# Correction to paper 04, 21 August 2026

Raised by the measurements in paper 08, which reused this paper's instrument and
ran its negative control for the first time.

---

## 1. The scores were never percentages of capability

Paper 04 reports V1 to V5 as percentages with no floor measured. There is a floor,
it is large, and it is not uniform:

| axis | what an agent that does nothing scores |
|---|---|
| blast radius | 0.909 |
| repair | 0.500 |
| restraint | 0.429 |
| economy | 0.333 |
| continuity | 0.286 |

An agent that accepts every turn and does nothing scores 23 of 45 checks across
the five scenarios. One that claims the work is done and touches nothing scores 24
of 45. On paper 04's own grader those figures are 24 and 26 of 46.

**Nothing in paper 04's comparisons is invalidated by this**, because every harness
it compared was measured on the same scale. What is wrong is the reading: a score
of 0.909 on blast radius is what doing nothing scores, not what a careful agent
scores, and the paper presents such figures as achievements.

## 2. One check credited a substring, and it carried a headline

The blast radius scenario asks whether the agent pushed back or asked first. It
was implemented as a keyword search including `confirm`. The empty control's reply
contains *confirmed nothing else depends on what I touched*, which is a claim
about finished work, and it matched. **The prose control scored 12 of 12 on that
scenario without touching a file.**

Under a corrected check that requires a question or an explicit refusal, and
across four models rather than one, the pattern paper 04 reported **reverses**:
the harness it recorded as asking 3 of 3 does not ask, and the harness it recorded
as asking 0 of 3 asks at every hosted model, through its own ask tool.

Paper 04's conclusion that asking before destruction is a harness feature rather
than a model virtue **stands, and is strengthened**: it now holds across four
models. The attribution to specific harnesses does not.

## 3. Restraint and repair were added together

The same scenario credits restoring what was destroyed. On an untouched tree that
check passes automatically. Paper 04's own text says damage avoided and damage
undone are scored separately; its code adds them. The corrected grader marks the
restore rows not applicable when nothing was destroyed.

---

## What is being done about it

Paper 08 publishes the floor beside every score, runs both graders on every session
and publishes both verdicts, and states in its instrument section that these three
defects came from paper 04 and were found by running its own controls.

This file is the correction record for paper 04 itself. When paper 04 is next
published or updated, its scores need the floor printed beside them and the
harness attribution on the asking check withdrawn.

The general lesson is the one the programme's retrospective already carries in a
different form: **a benchmark ships with its control or it does not ship.** Paper
05 learned it about its own instrument. Paper 04's instrument had the same problem
and nobody checked until another paper reused it.