RDLC Configuration

June 9, 2026 · View on GitHub

This document is installed by claude-rdlc-wizard into the consumer repo. It is the canonical for "what RDLC enforces here." Edit it freely — update-wizard preserves your customizations and only touches the metadata header.

Wizard Version Tracking

Property	Value
Wizard Version	0.3.2
Last Updated	2026-05-06
Claude Code Baseline	v2.1.154+ (required for `claude-opus-4-6[1m]`)
Recommended Model	`claude-opus-4-6[1m]` (Opus 4.6 max) for primary work, `gpt-5.5 xhigh` (Codex) for cross-model review
Recommended Effort	`max` for research drafting, `xhigh` for review and verification

RDLC Enforcement

This repository uses the RDLC Wizard to enforce:

1. Source Discipline

Every claim has a citation or URL
Sources ranked by tier (see "Source Hierarchy" below)
Source-at-first-mention; subsequent references inherit

2. Confidence Classification

Every claim labeled VERIFIED / SUPPORTED / INFERRED / UNVERIFIED
Labels are not optional; the hook blocks Write/Edit on research files without one
Confidence escalates only when new sources land — never silently

3. Cross-Model Review

Claude (primary author) drafts; a second model (Codex GPT-5.4 xhigh recommended) reviews
Mission-first prompt structure: mission, success, failure, audience, stakes
Convergence: 2 rounds is the sweet spot, 3 max

4. AI Slop Gate

Banned phrase list scanned via grep -E before commit
Hard fail (rewrite immediately) and soft warn (check context) tiers
Zero hits = pass

5. Audience-Firewall

Multi-deliverable repos enforce per-audience content boundaries
Private content (interview salary, internal mock questions) cannot leak to public deliverables
Enforced in the deliverable generator, not just the source pages

6. Fact Regression Tests

Every defect found becomes a permanent regression check
Test format: check_present and check_absent bash assertions on output files
The ratchet only turns forward — gates get stricter, never looser

Hooks Installed

Hook	Trigger	Purpose
`rdlc-prompt-check.sh`	Every prompt	RDLC baseline reminder
`rdlc-instructions-check.sh`	Session start	Validates RDLC.md exists; prompts setup if missing
`slop-scan-pretool.sh`	Before Write/Edit	Blocks AI slop additions
`confidence-required.sh`	Before Write/Edit on `research/`, `evidence/`	Requires confidence label on new claims
`source-required.sh`	Before Write/Edit on research files	Requires source-at-first-mention
`audience-firewall.sh`	Before Write to `output/`	Blocks audience-private content from public deliverables

Skills Available

Skill	Invocation	Purpose
RDLC	`/rdlc`	Full research workflow guidance
Setup	`/setup-rdlc`	Confidence-driven project setup
Update	`/update-rdlc`	Smart update with drift detection
Feedback	`/feedback-rdlc`	Privacy-first contribution loop

Source Hierarchy

Higher tiers carry more weight when claims conflict.

Tier	Source Type	Weight
1	Primary databases (DrugBank, ChEMBL, PubChem, openFDA, EDGAR, USPTO, FEC)	Highest — authoritative
2	Government records, legislative archives, court filings	Highest — authoritative
3	Peer-reviewed papers (PubMed, arXiv with peer review)	High
4	Established news outlets with editorial standards	High
5	Official organizational sites (statesproject.org, anthropic.com)	High
6	Professional profiles (LinkedIn, TheOrg, Crunchbase)	Moderate
7	Watchdog/advocacy sites (note bias)	Moderate
8	Podcasts/interviews with primary subjects	Moderate (primary source)
9	Wikipedia, secondary aggregators	Low — verify upstream
10	Social media, anonymous posts, screenshots	Lowest — attribute, do not assert

Confidence Vocabulary

Label	Meaning	Required Evidence
VERIFIED	Multiple independent sources confirm	≥2 tier 1-5 sources OR 1 tier 1-2 with cross-check
SUPPORTED	Single reliable source, no contradictions	1 tier 1-5 source
INFERRED	Logical deduction from verified premises	Reasoning chain documented; flag in text
UNVERIFIED	Source exists but not yet checked, or conflicting sources	Note "needs confirmation" in text
GAP (alt)	No source located, structurally unknowable (use for investigation domains)	Note "no source located" in text

Pick one alt for the fourth slot per project: UNVERIFIED if sources should exist, GAP if some facts cannot be recovered. Document the choice here.

AI Slop Gate

Every content change must pass a slop scan before commit.

Hard fail (rewrite immediately): deep dive, game-changer, cutting-edge, elevate, delve, tapestry, holistic, robust, paradigm, groundbreaking, streamline, empower, harness, unleash, unpack, pivotal, crucial, bigger picture, smoking gun, synergy, arguably, it's worth noting, at the end of the day, in today's world, needless to say

Soft warn (check context — legitimate use OK): leverage (as verb meaning "use"), comprehensive (as filler), landscape (filler vs "competitive landscape"), stakeholder (when generic), navigate (meaning "deal with"), facilitate (filler), best practices (generic), fundamentally, inherently

How to run:

bash scripts/slop_scan.sh

Zero hits = pass. Any hit = review and rewrite.

Allowlist: project-specific proper nouns (mission pillar names, direct quotes) can be added to .rdlc/slop-allowlist.txt to suppress false positives. Lines in that file are excluded from the scan.

Audience Firewall

Multi-deliverable repos must enforce per-audience boundaries. Each deliverable has its own assertion list of forbidden patterns.

Example mapping:

Deliverable	Audience	Forbidden Content
`output/private_*.html`	Internal	(none — full content)
`output/public_*.html`	External	Salary figures, mock-interview content, private contact info, internal review notes
`output/clinical_*.pdf`	Medical	Litigation language, manufacturer marketing claims

The check belongs in the generator (Python/script that builds the deliverable), not just the source markdown. PDFs and HTML get forwarded without their context — the boundary must hold at render time.

Cross-Model Review Protocol

For high-stakes deliverables (anything published, anything legally adjacent, anything tone-sensitive):

Pre-flight self-review. /code-review equivalent for prose. Document every check at .reviews/preflight-{review_id}.md. Don't waste a Codex round on issues you could catch yourself.
Mission-first handoff. Write .reviews/handoff.json with: mission (2-3 sentences), success criteria, failure criteria, audience profile, stakes, files_changed, verification_checklist (file:line specific).
Run the reviewer. Codex CLI at model_reasoning_effort=xhigh. Always xhigh — lower settings miss subtle errors.
Dialogue loop. Per-finding response, increment round, status PENDING_RECHECK. Convergence: 2 rounds is the sweet spot, 3 max. After 3 still NOT CERTIFIED → escalate.
Audience-as-reviewer variant. For tone-sensitive content, prompt the reviewer to adopt the audience's mindset: "If you were [audience], when do you stop reading?" — flushes out check-out moments document-level review never finds.

Full protocol: ~/xdlc/docs/cross-model-review.md.

Fact Regression Tests

Every defect becomes a permanent check. Test format:

check_present "Claim X exists" "expected pattern" "$DOC"
check_absent "No fabrication Y" "forbidden pattern" "$DOC"
check_absent_excl_review "No overclaim Z outside review section" "pattern" "$DOC"

Run: bash scripts/regression_test.sh

The test count grows monotonically. Test counts in docs go stale — verify dynamically or guard with a count assertion.

Confidence Levels (For Tasks)

Level	Meaning	Action
HIGH (90%+)	Know exactly what to do	Proceed after approval
MEDIUM (60-89%)	Solid approach, some unknowns	Highlight uncertainties
LOW (<60%)	Not sure	ASK before proceeding
FAILED 2x	Something's wrong	STOP and ask for help

Configuration Files

.claude/
├── settings.json
├── hooks/
│   ├── _find-rdlc-root.sh
│   ├── rdlc-prompt-check.sh
│   ├── rdlc-instructions-check.sh
│   ├── slop-scan-pretool.sh
│   ├── confidence-required.sh
│   ├── source-required.sh
│   └── audience-firewall.sh
└── skills/
    ├── rdlc/SKILL.md
    ├── setup/SKILL.md
    ├── update/SKILL.md
    └── feedback/SKILL.md

scripts/
├── regression_test.sh        # Fact regression suite
├── slop_scan.sh              # Banned-phrase grep
└── generate_deliverable.py   # Multi-document HTML generator (audience-firewalled)

research/                     # Topic-specific research files (lowercase_with_underscores.md)
evidence/
├── sources/                  # Raw source material with URLs
└── ...                       # Domain-specific subdirs
output/                       # Final deliverables (HTML, PDF)
.reviews/                     # Review artifacts (handoff.json, preflight-*.md)
.rdlc/
└── slop-allowlist.txt        # Project-specific exclusions from slop scan

Lessons Learned

(Empty at install. Populate as the project earns rules. Format mirrors ~/sdlc-wizard/SDLC.md "Lessons Learned" — one bullet per gotcha, cite the originating PR or incident date.)