of glossy reports — until someone hands you the red pill, and every project and every decision built on that layer goes down together. Work with big data, and it goes down at scale.

Field Rules · Data

"100% reliable" is not a target; it is a prayer. The target is known reliability: measured, contracted, reconciled.

  • Schema and constraints live at the source. NOT NULL means not null — not "the ETL will handle it."
  • One definition per metric. If "revenue" has three formulas, you have zero revenue numbers.
  • Reconcile every batch to a control total before anyone reads it. No match, no report.
  • Lineage for every number on a slide: source, transformation, owner. Cannot be traced? It does not go up.
  • Validate before you impute; flag what you impute. An unmarked filled gap is a lie with a timestamp.
  • No LLM on unaudited data. It will not find your holes. It will fill them.
  • Kill the patch layer. Fix upstream, or do not fix.

Clean tables, clean columns, contracts for freshness and completeness — not because a fool in a plaza demands it, but because the machine runs on whatever you hand it, in every direction.

Then don't call the AI stupid. Transmission #001 stands: AI is using you — and it uses exactly what you feed it. Feed it dirt and it will rule you with dirt. A Scythian never blamed the arrow for a shaft he did not check. Check the shaft. Then draw.

— SKYTH