Pipeline Auditor

Pipeline Auditor (draft audit + regression)

Purpose: a deterministic “regression test” for the writing stage.

It answers:

• did we leak placeholders or planner talk?
• did citation scope drift?
• did the draft fall back to generator voice (navigation/narration templates)?
• is citation density/health sufficient for a survey-like draft?

This skill is analysis-only. It does not edit content. For all survey-family profiles, style/citation-shape violations are blocking by default.

Inputs

• output/DRAFT.md
• outline/outline.yml
• Optional (recommended):
• outline/evidence_bindings.jsonl
• citations/ref.bib

Outputs

• output/AUDIT_REPORT.md
• output/TEMPLATE_RESIDUE_SCORECARD.json

What it checks (deterministic)

A150++ citation targets (used by the auditor):

• Per-H3: >=12 unique citations (deep: >=14).
• Global: >=150 unique citations across the full draft (recommended target: 165; deep floor: 165).

Course-paper targets:

• Per-H3: >=4 unique citations and >=1600 non-citation characters.
• Global: >=24 unique citations (recommended target: 32).

• Placeholder leakage: ellipsis (..., …), TODO markers, scaffold tags.
• Outline alignment: section/subsection order vs outline/outline.yml.
• Survey tables: require >=1 Markdown table for course_paper, >=2 for survey/deep (inserted by section-merger from outline/tables_appendix.md; index tables remain internal).
• Paper voice anti-patterns:
• narration templates (This subsection ..., In this subsection ...)
• slide navigation (Next, we move ..., We now turn to ...)
• unmistakable pipeline voice (this run, this workspace)
• ambiguous terms such as this pipeline, this stage, and quality gate

block only when the same sentence contains a Harness anchor such as a

checkpoint, Unit ID, Harness lock, attempt ledger, or template residue;

ordinary subject-matter uses remain non-blocking warnings

• Evidence-policy disclaimer spam: repeated “abstract-only/title-only/provisional” boilerplate inside H3 bodies.
• Meta survey-guidance phrasing: survey synthesis/comparisons should ....
• Synthesis stem repetition: repeated Taken together, ... and similar high-signal generator stems.
• Numeric claim context: numbers without minimal evaluation context tokens (benchmark/dataset/metric/budget/cost).
• Citation health (if citations/ref.bib exists): undefined keys, duplicates, basic formatting red flags.
• Citation-shape hard gate: no adjacent citation blocks ([@a] [@b]) and no duplicate keys inside one block ([@a; @a]). Mid-sentence citation ratio is >=20% for course_paper and >=30% for survey/deep.
• Citation scope (if outline/evidence_bindings.jsonl exists): citations used per H3 should stay within the bound evidence set.
• Whole-draft deterministic template residue: split English and CJK reader-facing prose, compare each sentence with fixed fragments from the template assets selected for this Run, and reject above the Workflow limit.
• Template source provenance: require output/FRONT_MATTER_CONTEXT.json to record the selected front-matter assets and hashes, then require the three template-owning Skill implementations to match .harness/harness.lock.json; missing provenance, legacy locks, and repository drift block acceptance.

The JSON scorecard records the measured ratio, counts, threshold, selected asset

hashes, heading-aware examples, and implementation-lock result. During normal

Harness execution, Completion projects its verdict and dimensions into

.harness/evaluations/ledger.jsonl, including failed Attempts, so Run Audit can

expose the latest measurement instead of reducing it to PASS/FAIL. The scorecard

file remains the complete evidence object; the ledger is intentionally smaller.

The current 10% limit is an initial policy target. The published Survey replay

completes the current 31-check contract at 0/226 residue, establishing

attainability for one retained Artifact set. Clean from-scratch execution,

unrelated topics, and cross-profile calibration remain open.

How to use the report (routing table)

Treat output/AUDIT_REPORT.md as a “what to fix next” router.

Common FAIL families -> responsible stage/skill:

• Placeholders / leaked scaffolds
• Fix: C2–C4 artifacts are not clean. Route to subsection-briefs / evidence-draft / writer-context-pack, then rewrite affected sections.

• Missing overview tables (below the selected profile's table minimum)
• Fix: ensure table-schema + appendix-table-writer produced outline/tables_appendix.md (>=1 course-paper table; >=2 survey/deep tables; citation-backed, no placeholders), then rerun section-merger.

• Planner talk in transitions / narrator bridges
• Fix: rerun transition-weaver (and ensure briefs include bridge_terms / contrast_hook), then re-merge.

• Narration templates / slide navigation inside H3
• Fix: rewrite the failing sections/S*.md via writer-selfloop (local, section-level) or subsection-polisher.

• Evidence-policy disclaimer spam
• Fix: keep evidence policy once in Intro/Related Work (front matter), delete repeats in H3 (use draft-polisher or local section rewrites).

• Citation scope drift (out-of-scope bibkeys)
• Fix: either (a) rewrite the subsection to stay in-scope, or (b) fix mapping/bindings (section-mapper → evidence-binder) and regenerate packs.

• Global unique citations too low
• Fix: citation-diversifier → citation-injector (NO NEW FACTS), then draft-polisher.

• Intro/Related Work too thin / too few cites
• Fix: rewrite the corresponding sections/S<sec_id>.md front-matter file via writer-selfloop (front-matter path) using dense positioning + method paragraph.

• Whole-draft template residue above the Workflow limit
• Fix: use each scorecard example's heading, section kind, section owner, and template owner to locate the responsible front matter, chapter lead, or H3 region; rewrite its source section, regenerate the merged draft, and rerun the auditor. Do not weaken or rename template assets to make an existing Run pass.

Prevention guidance (what upstream writers should do)

If you want the auditor to PASS *without* a heavy polish loop:

• Start each H3 with a content claim + thesis (avoid narration templates).
• Use explicit contrasts and at least one evaluation anchor paragraph.
• Embed citations per claim (avoid trailing cite dumps).
• Put evidence-policy limitations once in the front matter, not in every H3.

Script

Quick Start

• uv run python .codex/skills/pipeline-auditor/scripts/run.py --help
• uv run python .codex/skills/pipeline-auditor/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir>
• --unit-id <U###> (optional; for logs)
• --inputs <semicolon-separated> (rare override; prefer defaults)
• --outputs <semicolon-separated> (rare override; defaults write output/AUDIT_REPORT.md and output/TEMPLATE_RESIDUE_SCORECARD.json)
• --checkpoint <C#> (optional)

Examples

• Run audit after global-reviewer and before LaTeX/PDF:
• uv run python .codex/skills/pipeline-auditor/scripts/run.py --workspace <workspace>

Troubleshooting

Issue: audit fails due to undefined citations

Fix:

• Regenerate citations with citation-verifier and ensure citations/ref.bib contains every cited key.

Issue: audit fails due to narration-style navigation phrases

Fix:

• Rewrite as argument bridges (content-bearing handoffs, no navigation commentary) in the failing sections/* files, then re-merge.

Issue: audit fails due to "unique citations too low"

Fix:

• Run citation-diversifier to produce output/CITATION_BUDGET_REPORT.md.
• Apply it via citation-injector (edits output/DRAFT.md, writes output/CITATION_INJECTION_REPORT.md).
• Then run draft-polisher → global-reviewer → auditor.