0451 Software

0451-research-skills

Structured uncertainty quantification for agent outputs. CONACTU 4-axis confidence framework: Evidence, Model Fidelity, Completeness, Stability.

Embodied agent planning with integrated safety intention (SI) scoring and orthogonal feasibility+safety checks. For robots, vehicles, drones, game agents.

Search and retrieve academic papers from arXiv using their free REST API. No API key needed. Search by keyword, author, category, or ID.

Process arXiv PDFs into agent-notes — pymupdf4llm conversion, spawn 1 researcher per paper, max 3 concurrent, commit plans immediately.

Karpathy's LLM Wiki — build and maintain a persistent, interlinked markdown knowledge base from LLM-related sources.

Convert research PDFs to markdown (pymupdf4llm, not marker on M4), spawn researcher sub-agents, commit plans to agent-notes.

When and how to delegate non-trivial work to specialized persona sub-agents. Covers delegation triggers, context preparation, and result synthesis.

Safety evaluation for embodied agent plans — pre-execution checks for plans involving physical actions, tool use, or environment modification.

Apply Faberlens behavioral safety hardening to skills by comparing against the hardened-skills repo and selectively adopting guardrails.

Metacognitive monitoring and regulation prompts — MRO self-awareness loops, bias detection, mid-execution plan revision.

Adversarial red-team persona — find failure modes, stress-test assumptions, play devil's advocate.

Implementation persona — build working code with tests, practical and opinionated.

QA/inspection persona — verification, anti-pattern detection, challenge gate.

Research persona — deep dives, literature synthesis, domain reconnaissance.

Transform PSMAS DAGs of tasks into executable phases with checkpoint validation. For complex multi-step plans with dependency verification.

Evaluate whether a task request reflects benign, ambiguous, or potentially harmful intention before taking action.

Self-refinement loop using MRO three-role factorization — Monitor/Reasoner/Controller for iterative response improvement.

Comprehensive trip research pipeline: multi-agent delegation, mobile-formatted markdown output, format linter, and QA. For planning any trip from a task-list entry.