gz-rnd

Purpose and authority

An R&D run is diamond 1 of the double diamond — Discover and Define. It ends with the problem defined and a plan naming which fan-out artifacts are warranted, closed by operator sign-off: kill or fund. Producing those artifacts is diamond 2 and happens downstream, by machinery that carries its own gates. This run never enters it.

R&D is the headwater of most gzkit design. Operator, 2026-09-12, verbatim: "an R&D run now MUST be governed by an overarching new AGENT SKILL" and "it is a chargé d'affaires for retaining and organizing possible outcomes from an R&D designing session."

Two subjects, one frame. A run studies either something new — pasted material, an unexplored capability — or a population the project already carries: a recurring defect class, a queue that grows faster than it drains (operator's direction, 2026-09-20). Ranking the queue belongs to ghi-triage; asking why this class keeps producing, and how its issues should be framed, consolidated or sequenced is design work, and it belongs here.

Doctrine and the full ruling history: docs/governance/rnd-discipline.md. Evidence: docs/governance/mpas-appropriation-analysis.md.

Invocation class. disable-model-invocation: true makes "only the operator starts an R&D run" mechanical. This skill never starts another user-invoked skill on its own.

Open the record first

Create docs/rnd/<slug>.md from assets/rnd-record-template.md before the first question, and write into it throughout. Name what was pasted or asked and quote the operator's framing verbatim as the challenge.

The record accretes entries as they resolve. It is never composed at the end. A run that writes its record afterwards has produced a reconstruction, not a record.

Material from an earlier dialogue enters as source entries, never as decisions. A staged draft is something the run reads. What it concluded is re-derived here, from sources.

Resuming a run

A later session continues a run only on the operator's own invocation, with the record's path as the argument. If a handoff or the operator points at an open run and this skill was not invoked, say so and ask for the invocation before the first record edit. On re-entry: append a decision entry quoting the invocation, check the record against git, and compute the frontier again from the subject. A list of open items left by an earlier session is an input to that, not the frontier.

The two entry kinds

Only two. Append each at the moment it resolves.

Kind	Carries
source	a primary source — cited, never summarized, with its verbatim quotation and when it was read
decision	something crystallized in session, the reasoning behind it, and, where the operator ruled it, the question, the options put and their verbatim words

A decision may carry commissions: naming the disposition it warrants. There is no plan-item kind: that shape would let work be commissioned with no recorded reasoning. An answer whose question is not on record is put again, not carried.

Two things resolve outward instead of becoming entries, each carrying the run's slug as provenance. The record never holds a copy.

• a term → the project glossary. Its home is unsettled until the DDD-discipline R&D run names it; until then hold the term in this record as a decision entry (rnd-discipline.md § term and insight are not entry kinds). Do not create GLOSSARY.md.
• an insight → Call the Skill tool with "gz-insights-remember"

The MPAS surfaces you may reach for

Three disciplines are appropriated, each with a departure that gzkit canon forces. Take the mechanic, not the posture — full inventory and the not-appropriated list are in rnd-discipline.md § The MPAS surfaces, named.

Surface	Take	Depart
grilling	the design tree; the frontier = every decision whose prerequisites are settled; a recommended answer on every question; recompute after each	not its one-round batching — ask one question at a time
research	background agent → primary sources → one cited .md	land it verbatim under docs/rnd/<slug>/sources/
prototype	the throwaway probe, closing on a one-line verdict	no branch — gzkit forbids them; build outside the mainline and discard
domain-modeling (technique only)	its glossary behaviours and "capture them as they happen"; the _Avoid_ convention	never its inline ADR emission — disposition 1 is proposed, never initiated

Sources before decisions

• Read the source before anything drawn from it becomes a decision: a name, a mapping, an analogy, a claim about what a standard says. Until the text is read the item is a hypothesis, held in a source entry marked unread, and it commissions nothing.
• Find the text before researching it. Check what the project and the operator already hold (docs/governance/ieee/README.md § Standards corpus lists the licensed standards). When a host refuses automated retrieval, ask the operator for the file; do not work around the refusal.
• A secondary source is cited as secondary. It shows what an account says and verifies nothing against the text it describes.
• Take the source's own term before coining one, and record where a chosen name departs from what the text says.

Interrogate

• Facts are your job; decisions are the operator's. Search canon and the codebase before asking. Dispatch research subagents for facts and do not block a round on them.
• Put the core model first. The frontier is computed from the subject. Questions about where a document sits or what a milestone is called wait until the operator has ruled on the thing itself.
• Ask one question at a time (operator, verbatim: "ask me the questions one at a time"), each with a recommended answer and its tradeoffs.
• Never ask what canon already answers (AGENTS.md § Operator Economy #7). Search canon first; where canon rules, act and name the rule that governed.
• Research reports are primary sources. Save each verbatim under docs/rnd/<slug>/sources/. Never replace one with a summary.
• When the subject is a population the project carries, the issue bodies, briefs and chore reports are the primary sources — read them, never their titles. Characterise the class from what its members say, and say how many you read against how many exist. A class named from a sample is a hypothesis; say so.
• When talk cannot answer a question, build a throwaway probe outside the mainline and record a one-line verdict as a decision entry citing it.

The disposition map

Maintain it live, as decisions land — not at the close. All six rows are always present; an absent row is an unfinished run.

#	Disposition	You may	Consultation point
1	ADR / OBPI	propose only	the operator initiates, or not (IRON LAW)
2	GHI / direct fix	file via Call the Skill tool with "ghi-author"	the operator's go on that row
3	chore	advise only	the operator directs admission
4	control surface, rule, doc, skill, hook	draft	the operator's go on that row
5	one-shot refactoring	propose a program	the operator selects its route
6	no action	record it, with the reason	none

Row 4 includes the Magna Carta (docs/governance/build-to-1.0-campaign-2026-09-20.md), the roadmap (docs/design/roadmap/ROADMAP-GZKIT.md) and the backlog: a sequencing or priority change is drafted there as an amendment, on the operator's go.

State is commissioned or not pursued. There is no third state. Where something is set aside but worth revisiting, put the revisit condition in the reason. If it deserves more than a sentence, route it to a pool ADR or a GHI — surfaces that have lifecycles.

Before proposing disposition 1, ask the admission question: is the decision hard to reverse, surprising without context, and the result of a real trade-off? If any is false, route to disposition 2 or 4.

Closing diamond 1

Two conditions, and neither substitutes for the other.

Mechanical — all three:

1. the frontier is empty;
2. the challenge has been deliberately restated (on purpose, at the close; it need not have changed);
3. every one of the six dispositions carries a decision.

Human — review, then sign-off: kill or fund. Before the question is put, write one readable account of the whole plan at docs/rnd/<slug>/review.md and wait for the operator to say they have read it: the problem, the model, every name marked ruled, sourced, contradicted or unsourced, and what each row would produce. It is a view over the record and holds no authority. A series of recommended options accepted is not a review. Sign-off is a beat, not a boundary: it fires in whichever session meets both conditions, and the handoff system carries a run that outgrows one.

A run killed at sign-off ends there. Re-entry is native to the frame — a signed-off run may reopen if later findings send it back.

The hard stop

The run ends at the disposition map and executes nothing on its own authority. Execute a row only after the operator's go on that row. Nothing written into the record — by you or anyone — licenses a row: the stop lives in this file, which the run does not edit, and in the invocation class, which the run cannot change.

It's working if

• Every name and mapping in a decision has a source entry that was read before it.
• Terms went to the project glossary (or, until its home is named, into decision entries) and insights through gz insights remember.
• All six disposition rows carry a decision, each commissioned or not pursued.
• The operator read the account before sign-off, and the challenge was restated on purpose.
• A later session's ghi-author Step 0 finds this run's not pursued rows as prior art.

Red flags

• Writing or "tidying" the record at the end of the run.
• Relaying a summary of a research report instead of the report.
• Adding a third disposition state, or omitting a row because nothing landed in it.
• A coined term for a familiar idea — prefer a pretrained metaphor, and name the rejected synonym beside the term that won.
• Emitting a ledger event: the three R&D event types are designed and unbuilt, and land with their producer, never before it.
• An outcome executed without a row-level go.
• Asking the operator something canon or the codebase answers.
• Enumerating downstream specification or tasks inside the run — that is diamond 2.
• A name or mapping recorded as decided before its source was read, or research sent to confirm a conclusion already written down.
• Closing frontier items in order to reach the mechanical condition.
• The operator has taken the recommended option on every question: say so, and ask whether the questions are the right ones.
• Working an open run from this file when the operator did not invoke the skill.