G2g plan

G2G Plan

The versioned research plan is the steering wheel of a G2G programme: it translates tribunal verdicts into next actions, carries the human's approvals, and briefs whoever executes. This skill writes and iterates plans in g2g/05_plans/ (plan-v001.md, plan-v002.md, …). Requires an initialized g2g/ workspace; without one, offer to run g2g-setup or produce a standalone plan using the same anatomy.

Step 1 — Read the state

Read, in order: 00_programme.md (charter, severity, ID legend, collaborator), the highest-numbered plan in 05_plans/ (its approval record at the file end is the baseline — never renumber or reinterpret recorded approvals), the latest tribunal verdict(s) and cards in 03_records/ (the substitution tables there are the binding wording baseline), and 02_question-registry.md (seeds and genealogy for offense candidates).

Step 2 — Draft plan vN+1

Anatomy, in this order:

1. Header — version, date, status (draft — awaiting ticks), inputs (which verdicts/rulings/feedback this version digests).
2. Plain-language digest (§0) — half a page the human reads first. No internal jargon, no IDs without a gloss. What happened, what this version changes, what they need to do. Write it last, place it first.
3. Module lines, tick-to-approve:
   • Repair line R1…Rn (mandatory): one module per accepted tribunal finding. Self-containment rule: every module that responds to an objection carries a one-line plain restatement of that objection — "R3: (critic found the control window overlaps the event window, so the placebo is contaminated) → re-draw windows disjoint." A plan that says "fix per L8, see tribunal file" has outsourced its own comprehensibility; the human must be able to review without a decoder.
   • Offense lines A/B/C… (optional): from tribunal salvage and registry seeds, each with its expected cross-level mechanism in one sentence and its rotation-compliance note (both axes).
   • Every module gets a tick box. Approval is a tick: "R1 ✓ R2 ✓ A skip" completes review.
4. Decision bar — every open decision in the four-line format: what was done (one sentence of context) / what you must decide (a concrete choice, never an open question) / default ("takes effect unless changed", genuinely executable) / cost of not deciding — plus a blocks column naming what stalls while it waits. Three prohibitions: no "see §X" pointers that make the human fetch context; no missing defaults; no undefined jargon inside a decision item. Adequacy standard: decidable from this screen alone.
5. Frozen-wording appendix — sentences this plan carries forward from verdicts, quoted exactly from the substitution tables (qualifiers included). The plan may compress logistics, never wording.
6. Approval record (file end) — dated ticks and rulings as the human writes them. Append-only.

Step 3 — Version discipline

• Generating vN+1 triggers the archive prompt: list vN's handover items (task briefs, start prompts) that should move to the executed round's directory, and remind the human. Do not delete anything yourself.
• Entry files never name versions. If the project's CLAUDE.md (or equivalent) references plans, the phrasing is "highest version is current" — a hard-coded round number in an entry file once silently pointed a whole execution round at the previous round's spec.
• IDs follow the programme legend (00_programme.md); new series must be registered there first. Revisions use dotted numbers (R2.1), never accumulated primes.

Step 4 — The plan tribunal gate

Before a plan's propositions enter a frozen registry, or the plan (or any material derived from it) goes to an external collaborator: the plan stands tribunal — invoke g2g-tribunal in plan mode (ST series). The target is the new shapes and mappings this plan adds (formalizations, framings, outward translations), not the frozen results it cites. Exemptions: purely editorial revisions (wording, layout) don't trigger; a small change to an already-tried plan (no new shapes) doesn't trigger, but is logged. Field record: the first plan tribunal found fourteen dead sentences, all in the translation layer — the gate exists because that layer had no other guard.

Step 5 — Collaborator materials (when the charter names one)

• One-page brief: the plan's story in the collaborator's professional language — standard field terminology only; internal pet names never leave the building. Every results sentence passes the step-down rule: aligned sentence-by-sentence with frozen verdict wording, qualifiers intact.
• Call notes card: the questions to ask, the answers to have ready (including the standard answer to any question previously fumbled), decisions wanted from the call.
• Rulings archive: after contact, record the collaborator's decisions verbatim and dated in the plan's inputs section (and as standing rulings if they are standing) — collaborator rulings are argued with the human, never silently re-litigated by agents. A collaborator's spoken story receives the P2 treatment when formalized: multi-operationalization, [SIG] marks — their intuition has the same rank as the user's; the signature dies, not the story.
• Discipline invisibility: outward materials never showcase the internal machinery (frozen registries, tribunals, substitution tables) — results, nulls, caveats, and data gaps are reported in full; process narrative is not. Red line: a test added after results were known is never described outward as pre-specified. "Not saying" ≠ "saying the reverse."

Plan-first mode

On request ("plan-first", "no empirics yet", or a collaborator gate in force): iterate plan → digest → human/collaborator feedback → plan vN+1, with no empirical execution, until the frame is agreed. Each iteration still passes Step 4 before going outward. Cheaper than running studies nobody has agreed to want.

Handoff

End every run with: the digest, the tick list, the decision bar (with blocks), and — if an execution seat exists — the handover kit per references/collaboration-protocol.md. One session, one plan version; write the state, then stop.