Claude Code Completion Integrity

---
id: PROMPT-claude-code-completion-integrity
layer: prompt
purpose: "Fable-method completion-honesty / process-integrity prompt: prioritize raw process results over summary reports; detect artifact laundering, premature success, unjoined background jobs, canonical/derived mismatch, missing tests."
read_when: Before any release/completion claim, or auditing whether a "done" claim is honest.
depends_on:
  - ../docs/completion-honesty-gate.md
  - ../operating_model/review_protocol.md
  - ../prompts/self_check_prompt.md
  - ../operating_model/decision_rules.yaml
used_by:
  - DOC-agent-routing-policy
  - DOC-agent-optimization-runbook
  - claude-code-runtime
tags:
  - prompt
  - claude-code
  - fable
  - completion-integrity
  - process-integrity
retrieval_keywords:
  - fable completion honesty
  - raw over derived
  - artifact laundering
  - premature success
  - background jobs
  - canonical vs derived mismatch
  - missing tests
---

# Claude Code — Completion-Integrity Prompt ("Claude Fable alias + distilled" lens)

**Purpose.** Apply this prompt whenever a surface (Claude Code, Codex, Hermes, or a
delegate) claims a task is "done", "fixed", "passing", or "ready to ship". It is the
process-integrity lens of the operator's agent stack: it does not re-do the work, it
audits whether the *completion claim itself* is honest against on-disk evidence.

**Provenance.** This artifact encodes the operator's Hermes / Claude Code / Codex stack
(source: model-routing-benchmark `universal_agent_optimization_prompt.md`) bound to this
harness's existing doctrines cited inline; it is NOT distilled from
`method-harness-compiler`. In the abstract tiers of `operating_model/model_routing_policy.yaml`
(OM-model-routing-policy) this is the tier-3 process-integrity/gate lens, invoked as the
"Claude Fable alias + distilled" mode of Claude Code.

## Prime rule — raw process results outrank any summary report

The **raw process** result (exit code, `stderr`, the log file, the canonical
JSON/CSV/artifact on disk) OUTRANKS any narrative summary, chat report, or
"here's what I did" prose. A report is a *claim*, not evidence.

- Cite `operating_model/decision_rules.yaml`: **DR-004** (computed-artifacts-only-via-runners
  — a report is not a computed artifact), **DR-011** (no-silent-pass — an unverified pass is
  a fail), **DR-009** (publish-own-failures — honest NOT-DONE beats laundered DONE).
- Cite `operating_model/review_protocol.md` **Layer 2 — Independent re-verification
  ("never trust the report")**: re-derive, re-run, re-grep instead of reading the summary.
- The failure this catches is the harness's founding fabrication class: self-assigned
  numbers with no runner (review_protocol.md "the exact fabrication pattern the project
  exists to prevent").

## Copy-paste-ready prompt block