GPD's real value is enforcing physics verification discipline, not replacing your AI agent

GPD is a command layer that installs physics-research workflows into whichever AI coding agent you already use — Claude Code, Gemini CLI, Codex, OpenCode, or GitHub Copilot CLI. It does not replace those runtimes; it adds a structured command vocabulary on top of them. One npx invocation drops the commands into your runtime's config directory, and from that point forward you work through a fixed ladder: scope the problem, plan phases, execute them with specialist agents, verify the results, then write up.

The verification step is where GPD earns its physics-specific framing. Rather than treating correctness as a vibe check, it explicitly targets dimensional consistency, limiting cases, symmetry constraints, conservation laws, and numerical stability. That list is not decorative — it maps directly to the kinds of errors that survive informal review in theoretical and computational physics. The system also locks notation and sign conventions across up to 18 physics subfields for the lifetime of a project, which matters when a multi-phase calculation accumulates enough state that a convention drift becomes invisible.

The project structure is deliberate: Project → Milestone → Phase → Plan → Wave. Decimal phases like 2.1 exist specifically for urgent work that appears mid-milestone without disrupting the integer phase numbering. Waves inside a phase encode execution dependencies, so plans with no unmet prerequisites can run in parallel. This is not a novel scheduling idea, but applying it explicitly to physics research workflows — where a numerical convergence study might block a derivation that blocks a figure — is a reasonable design choice.

The README is candid about what GPD is not. It does not provide model access, API credits, or billing. It does not fabricate provider internals it cannot observe. The known limitations section is short but honest: runtime-internal tool and subagent detail is bounded by what the active provider exposes. The autonomy framing — Supervised mode as the default, Balanced as a graduation — treats the researcher as an advisor reviewing a graduate student's output, which is a more accurate mental model than "autopilot."

The citation file and the small but real list of published papers that acknowledge GPD suggest this is already in active research use, not just aspirational tooling. Version 1.2.2 is cited in the BibTeX block, and the four listed papers all carry 2026 arXiv identifiers, which puts the tool's production use roughly concurrent with its public release.

The publication pipeline — write-paper, peer-review, respond-to-referees, arxiv-submission — is the most speculative part of the system. The README is careful to bound what each command actually does versus what a researcher still owns, but the boundary between "GPD-authored outputs" and researcher judgment is exactly where agentic systems tend to blur in practice. That tension is not resolved here; it is just named.