The pipeline reproduces from the working data. Is that not enough?

It proves the second half of the journey. Reproducing from the archived export is the claim that matters, because the working data is precisely where the unrecorded decisions were made.

What has to be archived?

The raw bytes, under a name never reused, never edited in place, and never copied into a report directory or a chat message. The path is what code reads, and the file itself stays in the landing zone.

Does an interactive step ever have a place?

For exploring, yes. For anything a published figure rests on, the step becomes a script before the figure ships, or the figure is published labelled unverified with the reason stated.

Failure catalogue · The wrapper betrays the numbers · 49 of 51

OBSERVED FAILURE MODE

The step nobody can run again.

The first transformation, raw export to working data, is often done by hand in one session: a few edits, a filter, a saved copy. When that step is never written down as code, no figure downstream can be reproduced from the source, however clean the rest of the pipeline is. Provenance ends at the gap, and the gap is invisible.

What we saw

In a Meridian Journal Group workspace, the path from raw export to working data was done interactively and never preserved. Everything after it was scripted, gated and reproducible: the pipeline ran cleanly from the working data, and every anchor tied. Nobody could show how the working data had been produced from the export sitting in the archive beside it. The pack answered how do I know these are right with a chain that stopped one link short of the bytes, which is the same as not answering.

Why it passes a glance

The manual step happens before anything is under review, and it leaves behind a file that looks exactly like a generated one. Everything downstream reproduces perfectly from that file, so a reproduction test goes green while proving only the second half of the journey. The gap is visible in the folder layout, which is where nobody looks for a correctness defect.

What addresses it

The workspace-architecture skill draws the boundary that prevents it: raw stays immutable in the landing zone, processed is regenerable from raw by code, and merging the two means you cannot re-derive after fixing an extractor bug or prove which source file produced a published figure. Doctrine principle 1, An AI never authors a number, requires plain inspectable code for every figure, and a step only a person can repeat authors the input to all of them.

Check your own file in two minutes

Delete the working data, run the pipeline from the archived raw bytes, and confirm the figures come back identical.
Confirm every raw file is stored under a name never reused and never edited in place.
Name the script that produced each processed file, and write one for any file that cannot name one.

What this does not catch

Reproducing the transformation proves the numeric path re-runs to the same values, not that the values are right. A wrong column or a wrong filter reproduces perfectly, forever. The definition printed beside each number, not the reproduction, is what addresses meaning.

Quick answers

The pipeline reproduces from the working data. Is that not enough?: It proves the second half of the journey. Reproducing from the archived export is the claim that matters, because the working data is precisely where the unrecorded decisions were made.
What has to be archived?: The raw bytes, under a name never reused, never edited in place, and never copied into a report directory or a chat message. The path is what code reads, and the file itself stays in the landing zone.
Does an interactive step ever have a place?: For exploring, yes. For anything a published figure rests on, the step becomes a script before the figure ships, or the figure is published labelled unverified with the reason stated.

Nearby failures

The hash that disproves itselfThe verification is theater The estimate outside its own intervalThe wrapper betrays the numbers The refresh time bombThe wrapper betrays the numbers

Last updated 2026-09-02 · Dylan, founder · one of 51 observed failure modes, every one seen in a real build or in our own audits, none invented.

Get the free skill → The doctrine All 51 failure modes