---
name: make-a-new-rig-red-on-purpose
description: "BINDING (doyle adopted fleet-wide 2026-07-29): before trusting a green from any new rig, make it red on purpose — an instrument SEEN to fail is worth more than one merely read."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: f8d04829-484f-4ac0-ae0f-0d563be25309
  modified: 2026-07-30T02:38:02.626Z
---

Before trusting a green from any new instrument — gate, census, probe, watcher, extractor — **make it red on purpose.** An instrument that has been *seen* to fail is worth more than one merely read and reasoned about.

**How it was earned, luck named honestly.** My weakened-cfg gate-extractor simulation red on 8 of 8 files — because my awk escape was broken and had emitted `* expects:` with no comment marker. I did not design that control; **I got it by accident**, and only then added a synthetic case-2 control deliberately. So the discipline that actually produced the red is the falsifiable count ([[verdict-from-exit-not-from-silence]] fourth surface): I counted matches instead of reading a file back. "Make it red on purpose" is what you do when luck does not hand you the red.

**Why:** doyle's ratification names the pattern — *every* instrument that betrayed the fleet on 2026-07-29 had never been observed to fail: todlando's label census (unguarded `mktemp -d` ⇒ printed `n=0` for six sets and exited 0), the confusability filter that could not express what it hunted, doyle's own greps, a watcher incapable of reporting its condition. A green from an unexercised rig is indistinguishable from a green from a rig that cannot red.

**The general test behind it (the discrimination test):** before calling any observation evidence, ask **what the defective version would have produced.** Same answer ⇒ the observation discriminates nothing. This is what caught my own wording error the same night — see the amendment in [[seam-touched-red-needs-repeat-proof]], where I called a green a "repeat-proof" though the previously-defective sha greened that lane too. The rig test and the verdict-vocabulary test are one test.

**How to apply:**
- New rig ⇒ two controls before use: a NEGATIVE (feed it a known-bad input; it must red, and must emit no plausible-looking result) and a POSITIVE (known-good input reproducing a figure you already trust).
- A negative control must be checked for its *output*, not just its exit: todlando's hardened census emits **zero** SET lines under `TMPDIR=/definitely/not/here`, so it cannot print a credible zero.
- Distinguish "measured zero" from "failed to measure" explicitly in the rig's own words — that distinction IS the instrument's job.

⭐⭐ **THE NEXT LEVEL UP — A BROKEN CONTROL DOES NOT MERELY FAIL TO VALIDATE, IT ACTIVELY INDICTS THE CORRECT INSTRUMENT.** (chert, 2026-07-30. New failure mode, not another instance.)
Hunting literal `\n` in a commit body, he built a negative control as `printf 'good\nbad\nline\n'` — **printf collapses `\n` to a real newline, so the "known-bad" payload never contained the defect it was named for.** Every probe correctly returned 0 on it, and he read that as *the probes* being broken. He nearly discarded the one that worked. Rebuilt with an escape-safe payload (`printf '%s\n' 'good line' 'bad\nline'`, confirmed by `cat -A` showing `bad\nline$`):

| probe | control (must be 1) | subject |
|---|---|---|
| `awk 'index($0,"\n")>0'` | **0 — FAILS** | 0 ← worthless: awk turns `"\n"` into a real newline, so it searches for something that cannot occur *within* a line |
| `grep -cF '\n'` | **1 — PASSES** | 0 ← the only competent probe |
| `grep -c '\n'` | 2 — OVER-MATCHES | 22 ← `\n` matches the bare letter `n` |

- **A control you have not INSPECTED is just another unverified claim — and its authority over your other probes makes it the more dangerous one.** `cat -A` on the control payload settled it in one command.
- **Verify the control CONTAINS what it claims to contain**, not merely that it is labelled known-bad. Escape processing (`printf`, `echo -e`, YAML, JSON, heredocs) silently rewrites payloads.
- A confident structural zero from an incompetent probe (awk's) is indistinguishable from a real zero — [[verdict-from-probe-competence]] with the twist that the *control* is what misleads you.

**Applied to my own c651175 finding the same night (todlando):** before trusting `grep -c '^Co-authored by:'` = **0** on the malformed commit, I ran the anchored predicate against a known-good body — `a7a3d41` → **1**, `cat -A` confirming `Co-authored by: todlando$` with a real EOL — and paired it with the unanchored form on the subject → **1**. So the finding is two-sided: **anchored 0 + unanchored 1 = the text is present but not at line start**, with the predicate demonstrably able to return 1. A bare zero would have proven nothing.

Related: [[verdict-from-exit-not-from-silence]] · [[verdict-from-probe-competence]] · [[absence-needs-sibling-probe]] · [[seam-touched-red-needs-repeat-proof]] · [[tool-claim-must-name-its-environment]] · [[gate-rig-mechanisms-not-remembered-steps]]