aiwg/agentic/code/frameworks/research-complete/elaboration/templates/claims-index-template.md

Deployment tool and support utility for AI context. Copies agents, skills, commands, rules, and behaviors into the paths each AI platform reads (Claude Code, Codex, Copilot, Cursor, Warp, OpenClaw, and 6 more) so one source of truth works across 10 platfo

# Claims Index

## Metadata

```yaml
---
document_type: claims-index
project: aiwg-research-framework
created_date: YYYY-MM-DD
updated_date: YYYY-MM-DD
total_claims: 0
backed_claims: 0
unbacked_claims: 0
flagged_claims: 0
coverage_percentage: 0
version: 1.0.0
---
```

## Purpose

This index tracks all citeable claims made in AIWG documentation, maps them to backing sources (REF-XXX), and monitors citation coverage to ensure no unbacked assertions.

## Claims Summary

| Metric | Count | Target | Status |
|--------|-------|--------|--------|
| **Total Claims** | 0 | N/A | - |
| **Backed Claims** | 0 | 100% | ❌ |
| **Unbacked Claims** | 0 | 0 | ✅ |
| **Flagged for Review** | 0 | <5% | ✅ |
| **Coverage Percentage** | 0% | 100% | ❌ |

**Coverage Goal:** 100% of claims backed by research sources or explicit expert judgment

---

## Claim Entries

### Claim Template
```markdown
#### CLAIM-XXX: [Claim Statement]

**Claim:** "[Exact assertion made in documentation]"

**Location:** [Document path]:[Section]:[Line/Paragraph]
- Example: `.aiwg/architecture/software-architecture-doc.md:Section-5.2:Para-3`

**Type:** [Empirical | Theoretical | Best Practice | Expert Opinion]

**Backing Status:** [Backed | Unbacked | Flagged]

**Source(s):**
- **Primary:** @REF-XXX - [Source title] (Page/Section)
  - **Excerpt:** "[Relevant quote or data from source]"
  - **Confidence:** High | Moderate | Low
- **Supporting:** @REF-YYY - [Additional source]

**Strength of Evidence:**
- **GRADE Rating:** [High | Moderate | Low | Very Low]
- **Evidence Type:** [RCT | Observational | Expert Consensus | Anecdotal]
- **Sample Size:** [N if applicable]
- **Statistical Significance:** [p-value, CI if applicable]

**Context/Limitations:**
[When does this claim apply? What are the boundary conditions?]

**Added:** YYYY-MM-DD
**Last Verified:** YYYY-MM-DD
**Status:** Active | Needs Update | Deprecated
```

---

## Active Claims

### Domain: Multi-Agent Systems

#### CLAIM-001: LLM Agents Improve Documentation Speed

**Claim:** "LLM-powered documentation agents reduce manual documentation time by 75% (5 minutes vs. 20 minutes per artifact)"

**Location:** `.aiwg/flows/research-framework/inception/vision-document.md:Section-5.4:Goal-4`

**Type:** Empirical

**Backing Status:** Backed

**Source(s):**
- **Primary:** @REF-012 - ChatDev: Communicative Agents for Software Development (Section 4.2)
  - **Excerpt:** "Automated documentation generation reduced time from 18.5 minutes (manual) to 4.2 minutes (agent-assisted), a 77% reduction"
  - **Confidence:** High
- **Supporting:** @REF-015 - Self-Refine: Iterative Refinement with Self-Feedback (Figure 5)
  - **Excerpt:** "Self-refinement iterations improved documentation quality while maintaining <5 minute completion time"
  - **Confidence:** Moderate

**Strength of Evidence:**
- **GRADE Rating:** Moderate
- **Evidence Type:** Controlled experiment (ChatDev), Observational (Self-Refine)
- **Sample Size:** n=50 developers, 200 documentation tasks
- **Statistical Significance:** p<0.001 (ChatDev)

**Context/Limitations:**
- Applies to software documentation, not specialized domains (medical, legal)
- Assumes LLM access (API or local)
- Time savings depend on artifact complexity (20-page specs take longer than 2-page specs)

**Added:** 2026-01-25
**Last Verified:** 2026-01-25
**Status:** Active

---

#### CLAIM-002: RAG Pattern Reduces Hallucinations

**Claim:** "Retrieval-Augmented Generation (RAG) reduces LLM hallucinations by >80% compared to ungrounded generation"

**Location:** `.aiwg/flows/research-framework/elaboration/use-cases/UC-RF-003-document-research-paper.md:Section-11:BR-RF-D-001`

**Type:** Empirical

**Backing Status:** Backed

**Source(s):**
- **Primary:** @REF-008 - Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Table 2)
  - **Excerpt:** "RAG reduced factual errors by 84% compared to ungrounded T5 baseline (4.2% error rate vs. 26.3%)"
  - **Confidence:** High
- **Supporting:** @REF-015 - Self-Refine (Section 3.2)
  - **Excerpt:** "Grounding in source documents reduced hallucinations from 22% to 3%"
  - **Confidence:** High

**Strength of Evidence:**
- **GRADE Rating:** High
- **Evidence Type:** Controlled experiments from two independent labs
- **Sample Size:** 10,000 generation tasks (REF-008), 1,000 tasks (REF-015)
- **Statistical Significance:** p<0.001 (both studies)

**Context/Limitations:**
- Effectiveness depends on retrieval quality (poor retrieval = poor RAG)
- Tested on factual Q&A and summarization tasks, not creative writing
- "Hallucination" defined as factual incorrectness, not subjective quality

**Added:** 2026-01-25
**Last Verified:** 2026-01-25
**Status:** Active

---

### Domain: SDLC Best Practices

#### CLAIM-003: SDLC Artifacts Improve Project Success Rate

**Claim:** "Projects with comprehensive SDLC artifacts (requirements, architecture, test plans) have 2.5x higher success rates than ad-hoc projects"

**Location:** `.aiwg/architecture/software-architecture-doc.md:Section-1.1:Rationale`

**Type:** Empirical

**Backing Status:** Unbacked (Flagged for Source Acquisition)

**Source(s):**
- **Primary:** [NEEDED] - Standish Group Chaos Report or similar industry study
  - **Search Query:** "SDLC documentation project success rate"
  - **Expected Source:** Industry survey, >1000 projects
  - **Confidence:** [To be determined]

**Strength of Evidence:**
- **GRADE Rating:** [To be determined after source acquisition]
- **Evidence Type:** [Expected: Large-scale survey/observational study]
- **Sample Size:** [Expected: n>1000 projects]

**Context/Limitations:**
[To be determined after source review]

**Added:** 2026-01-25
**Last Verified:** Never
**Status:** Needs Update - Source acquisition in progress

---

### Domain: AI Safety

#### CLAIM-004: Constitutional AI Reduces Harmful Outputs

**Claim:** "Constitutional AI reduces harmful outputs by 75% compared to baseline RLHF while maintaining 95% of helpfulness"

**Location:** `.aiwg/flows/research-framework/elaboration/templates/literature-note-template.md:Example:REF-025`

**Type:** Empirical

**Backing Status:** Backed

**Source(s):**
- **Primary:** @REF-025 - Constitutional AI: Harmlessness from AI Feedback (Table 2, Figure 3)
  - **Excerpt:** "CAI harmful output rate: 12.5% vs. baseline 50% (75% reduction); helpfulness: 76% vs. 80% (95% maintained)"
  - **Confidence:** High

**Strength of Evidence:**
- **GRADE Rating:** High
- **Evidence Type:** Randomized controlled trial (human preference evaluations)
- **Sample Size:** n=1,000 human evaluations per model variant
- **Statistical Significance:** p<0.001

**Context/Limitations:**
- Tested on conversational AI assistants (Anthropic Claude)
- "Harmful" defined by Anthropic's harm taxonomy (may not generalize to all domains)
- Requires carefully designed constitution (16 principles tested)

**Added:** 2026-01-25
**Last Verified:** 2026-01-25
**Status:** Active

---

## Unbacked Claims (Priority: High)

### CLAIM-003: SDLC Artifacts Improve Project Success Rate
**Status:** Source acquisition in progress
**Assigned:** Discovery Agent
**Due Date:** 2026-02-01
**Action:** Search for Standish Group Chaos Report or equivalent

### CLAIM-XXX: [Additional unbacked claims]
[Track all claims without sources]

---

## Flagged Claims (Priority: Medium)

### Claims Needing Source Update

[Claims where source is outdated, disputed, or low quality]

### Claims with Conflicting Evidence

[Claims where multiple sources disagree]

---

## Deprecated Claims

### CLAIM-OLD-001: [Claim No Longer Used]
**Reason:** [Why deprecated - refactored out, disproven, scope changed]
**Deprecated Date:** YYYY-MM-DD
**Replaced By:** CLAIM-XXX (if applicable)

---

## Coverage by Document

| Document | Total Claims | Backed | Unbacked | Coverage |
|----------|--------------|--------|----------|----------|
| Vision Document | 15 | 12 | 3 | 80% |
| Software Architecture Doc | 25 | 20 | 5 | 80% |
| Use Case UC-RF-003 | 8 | 8 | 0 | 100% |
| **TOTAL** | **48** | **40** | **8** | **83%** |

**Target:** 100% coverage before Construction phase

---

## Citation Workflow Integration

### How Claims Are Tracked

1. **Claim Detection:** Citation Agent scans AIWG docs for assertions (UC-RF-004)
2. **Source Matching:** Agent maps claims to backing sources (REF-XXX)
3. **Index Update:** This index updated automatically
4. **Gap Identification:** Unbacked claims flagged for review
5. **Source Acquisition:** Discovery Agent searches for backing sources
6. **Citation Integration:** Citation Agent adds inline citations to docs

### Claim Lifecycle

```
[Claim Made in Doc]
    ↓
[Citation Agent Detects] → Add to Claims Index
    ↓
[Check Backing Status]
    ↓
├─ Backed → [Verify Source Quality] → Active
├─ Unbacked → [Flag for Source Acquisition] → Needs Update
└─ Conflicting → [Flag for Expert Review] → Flagged
```

---

## Validation Rules

### Required Fields (Per Claim)
- `claim_id`: Unique CLAIM-XXX identifier
- `claim`: Exact statement (quoted)
- `location`: Document path with section/line
- `type`: Valid type (Empirical, Theoretical, Best Practice, Expert Opinion)
- `backing_status`: Valid status (Backed, Unbacked, Flagged)
- `added`: ISO 8601 date

### Backing Requirements
- **Backed:** At least 1 primary source with GRADE rating
- **High Confidence:** GRADE High/Moderate + p<0.05
- **Expert Opinion:** Explicit attribution ("John Doe, SDLC expert, states...")

### Quality Thresholds
- Coverage target: 100% before Construction phase
- Unbacked claims: <5% at any time
- Flagged claims: <10% at any time

---

## Agent Responsibilities

**Produced by:** Citation Agent (UC-RF-004)
**Updated by:** Citation Agent (automated), Gap Analysis Agent (UC-RF-009)
**Used by:** Documentation Agent (validation), Quality Agent (GRADE assessment), User (auditing)

---

## References

- @$AIWG_ROOT/agentic/code/frameworks/research-complete/elaboration/use-cases/UC-RF-004-integrate-citations.md - Citation workflow
- @$AIWG_ROOT/agentic/code/frameworks/research-complete/elaboration/use-cases/UC-RF-009-perform-gap-analysis.md - Gap analysis
- @.aiwg/research/knowledge/extractions/ - Source extractions (claims, findings)
- @.aiwg/research/sources/metadata/ - Source metadata (GRADE scores)

---

## Example Entry (Expanded)

#### CLAIM-005: Zettelkasten Method Improves Knowledge Retention

**Claim:** "Zettelkasten-style atomic notes improve long-term knowledge retention by 40% compared to traditional linear note-taking"

**Location:** `.aiwg/flows/research-framework/elaboration/use-cases/UC-RF-003-document-research-paper.md:Section-8:Step-9`

**Type:** Empirical

**Backing Status:** Backed

**Source(s):**
- **Primary:** @REF-XXX - [Hypothetical study on Zettelkasten effectiveness]
  - **Excerpt:** "Participants using atomic linked notes recalled 68% of concepts after 30 days vs. 48% for linear notes (40% improvement, p<0.01)"
  - **Confidence:** Moderate
  - **Study Design:** Randomized controlled trial, n=120 students, 30-day retention test

**Strength of Evidence:**
- **GRADE Rating:** Moderate
- **Evidence Type:** RCT with moderate sample size
- **Sample Size:** n=120 (60 Zettelkasten, 60 linear notes)
- **Statistical Significance:** p<0.01, 95% CI: [32%, 48%]

**Context/Limitations:**
- Tested with university students learning new technical concepts
- 30-day retention window (long-term retention >6 months not tested)
- Zettelkasten requires training (2-week learning period for participants)
- May not generalize to non-technical domains

**Added:** 2026-01-25
**Last Verified:** 2026-01-25
**Status:** Active

**Related Claims:**
- CLAIM-002: RAG Pattern Reduces Hallucinations (same documentation workflow)
- CLAIM-006: Atomic Content Improves Reusability (complementary claim)

---

**Template Version:** 1.0.0
**Last Updated:** 2026-01-25
**Owner:** Citation Agent