G2g setup

G2G Setup

Initialize a Grand-to-Ground research programme. G2G treats the researcher's grand theoretical framework as a question generator, and converges each cycle onto one identifiable, falsifiable question. This skill creates the persistent state that every other G2G skill reads and writes.

Step 1 — Check for an existing programme

Look for a g2g/ directory in the working folder. If it exists and contains 00_programme.md, tell the user a programme already exists, summarize its charter in 3 lines, and ask whether to (a) resume it, (b) archive it and start fresh, or (c) abort. Never silently overwrite. v1 → v2 migration: if the programme predates v2 (no severity tier in the charter, no 05_plans/), offer to migrate: add the missing files and charter fields with defaults, touch nothing that exists. Old records are never rewritten.

Step 2 — Intake interview

Ask the user the following, in one batch (use the platform's structured-question UI if available; otherwise plain text). Do not proceed until you have answers to 1–3; use stated defaults for the rest if skipped.

1. Vision. "Describe your grand framework or research vision in plain language — hunches, end-state claims, and all. Rambling is fine; formalization happens later." (Free text. This is the P0 input.)
2. Discipline and layers. "Which field is this, and what are the 2–4 layers of your framework?" Offer examples: economics/finance → micro behavior / distributional / macro; sociology → individual / institutional / societal; political science → voter / party / regime.
3. Protected assumptions (hard-core candidates). "Which 2–5 assumptions is your programme built on — the claims you want protected while everything else gets attacked?" Explain in one sentence: G2G protects a Lakatosian hard core but logs every attack on it and audits programme health over time.
4. Declared assumptions (exempt working premises). "Beyond the core: any working assumptions you want to declare rather than defend — premises you'd list at the top of a paper, the way a theorist assumes purchasing-power parity? These are exempt from refutation: the tribunal only checks that reasoning stays inside them." (Default: none; they can be declared any time later.)
5. Severity default. "How hard should the tribunal hit by default? exploratory (hunches and essays — 3 objections per critic, one defense round) / standard (working research — 5 objections, two rounds) / submission (anything leaving the building — 5 objections, three rounds, independent defender with data recomputation)." (Default: standard. Any single tribunal can override.)
6. Field conventions. "What are your target venue's actual conventions — anything reviewers there never do or always expect? (e.g., 'pre-registration rituals are rare in empirical finance')." Seed each answer as a standing ruling so critics calibrate to the venue. (Default: none.)
7. External collaborator. "Is there a supervisor or co-author whose rulings should be archived and whose materials (one-pagers, call notes) the plan skill should prepare? Name and role." (Default: none.)
8. Evidence environment. "What data, archives, or natural experiments can you plausibly access?" (Default: "to be determined during first grounding".)
9. Budget and thresholds. Scoring pass line (default 18/25 with falsifiability ≥ 3/5 and sellability ≥ 2/5), tribunal rounds per question (per severity tier), critics count (default 3).
10. Rotation menu. Show the discipline's default theoretical-framework menu from references/rotation-menus.md and ask if they want to add or remove entries. Note that rotation runs on two axes: framework family AND data domain.

Step 3 — Scaffold the workspace

Create the g2g/ directory with these files, filling templates from references/templates.md with interview answers:

• g2g/00_programme.md — charter: vision verbatim, discipline, layers, severity default, field conventions, collaborator, thresholds, rotation menu (two-axis), ID legend (pre-filled with the reserved series; every new letter series must be registered here at first use), current phase.
• g2g/01_hard-core.md — hard core marked RATIFIED after Step 4; Declared assumptions section (exempt tier, with quick-scan dates and the silent attack ledger); Protective belt; Attacks on the core log; Standing rulings register (U-series, seeded from question 6); Corrected facts; Pending axiom proposals.
• g2g/02_question-registry.md — registry table (ID, question, framework family, data domain, status, card link, expected cross-level mechanism) plus a seeds section. The two-axis genealogy line lives here.
• g2g/03_records/contradiction-log.md — empty, with entry format.
• g2g/03_records/literature-ledger.md — the P1 incremental-mode asset: nearest neighbors, one-line verdicts, last-checked dates.
• g2g/03_records/ — conclusion cards accumulate as card-CXXX.md; tribunal verdicts as tribunal-XX.md / tribunal-ST-XX.md; re-score appendices alongside their cards.
• g2g/04_log.md — session log; write the first entry (date, "programme initialized").
• g2g/05_plans/ — versioned research plans (plan-v001.md, …) from the g2g-plan skill.
• g2g/06_process-defects.md — a living file, mostly written for the human: process defects discovered in use, and the human-side rules that fix them. Pre-seed it with the four-line decision format and session-hygiene rules from the template.

Step 4 — Fact-check before ratifying

Hard core: for each candidate axiom, run a quick web check for direct factual conflicts (a false "fact" planted in the core will be defended brilliantly forever after). Report conflicts with sources and let the user decide: revise, keep with a caveat, or demote to the belt. Record corrections. Only then mark RATIFIED.

Declared assumptions: run a lighter one-time screen — factual howlers and carrier errors only (does the named institution/channel/mechanism exist and carry what the assumption needs?). Do not litigate plausibility; that is the whole point of the tier. Record the scan date. After this, attacks on these assumptions go to the silent ledger only.

Step 5 — Optional: project entry-point integration

Offer to append a short G2G section to the project's CLAUDE.md (or equivalent entry file) so any future session finds the programme. The snippet must follow the entry-file rule: no version or round numbers ever — always "the highest-numbered file is current"; and it should name the reading order (00_programme.md → latest plan → latest verdict).

Step 6 — Confirm and hand off

Summarize the charter in ≤6 lines (include severity default and any rulings seeded). Tell the user: "Programme initialized. Say run a G2G cycle to start cycle 1, draft a research plan for the plan layer, or G2G status anytime for programme health." Do not start a cycle without being asked.

Invariants (apply to every G2G skill)

• The human ratifies axioms, arbitrates disputes, and reviews every verdict. AI proposes; the human disposes — and scores inform; the human disposes too: no automatic demotion or burial.
• Errors belong to the human; the AI's pride is fidelity, divergence, and professionalism — never holding the same view as most of the world.
• Nothing enters the hard core unverified. All guesses are marked as guesses; all agent-added shapes are marked [SIG].
• Declared assumptions are exempt from refutation; deviations from them are not.
• State lives in g2g/ as plain Markdown — portable, inspectable, versionable. Any decision not written into a file did not happen.