---
name: precise-claim-is-falsifiable-vague-is-useless
description: "Doctrine (doyle 2026-07-30, ranked above the night's other lessons): a precisely-stated claim is falsifiable by its own author later; a vague one is permanently safe and permanently useless"
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 190bd6db-cddf-4d69-b960-51e892a44257
  modified: 2026-07-30T02:00:10.632Z
---

**A precisely-stated claim is falsifiable by its own author later. A vague one is permanently safe — and permanently useless.** doyle ruled this the lesson of the whole milestone-A guard thread, explicitly ranking it above the conclusions the thread actually reached.

The case that produced it: todlando supplied the **source verification** that upgraded a claim about the worker guard's diagnostics ("all four child outcomes emitted before the destructive reaper, L244 before L246"), and then supplied the **measurement that downgraded it** — and the second was findable ONLY because the first was stated precisely enough to be tested against a log. Compare the hedge that would have been "permanently safe": *"the guard improves diagnostics."* Nothing can check that.

**Why:** every correction that night landed on a specific claim, and **not one landed on a hedge** — not because the hedges were right, but because they could not be reached. So the instinct to soften a claim to avoid being wrong is the instinct to make it unfalsifiable, **which is worse than being wrong and then corrected.**

⭐⭐ **THE DECISION-PROCEDURE FORM, same day: WHEN YOUR PRE-COMMITTED BRANCHES ALL MISS REALITY, NAME THE GAP — DO NOT STRETCH A BRANCH TO COVER IT.** doyle pre-committed three arms for golden r3's last leg (success before a derived ~01:41:08Z mark / failure at the mark with a timeout diagnostic / failure otherwise). I measured the live state at 01:42:35Z — step 10 still `in_progress`, past the mark — which **none** of the three named, and said so. He opened a FOURTH branch ("still running, past the mark") that explicitly licenses nothing: not failure (the mark was DERIVED from the *other* box's `WAIT_SECS: 900`, and its non-transfer was the stated caveat, so b outliving it is the caveat working), not health ("b is still slow" discriminates the healthy mirror from a rendezvous timeout exactly as poorly at 21 minutes as at 13). The only new information was **a borrowed number falsified**. ⚖ Why stretching is the trap: a state nobody planned for gets handled by the arm it least resembles, and **the stretch is invisible afterwards because the branch text still reads as if it applied** — the plan becomes unfalsifiable after the fact, which is this entry's defect in decision-procedure clothing. Duration is a non-discriminating observation and more of it is still none.

⚖ **The companion failure — the same discipline SKIPPED on claims that feel administrative — is banked separately in [[audit-the-boring-claims]] (four fabrications in one night: a filename, a draft's existence, a delivery mapping, a subordinate clause), and the channel two of them arrived through is [[commune-is-testimony-not-an-artifact]].** Read those two with this one; the operational trigger they share is: say "confirmed"/"verified"/"delivered" ⇒ name the command and the sha in the next sentence, or withdraw the claim rather than defend it.

**How to apply:** state claims with the file, line, sha, count, or ordering that makes them checkable, and accept that you are the most likely person to falsify them later — that is the claim working, not the claim failing. When you catch yourself reaching for a softer wording, ask whether the softening removes the handle a peer would grab to test it; if so, keep the sharp version and let it be corrected. Kin: [[make-a-new-rig-red-on-purpose]] (an instrument seen to red), [[absence-needs-sibling-probe]] (a peer's READING is a claim, audit by re-measuring), [[name-the-file-and-sha-a-condition-came-from]] (a shared noun is not a shared claim), [[verdict-from-probe-competence]].