---
name: derived-figure-without-its-formula-dies-with-the-context
description: "A number DERIVED (not emitted) whose formula lives only in the measurer's context becomes unfalsifiable at the next context reset — the figures survive in the record, the bound does not"
metadata: 
  node_type: memory
  type: feedback
  originSessionId: cea05fb0-78f4-496c-bdf5-5c1a3aec65fe
  modified: 2026-08-05T04:05:02.556Z
---

A figure you **derive** — computed from timestamps, diffs, or log arithmetic — carries its bound
only in your head unless the bound is written at the point the number is produced. The **numbers**
get recorded (a peer files them, a register cites them); the **formula** does not. One `/clear`
later the series is unfalsifiable, and nobody can tell because the figures look exactly as
authoritative as the day they were measured.

**Measured, 2026-08-05 (v0.55.0 golden batch).** The twohost "margin series" — six draws, `+60`,
`-0.6` RED, `+59.4`, `+60`, `+59.59`, `+63.89` — was mine. doyle asked for a seventh on the new run.
The extraction refused: grepping `margin|deadline|serve window|slack|headroom|budget` against a
**known-good control run** returned ZERO, and `margin` is absent from `golden.yml` with every
`crates/` hit unrelated (terminal right-margins, `ROUND_DRAIN_MARGIN` in `pump/mod.rs`). So the six
were never emitted — they were derived against a bound **nobody wrote down**. IR-29 attributes the
series to me; doyle recorded the figures and never held the bound; my context had cleared. Between
the two of us the derivation was GONE. **I asked doyle for the provenance of my own measurement.**

**Why:** ⭐⭐ this is the failure mode that survives every other discipline. The series was
positive-controlled, its markers cited by text, its numbers filed in the register — and it still
died, because *every check was applied to the numbers and none to the formula*. The series' whole
value was its **bimodality**; a seventh figure measured against a different bound would have looked
comparable, entered the series, and corrupted exactly that structure invisibly. A wrong number gets
distrusted; a plausible number of unknown provenance gets **promoted**. Same shape as
[[announce-token-from-the-failing-frame-not-the-register]] and
[[a-predicate-without-its-tool-is-not-evidence]] — the artifact is fine, its *derivation* is what
was never captured. Cousin of [[cancelled-measurement-leaves-labelled-hole]] (that hole is labelled
on purpose; this one labels itself only if you notice).

**How to apply:**
- **Prefer an EMITTED token to a derived one.** If a quantity matters enough to track across runs,
  it belongs at an emit site with its two events named there — not in a measurer's arithmetic.
  doyle filed exactly this as the register fix (`@e8805f7`), not as a chat conclusion.
- **When you must derive, write the bound WITH the number, every time** — "node → ladder complete"
  beats "183.50s" and costs six words. A span with named endpoints is re-derivable by a stranger;
  a bare figure is not.
- **Refuse the next draw rather than guess the formula.** Closing a series at N draws with a
  LABELLED HOLE is strictly better than an N+1 that may be a different quantity wearing the same
  name. doyle ratified the refusal as the ruling.
- **Do not report a derived figure under a borrowed label.** Report what you can define. I reported
  a *ladder span with named endpoints* (183.50s batch vs 259.21s control) and explicitly did NOT
  call it "the margin" — different quantity, so a different word.
- **A control that REFUSES is doing its job.** The zero-hit grep against a known-good run is what
  made the missing formula visible at all; without the control arm it would have read as "the log
  changed." See [[make-a-new-rig-red-on-purpose]], [[verdict-from-probe-competence]].