v0.2.0 — the field-hardened release

Methodology v1.0 → v2.0 · Plugin v0.1.0 → v0.2.0

G2G just spent a month running hot on a live finance research programme: five adversarial tribunals, three full research cycles, six conclusion cards, six generations of research plans — with every process defect logged the day it bit. v0.2.0 is those logs, turned into machinery. (This release was itself tribunal-reviewed before shipping, per its own rules.)

The headline finding, and the redesign it forced

Across five tribunals in a single month, what died under adversarial review was almost never the researcher's idea — it was the shape the agent added during formalization, and the qualifier that vanished between the internal record and the outward text. So v2 re-aims the firepower:

• Declared assumptions (new tier). Premises you state as assumptions — the way a theorist lists them before a proof — are exempt from refutation. The tribunal checks that reasoning stays inside them, not whether they're true. Attacks on them go to a silent ledger, surfaced at audits.
• Severity dial. Exploratory / standard / submission — you choose how hard the court hits. Depth changes; honesty doesn't (quotas, sourcing, and no-endorsement hold at every tier).
• Parallel deduction. N independent deducers, merged only after each brief is frozen — the cheap correction for LLM path dependence.
• Sentence-level verdicts. Kill tables with mandatory replacement language; three-value verdicts (alive / dead / undecidable); a patrol on the five verbs of overreach.
• An independent defender (g2g-defender, new agent) that recomputes every computable allegation from your data files instead of arguing — and reports errors on its own side with equal prominence.
• The plan layer (g2g-plan, new skill). Versioned, tick-to-approve research plans: modular repair/offense lines, a four-line decision bar showing what each open decision blocks, plan-level tribunals (ST series) before anything freezes or ships, and collaborator one-pagers aligned sentence-by-sentence with frozen verdict wording.
• Five-axis scoring adds sellability — a paper sells one idea; the loop now notices before the referee does. And explicitly: scores inform, the human disposes — no automatic demotions, ever.
• Human-side protocol, written down: the four-line decision format, ID discipline (no more R‴2), session hygiene, and the two-seat planning/execution protocol with handover kits.

Install / upgrade

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

Cowork (desktop): open grand-to-ground.plugin from this release's assets. Existing v1 workspaces migrate automatically on first run — new files are added, old records are never rewritten.

Full change list with the field evidence behind each item: CHANGELOG.md. Methodology paper: G2G_Methodology_v2.pdf in this release's assets (bilingual EN/中文).

v0.1.0 — first release

First public release of Grand-to-Ground (G2G) — an adversarial, Lakatos-inspired research loop for AI-assisted social science.

Install (Claude Code)

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

Install (Claude Cowork, desktop)
Download grand-to-ground.plugin below and open it — Cowork shows an install prompt.

What's inside: 4 skills (setup, cycle, tribunal, status) and 4 agents (a deducer plus three independent critics). The human researcher is the final arbiter. MIT licensed.