Grand-to-Ground (G2G)

An adversarial, Lakatos-inspired research loop for AI-assisted social science.

AI research systems that work in biology and machine learning assume a lab exists to close the loop. Social science has no lab — its experiments are history, its theories are built by deduction, and its hardest step is choosing the right question. G2G is a research methodology, packaged as a Claude plugin, for exactly that space: treat your grand framework as a question generator, converge each cycle onto one identifiable, falsifiable question, and let independent AI critics attack everything except what you explicitly chose to protect.

Built for economics and finance first; the loop generalizes to any social science. v0.2.0 is the field-hardened release: every change comes from one month of intensive live use on a live finance research programme — five tribunals, three full cycles, six conclusion cards, six generations of research plans — with each lesson logged before it became a feature.

P0 vision → P1 recon (gate; incremental after cycle 1) → P2 formalize (multi-operationalized)
→ P3 axiomatize (hard core + declared assumptions + belt) → P4 conjecture
→ P5 parallel deduction (N independent deducers → merge) → P6 tribunal (3 critics ⇄ defender, severity dial)
→ P7 ground (3 questions) → P8 score (5 axes) · human disposes · record · loop
                └── convergence cut: the identifiable small question ──┘

What's inside

All programme state lives in a g2g/ folder of plain Markdown in your project — portable, inspectable, versionable.

Install

Claude Code

/plugin marketplace add HongtaoMa1/grand-to-ground
/plugin install grand-to-ground@grand-to-ground

Claude Cowork (desktop): download grand-to-ground.plugin from Releases and open it — Cowork shows an install prompt.

Manual (any Claude with skills support): copy the skills/* folders into your skills directory and agents/* into your agents directory.

Quickstart

1. Install, then say: "Set up G2G." You'll be asked for your research vision (rambling is fine), your framework's layers, the assumptions you want protected — and, new in v2: the assumptions you simply want declared (exempt from refutation, like a theorist's assumption list), how hard the tribunal should hit by default, and your field's actual conventions.
2. Say: "Run a G2G cycle." Watch independent deducers disagree, three isolated critics attack, a defender recompute the numbers, and a verdict hand you exact surviving sentences — plus replacement language for what died.
3. Say: "Draft a research plan." You get a tick-to-approve plan: modules you can clear with "R1 ✓ R2 ✓", a decision bar you can settle from one screen, and a one-pager for your supervisor in which every sentence matches a frozen verdict.
4. Anytime: "G2G status" for programme health, or "Run a tribunal on this draft" to stress-test anything — no setup needed.

Design principles

• AI proposes; the human disposes. You ratify axioms, arbitrate disputes, and review every verdict. And in v2, explicitly: scores inform; the human disposes — no automatic demotion, no burial by objection pile-up. Errors belong to the human; the AI's pride is fidelity, divergence, and professionalism.
• The tribunal attacks the execution, not your premises. Field evidence was unambiguous: what dies under adversarial review is almost never the researcher's story — it is the shape the agent added and the qualifier the summary dropped. v2 aims the firepower there: declared assumptions are exempt and ledgered, agent additions carry [SIG] marks, and verdicts kill sentences, with replacement language.
• Anti-sycophancy is mechanical, not aspirational. Critics carry minimum-objection quotas, endorsement bans, and fresh contexts per round. The severity dial changes depth, never honesty.
• Anti-dogma rules (R1–R4). Framework families — and now data domains — rotate between cycles; registered questions can't be re-cut; every conclusion must link two layers; every cycle must propose a cross-level mechanism.
• Lakatos, operationalized. A protected hard core, a declared-assumption tier, an attackable belt; attacks on protected tiers are logged and audited — the status skill will tell you, bluntly, whether your programme is progressive or degenerating.
• Falsifiability is scored, every round. And so, now, is sellability — a paper sells one idea; the loop notices before the referee does.

What G2G is not

Not autonomous science — it will not write your paper or replace your judgment; by design it declines the autonomy race. Not a cheerleader — the tribunal does not flatter, by construction. Not a model picker — it is a discipline for the loop around whatever models you use. The methodology is the cheap part; the discipline — verify, ground, stop — is the work.

Methodology

The full methodology paper is in docs/METHODOLOGY.md (v2.0; the bilingual PDF is distributed alongside releases). Related work: Google DeepMind's Co-Scientist (Nature, 2026), Sakana's AI Scientist-v2, Korinek's AI Agents for Economic Research (NBER w34202), and the Stanford ideation study (arXiv:2409.04109). G2G differs by targeting deduction-first, lab-free disciplines and by making the human gates and anti-dogma rules mechanical.

See CHANGELOG.md for what changed in v0.2.0 and the field evidence behind each change.

Provenance

This plugin was designed and built with AI research agents, using the loop it implements — and v2.0 was revised the same way: the methodology draft stood its own tribunal before release. Cowork environment note: when verifying files written to a mounted folder, read them back with a host-side file tool — re-staging a file that already exists can return a cached copy.

Citation

If G2G helps your research, cite it (see CITATION.cff):

Ma, R. (2026). Grand-to-Ground (G2G): An adversarial, Lakatos-inspired research loop for AI-assisted social science (v0.2.0). https://github.com/HongtaoMa1/grand-to-ground

License

MIT © 2026 Rick Ma