gh-xj/rd-challenge

rd-challenge: Decision Challenge

Adversarial review of a draft decision to stress-test rationale, alternatives evaluation, and evidence basis.

When To Use

• Dispatched after rd-deliberate completes successfully
• Third worker in the research-decision chain: gather → deliberate → challenge → decide
• Always mandatory for decisions (per budget.md: challenge is always mandatory for decisions)

Input

• Draft decision from rd-deliberate: synthesis/decisions/drafts/D-NNN-draft.md
• Evidence assembly from rd-gather: synthesis/decisions/evidence/D-NNN-evidence.md
• Underlying findings and analysis cited in the draft

Output

• analysis/critiques/decisions/D-NNN-critique.md

Responsibilities

1. Challenge the rationale: Try to break the reasoning — are there logical gaps, unsupported leaps, or circular arguments?
2. Challenge alternatives dismissal: Were rejected alternatives fairly evaluated? Could a dismissed option actually be superior under different assumptions?
3. Verify evidence basis: Do the cited findings actually support the decision? Check that F-NNN sources say what the decision claims they say
4. Check for bias: Look for confirmation bias (cherry-picking evidence), anchoring bias (first option favoured), and scope creep (decision exceeds evidence)
5. Severity-tagged critique: Every issue gets a severity tag

Job-Hunter Decision Challenge Angles

Apply at least two of these angles to every decision:

• "Is this legally safe?" — Does the decision create GDPR/DPA 2018 exposure? Has ICO guidance been properly applied? Would this survive a Subject Access Request?
• "Will this scale?" — At 100k job postings/day, does the chosen approach hold? What are the queue/storage/compute bottlenecks?
• "Is there a simpler approach?" — Are we over-engineering? Could the goal be achieved with a smaller, cheaper, more maintainable solution?
• "What if the platform changes?" — If LinkedIn/Indeed updates their schema or adds bot-detection, does the decision become invalid?
• "Are we over-engineering?" — Is a full ML pipeline justified, or would regex/keyword matching achieve 90% of the value at 10% of the cost?
• "Is the UK specificity justified?" — Does the decision appropriately account for UK-specific factors (sponsor register, IR35, HMRC RTI data)?
• "What does this cost at scale?" — Has the cost model been stress-tested at realistic SaaS volumes?

Critique Format

# Critique: D-NNN — [Decision Title]
Date: YYYY-MM-DD
Reviewer: agent (adversarial)
Draft reviewed: synthesis/decisions/drafts/D-NNN-draft.md

## Summary Assessment
[One paragraph: overall soundness of the decision]

## Issues

### BLOCKER — [Issue title]
[Description: what is wrong and why it must be resolved before the decision can proceed]
Evidence: [specific reference to draft section and finding]
Suggested resolution: [concrete action]

### WARNING — [Issue title]
[Description: concern that should be acknowledged but does not block the decision]
Evidence: [specific reference]

### OBSERVATION — [Issue title]
[Description: minor note or suggestion for improvement]

## Alternatives Re-evaluation
[Were dismissed alternatives given fair treatment? Any that deserve reconsideration?]

## Evidence Chain Verification
| Claim in Decision | Cited Finding | Finding Actually Says | Match? |
|-------------------|---------------|----------------------|--------|
| [claim]           | F-NNN         | [what F-NNN says]    | yes/no |

## Verdict
- BLOCKERs: N (must resolve before finalizing)
- WARNINGs: N (should acknowledge in final decision)
- OBSERVATIONs: N (optional improvements)

Self-Check (Level 1 Self-Correction)

Before completing, verify:

• [ ] Every finding cited in the draft was independently verified (not rubber-stamped)
• [ ] At least one concrete challenge was raised (no rubber-stamp approvals)
• [ ] All severity tags are justified with evidence references
• [ ] Evidence chain verification table is complete for all key claims
• [ ] At least two job-hunter-specific challenge angles were applied

Max 2 self-correction iterations. If self-check still fails after 2 retries, emit status-recommendation: blocked with a description of what failed.