Neuro Hound

Primary: Produce the weekly briefing

1. Fetch from all configured sources (PubMed, RSS feeds, Tavily wideband)
2. Pre-filter with domain-specific regex (fast, free, deterministic)
3. Deduplicate against scored history
4. Score each candidate with LLM (1–10, category, assessment, vaporware flag)
5. Cluster scored items into 2–5 themes with significance ratings
6. Write an executive brief (TL;DR, themes, alerts, what-to-watch)
7. Review the brief via Reflection Pattern (reviewer LLM critiques and adjusts)
8. Produce HTML report, dashboard, markdown, alerts JSON, full results JSON
9. Log to MLflow (parameters, metrics, artifacts)

Meta: Maintain and evolve source coverage

10. Company discovery: After scoring, extract new company names from high-scoring Tavily results. Write candidates to discoveries.yaml for human review and promotion to the watchlist.
11. Domain discovery: Track which web domains yield high-scoring items across runs. Propose new RSS sources when a domain is consistently valuable.
12. Source health monitoring: Track per-source yield (items fetched, in-scope count, last hit date). Flag cold sources that haven't produced in-scope items in 30+ days.
13. Vocabulary maintenance: Extract domain terminology from high-scoring papers and propose additions to vocabulary.yaml. The PubMed query and regex pre-filter are constructed dynamically from this vocabulary. Term counts self-stabilize as domain vocabulary is finite. A configurable max_terms_per_category limit prevents unbounded growth.

Status: Goals 10–13 are now handled by the meta_reflect ReAct agent (Phase 9). The LLM reasons about whether and when to pursue these goals based on the pipeline's output. See nodes/meta_reflect.py and docs/ADR-001-agent-specification.md.

Tools Available

The skill has access to the following tools, implemented as Python modules:

Workflow

fetch_pubmed → fetch_rss → fetch_tavily → save_registry
    → prefilter (regex + dedup)
    → [conditional: skip LLM if nothing in-scope]
    → score_items (LLM × N items)
    → summarize_themes (cluster + significance)
    → write_brief (executive briefing)
    → review (Reflection Pattern + dedup update)
    → meta_reflect (ReAct agent — vocabulary, sources, companies, coverage)
    → outputs (HTML, dashboard, markdown, JSON, MLflow)

Implemented as a LangGraph StateGraph with a conditional edge after pre-filtering: if nothing passes regex, the LLM pipeline is skipped entirely (no API cost on quiet weeks).

The meta_reflect node is a genuine ReAct agent: a multi-turn Thought → Action → Observation loop where the LLM reasons about the pipeline's output and decides which self-improvement tools to invoke. It may call zero tools (quiet week) or several (vocabulary gaps, cold sources, new companies). Max 5 iterations per run.

Scoring Criteria

Items are scored 1–10 by the LLM using criteria defined in prompts.yaml:

Categories: implantable_bci · ecog_seeg · stimulation · materials · regulatory · funding · animal_study · methods · out_of_scope

Behavioral Constraints

These are inherited from SOUL.md and enforced in prompts and code:

• No hallucination. If a source isn't found, say so. Never fabricate links.
• Peer-reviewed over press releases. Prefer primary literature. Flag vaporware and marketing.
• Verify before alerting. Priority 9–10 alerts only for human clinical milestones or FDA regulatory milestones.
• Evidence-based opinions. Have opinions on scientific validity, but ground them in what the data actually says.
• Concise over comprehensive. The executive brief should be readable in 5 minutes. Details go in the scored items list.

Self-Assessment

After each run, the review node evaluates:

1. Are significance assessments calibrated? (Overhyped or underappreciated?)
2. Are items correctly categorized? Any miscategorizations?
3. Did the analyst miss themes or connections between items?
4. Any vaporware or marketing that slipped through?
5. Which 1–3 items are genuinely most important this week?

The reviewer can adjust scores, flag missed signals, and call out vaporware. Quality score (1–10) is logged to MLflow for tracking over time.

Configuration

All runtime parameters live in two YAML files (no code edits needed):

config.yaml — Sources, models, company watchlist, curated industry sources, Tavily queries, MLflow settings. This is the single source of truth for what the agent monitors and how.

prompts.yaml — All LLM prompts as templates with {variable} placeholders. Edit here to iterate on analysis quality. Prompt text is logged to MLflow for A/B tracking.

vocabulary.yaml — Domain vocabulary organized into primary terms (domain-defining) and qualifier terms (relevance-filtering). The PubMed query is constructed dynamically at runtime. Terms are extracted from representative papers; provenance is tracked. A max_terms_per_category setting (default: 0 = no limit) provides a configurable ceiling, though term counts naturally self-stabilize as domain vocabulary is finite.

Company Watchlist (config.yaml)

Tracked companies get automatic Tavily queries and RSS feeds (if Substack URL provided). New companies can be added manually or promoted from discoveries.yaml (auto-generated after each run).

Curated Industry Sources (config.yaml)

Websites without RSS (e.g., Neurofounders, IEEE Spectrum) are searched via Tavily site: queries.

Invocation

# Full pipeline
python3 skills/neuro_hound/run.py --days 7

# Phase 1 only (regex scoring, no LLM cost)
python3 skills/neuro_hound/run.py --days 7 --phase1-only

# With specific models
python3 skills/neuro_hound/run.py --days 7 --model gpt-4o --reviewer gpt-4o

# From Telegram (via OpenClaw)
/neuro_hound --days 7

Outputs

Environment Variables

Evolution

See docs/ADR-001-agent-specification.md for the full architecture rationale.