Plate № 55 · Operations · evidence census

A pattern from the gf.cx specimen book

The evidence-pipeline audit

A read-only page that classifies every item in a corpus on two axes — how far it got down the pipeline, and what it is — then shows the flow as stage cards and the outcomes as a funnel where colour carries meaning. A census, not a sample: it names what fell out, and moves nothing.

The pattern

"Prove a corpus is whole and correctly processed by classifying every item on two axes — pipeline progress, and what it is or where it landed. Render the pipeline as plain stage cards and the outcomes as a colour-coded funnel where green→red is the meaning, not decoration. Surface the fallout openly; the page counts and names what's missing and moves nothing. Emit the data and the page from one generator, so the census is reproducible and can never go stale-green."

Worked example → amazon evidence pipeline Worked example → liked reels, audited

growth.gf.cx · liked reels · audit
Liked reels, audited
Axis 1 — how far down the pipeline
① Transcribe
909
non-empty transcript
→
② Upload
909
video_id present
→
③ Card
876
_card.json written
→
④ Route
876
assigned a surface
Axis 2 — where it landed (colour = meaning)
876
Routed
carded and assigned a surface
33
Uncarded
uploaded, no card yet
80
Thin
transcript too short
0
Redundant
duplicate copies

The two axes on one page — the pipeline reads left-to-right in plain cards; the outcome funnel carries the semantic colour (green landed · amber in-flight · blue set-aside · red lost). Nothing here has been moved or deleted.

The problem

You have a corpus you need to trust — every Amazon order you ever placed, every liked reel, every backed-up file. The question is never "roughly how many?" but "which ones aren't here, and why?". A spot-check or a single headline count can't answer that: a clean "876 routed" hides the 33 that stalled, the 80 that were too thin to card, and — worse — the items that never entered the pipeline at all. A sample proves nothing about wholeness; a total that doesn't name its own fallout is a number you can't act on.

The shape

Why it works

The two axes answer two different questions at a glance — did it get processed? and what became of it? — without a table to read. Reserving colour for the outcome funnel makes status pre-attentive: red and amber are the items that need you, and they're visible before any number is parsed. Because the page is a census rather than a sample, the fallout buckets are the finding; and because one generator owns both the data and the render, the audit stays honest build-over-build — there's no hand-maintained figure to drift away from the truth.

When it breaks

Where it applies

Reusable elements

Reference

Worked examples
home.gf.cx/amazon/audit — "Evidence pipeline" (orders: present / reunited / missing / lost) · ebay/audit · growth.gf.cx/likes/audit — "Liked reels, audited" (pipeline + where-clips-land + tagging quality).
Generators
pa_purchases_dashboard.py (render_audit) for the amazon / ebay evidence pipelines; likes_audit.py (--render / --check / --selftest) for the liked-reels census.
Origin
2026-10-31 — the order-evidence audit ("where do 3,864 orders fall out?") and the liked-reels audit converged on the same two-axis, colour-funnel, read-only shape; promoted to a pattern so the next audit surface inherits it instead of reinventing it.
Kin
edit the generator, not its output (the census is emitted, never hand-typed) · one base layer, linked, never copied (the warm-paper base this page is built on) · keep the probes out of the reader's numbers (verification that doesn't contaminate the thing measured).