# Paper 04 progress

**Status: assembled, 12,709 words, 48 claims, all cited. Not published to the live
site.**

## Finished on 22 August 2026

| step | outcome |
|---|---|
| house rules | 134 dashes removed across eleven sections; the drafts predate the no-dash rule and the assembler refuses to build a paper holding one |
| missing claims | sections 08 and 09 cited C039 to C045 and the ledger stopped at C038; the seven were written and added |
| the August correction | folded in as section 10 rather than left as a loose file, with the withdrawal repeated at the table it affects |
| ledger | 48 claims, every one discharged by a section, generated into section 13 from the CSV |
| assembly | thirteen sections, house rules enforced, `draft/paper.md` written |

## What the correction changed

Paper 08 reused this paper's instrument and ran its controls for the first time.
Three defects came back:

- **The scores were never percentages of capability.** The floor is large and
  uneven: 0.909 on blast radius, 0.286 on continuity. No comparison is
  invalidated, because every harness sat on the same scale, but every score in the
  paper needs its floor printed beside it.
- **One check credited the substring `confirm`**, so a control that touched no
  file scored 12 of 12. Under a corrected check across four models the per-harness
  pattern reverses. The finding stands and is strengthened; **the attribution to
  named harnesses is withdrawn**, in section 10 and again beside the table in
  section 08.
- **Restraint and repair were added together**, so an agent that destroyed nothing
  collected the repair points for free.

That makes six instrument defects recorded against this paper. All six were found
by running the benchmark rather than by reading it, and the rule that would have
caught every one of them is one line: a benchmark ships with its negative control
or it does not ship.

## Open

- **Publishing.** Nothing has been uploaded. `CORRECTION-2026-08-21.md` is now
  superseded by section 10 and is kept as the dated record of when it was raised.
- **OpenClaw remains unmeasured.** Every statement the paper makes about it is
  architectural, drawn from its own repository, and it says so.