Skip to content

cross-domain-engine-v2.md

Latest commit

 

History

History
54 lines (41 loc) · 2.82 KB

cross-domain-engine-v2.md

File metadata and controls

54 lines (41 loc) · 2.82 KB
Raw

Cross-domain Rule Engine v2

Why: LLM-generated cross-domain discovery produces too many false positives ("both domains mention transformers"). Deterministic rules are auditable, predictable, and encode the user's actual cross-domain hypotheses rather than letting the model guess.

What changed in v2

v1 v2
Rules 2 (R001, R003) 7 (R001-R007)
Config source active-config.json Built into the script (no config dependency)
LLM significance none One sentence per insight batch
GitHub-repo convergence not detected R006 reads compute-gh-adoption.py output
Hypothesis-driven transfer not supported R007 (LLM generates AI→VLA transfer hypotheses with per-cycle cap)

The 7 rules

ID Label Direction Trigger
R001 vla-technique-cross VLA → AI App Technique keywords (diffusion, flow matching, transformer, RLHF, quantization, distillation, …) appearing in ⚡ / 🔧 / 📖 VLA items
R002 ai-framework-to-vla AI App → VLA New agent / orchestration frameworks (LangGraph, CrewAI, AutoGen, …) being adopted in robotics stacks
R003 ai-embodied-to-vla AI App → VLA Embodied-AI papers from generalist labs reaching VLA practitioners
R004 vla-foundations-to-ai VLA → AI App Foundation-model and pre-training methods originating in robotics
R005 paradigm-fusion bidirectional Both domains converging on the same paradigm in the same week
R006 github-repo-convergence bidirectional ≥3 monitored repos touching shared dependency / method / benchmark in the same week
R007 hypothesis-driven-transfer AI → VLA LLM generates N transfer hypotheses per cycle, capped to avoid spam

Output

memory/cross-domain-insight.json — append-only, 60-day retention. Each insight:

{
  "date": "2026-05-12",
  "rule_id": "R005",
  "label": "paradigm-fusion",
  "items": [
    {"title": "...", "rating": "⚡", "domain": "vla"},
    {"title": "...", "rating": "🔧", "domain": "ai_app"}
  ],
  "significance": "Both domains converged on diffusion-flow-matching this week — VLA via Pi0.5, AI app via Flux Schnell — suggesting the technique is moving from research to production tooling."
}

The significance line is the only LLM-generated content; rule matching itself is deterministic.

When to disable a rule

If a rule generates too much noise for your domain, comment it out in RULES at the top of cross-domain-rule-engine.py. Each rule is self-contained — no cross-dependencies.

Consumption

  • Weekly / biweekly reports read cross-domain-insight.json directly.
  • The web push step lifts the most-significant items into the public dashboard.
  • The MCP search_signals tool can grep the file by keyword.