dr-report

You are synthesizing research findings into the final deliverable.

Accepts optional arguments: /dr-report, /dr-report --provenance.

Steps

1. Load Context

Read .research/PROJECT.md for:

• The original research question / mission
• The decision context (what the user will do with this)
• The desired artifact type (strategic brief, landscape map, scorecard, etc.)

2. Check Completeness

Read .research/ROADMAP.md. For each phase:

• If COMPLETE, proceed
• If IN_PROGRESS or NOT_STARTED, warn the user: "Phase {N} is not complete. The report will have gaps. Continue anyway? Or run /dr-run first."

3. Load All Data

Read every file in data/. For each:

• Parse the entries
• Note total count, completeness rate, CONFLICTING/UNVERIFIED values
• Group by category or phase as appropriate

Merge adversary sidecar files: Also read any data/*.adversary.json sidecar files. Merge adversary verdicts into the data entries by matching task_id and field_name to the provenance envelopes.

• Error handling: If a sidecar file contains malformed JSON, skip it, log "[WARNING] Skipped malformed sidecar: {filename}", and continue processing other files. Do not crash.
• Update adversary_verdict, adversary_evidence, confidence_delta, verified_at, and survival_score in each matching provenance envelope.

4. Generate Report

Write output/final-report.md with this structure:

# {Research Title}

Generated: {ISO date}
Source: deepresearch autonomous research project

## Executive Summary

{2-3 paragraphs directly answering the research question from PROJECT.md.
This is the most important section. Be specific and actionable.
Reference data points and counts.}

## Methodology

- Research approach: {from ARCHITECTURE.md ADRs}
- Total entities researched: {count}
- Data sources: {count unique source URLs}
- Phases completed: {N of M}
- Overall data completeness: {percentage}

## Key Findings

### {Phase 1 Title}

{Findings from Phase 1 data. Include specific numbers, top entries, patterns.
Reference the data files for details.}

### {Phase 2 Title}

{Findings from Phase 2 data.}

{...repeat for each phase...}

**Adversary highlights in findings:** When presenting claims in Key Findings sections:
- Claims with `adversary_verdict: "refuted"` → mark with `[REFUTED]` and strikethrough: `~~claim text~~ [REFUTED]`
- Claims with `adversary_verdict: "weakened"` → mark with `[WEAKENED]` and show adversary counter-evidence inline: `claim text [WEAKENED — adversary found: "counter-evidence quote"]`
- Claims with `adversary_verdict: "confirmed"` → present normally (no special marking)
- Claims with `adversary_verdict: "pending"` or `"unverifiable"` → present normally

At the end of each Key Findings section, add a **Contested Claims** subsection listing all refuted/weakened claims with their adversary evidence:

```markdown
#### Contested Claims

| Claim | Verdict | Adversary Evidence | Survival Score |
|-------|---------|-------------------|----------------|
| Artisan AI raised $25M | REFUTED | "Crunchbase shows $12M total funding" | 0.00 |
| Regie.ai has 80 employees | WEAKENED | "LinkedIn shows ~120 employees" | 0.42 |

If no adversary sidecar files exist, skip all adversary highlights.

Patterns and Insights

{Cross-cutting observations that span multiple phases. What trends emerged? What was surprising? What does the data suggest that wasn't in the original question?}

Gaps and Limitations

• {N} entries marked INCOMPLETE — could not find full data
• {N} entries marked INACCESSIBLE — sources were unavailable
• Fields with high NOT_FOUND rates: {list}
• {N} CONFLICTING values unresolved
• {Any phases not completed}

Recommendations

{Based on the decision context from PROJECT.md. What should the user do with this information? Prioritized, specific, actionable.}

Appendix

Data Sources

{Count of unique URLs. Top domains referenced.}

Completeness by Field

Research Quality

• Eval pass rate: {from latest eval-history file, if exists}
• Researcher iterations: {count of versions in eval-history}

### 5. Adapt to Artifact Type

Based on the done condition from PROJECT.md:
- **Strategic brief:** Focus on executive summary and recommendations
- **Landscape map:** Include a category breakdown table with key metrics per entity
- **Opportunity scorecard:** Add a scoring matrix with weighted criteria
- **Spreadsheet:** Additionally generate `output/data-export.json` with all entries merged and normalized

### 6. Provenance Export (--provenance flag)

If the user passed `--provenance`:

Write `output/provenance-chain.json` alongside the markdown report. This is the full epistemic audit trail.

Structure:
```json
{
  "generated_at": "2026-04-10T14:30:00Z",
  "trust_score": 0.74,
  "survival_rate": 0.87,
  "claims": {
    "P1.01:funding_total": {
      "claim_text": "Artisan AI raised $12M",
      "source_url": "https://crunchbase.com/organization/artisan-ai",
      "extraction_method": "direct_quote",
      "cross_ref_count": 2,
      "cross_ref_urls": ["https://techcrunch.com/artisan-ai-funding"],
      "confidence_score": 0.9,
      "volatility_class": "medium",
      "adversary_verdict": "confirmed",
      "adversary_evidence": null,
      "verified_at": "2026-04-10T14:32:00Z",
      "survival_score": 0.9
    }
  }
}

• Key format: {task_id}:{field_name} for easy lookup
• Flatten all provenance envelopes from all data files + sidecar verdicts into one exportable artifact
• trust_score: mean survival_score across all adversary-reviewed claims
• survival_rate: (confirmed + weakened) / (confirmed + weakened + refuted)
• If no adversary data exists, set trust_score and survival_rate to null

7. Print Summary

After writing the report:

Report written to output/final-report.md

{word count} words, {section count} sections
{entity count} entities across {category count} categories
{source count} unique sources referenced
Trust: {survival_rate}% survived | Score: {trust_score} (or "N/A" if no adversary data)

The report directly answers: "{original research question}"

If --provenance was passed, also print: Provenance chain written to output/provenance-chain.json ({claim count} claims)

Error Handling

• If no data files exist, tell the user to run /dr-run first.
• If PROJECT.md is missing, tell the user to run /dr-new first.
• If data is sparse (<10 entries total), warn that findings may not be representative.