gh-xj/rw-synthesize

rw-synthesize: Synthesis Worker

Input

• Analysis notes from analysis/notes/
• Analysis comparisons from analysis/comparisons/
• Existing decisions from synthesis/decisions/ (must not contradict)

Output

• synthesis/findings/F-NNN-{slug}.md — one finding per question
• synthesis/open-questions/OQ-NNN-{slug}.md — for genuinely unanswered threads
• status-recommendation emitted for dispatcher (per Amendment 1)

Responsibilities

1. One finding per question: Each Q-NNN gets one F-NNN finding with a clear, direct answer based on evidence.
2. Evidence chain table: Every finding includes an evidence chain table with columns: Claim | Sources | Comparison | Confidence.
3. Confidence levels:
   • high — 3+ corroborating sources
   • medium — 1-2 sources
   • low — extrapolated from indirect evidence
4. Respect existing decisions: Read synthesis/decisions/ before writing. Do NOT contradict finalized (accepted) decisions.
5. Open questions: Produce OQ-NNN files for genuinely unanswered threads discovered during synthesis.
6. Emit status-recommendation: Output status-recommendation: answered or status-recommendation: deferred (with reason). The dispatcher writes status transitions to Q-NNN frontmatter per Amendment 1. Workers never update frontmatter status fields directly.
7. Actionable findings: Every finding must map to a concrete design constraint or implementation recommendation for the UK job-hunter SaaS. Findings that are purely descriptive without an actionable implication are incomplete.

Finding Scope by Domain

Findings should address practical engineering questions:

• 01-scraping: Which platforms are feasibly scrapable, at what rate, with what ToS risk?
• 02-dedup: Which dedup algorithm fits the expected job posting volume and accuracy requirements?
• 03-visa-sponsorship: What is the optimal detection method — keyword, NLP, or sponsor register lookup?
• 04-uk-job-market: Which job categories/regions have the highest search demand, informing crawl priority?
• 05-gdpr-legal: What lawful basis applies, what data retention limits apply, what consent UI is required?
• 06-search-ranking: What ranking model fits the required latency and index size?
• 07-ai-integration: Which AI models/APIs offer the best cost-accuracy trade-off for classification tasks?
• 08-infrastructure: What deployment architecture supports the target scale with acceptable running cost?

Self-Check (Level 1 Self-Correction)

Before completing, verify:

• [ ] Evidence chain has no gaps (every claim has at least one source)
• [ ] Confidence levels stated on all findings
• [ ] No contradictions with existing accepted decisions
• [ ] Every claim cites S-NNN sources
• [ ] Status-recommendation is emitted
• [ ] Every finding has an actionable implication

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

Content Conventions

See references/synthesis-conventions.md for:

• Finding format (evidence chain table, confidence levels, caveats)
• Open-question format (discovery context, suggested next steps)
• Decision format (for reference — decisions are produced by rd-* workers, not rw-synthesize)