---
name: rd-decision-brief
description: Create evidence-first R&D decision packets with an assumption register, three competing paths, a candidate small discriminating experiment, precommitted go/pivot/stop thresholds, and a human decision record. Use for scientific, formulation, materials, product, or AI R&D planning; hypothesis comparison; experiment prioritization; go/no-go reviews; or auditing an AI-generated research recommendation.
license: MIT
---

# R&D Decision Brief

Turn an uncertain R&D question into four durable Markdown files that make evidence, assumptions, disagreement, and the next experiment auditable. Optimize for learning before commitment, not for producing a confident-sounding answer.

## Required outputs

Create one folder containing:

1. `DECISION_BRIEF.md` — question, constraints, three paths, tradeoffs, and current decision.
2. `ASSUMPTION_REGISTER.md` — stable IDs for facts, inferences, unknowns, and constraints.
3. `EXPERIMENT_CARD.md` — a candidate small experiment that can change the decision.
4. `SHARE_DRAFT.md` — an English and Chinese draft that requires human confidentiality review before publication.

Use the templates in `assets/`. For a new packet, run:

```bash
python3 scripts/new_brief.py <output-folder> --title "<decision title>" --owner "<decision owner>"
```

Do not overwrite an existing packet unless the user explicitly requests it.

## Workflow

### 1. Frame the decision

Capture the decision that must be made, the deadline, the decision owner, and the cost of being wrong. Identify hard limits for safety, regulation, privacy, budget, time, equipment, materials, and data.

Ask only questions whose answers can materially change the paths or experiment. Ask at most five in one batch. If the user cannot answer, record the item as `UNKNOWN` and continue without silently filling it in.

### 2. Run the assumption gate

Classify every material statement using the rules in `references/brief-format.md`:

- `FACT`: directly observed or supported by a traceable source.
- `INFERENCE`: a reasoned interpretation that may be wrong.
- `UNKNOWN`: missing information.
- `CONSTRAINT`: a fixed boundary for this decision.

Give each item a stable ID such as `A-001`. Record its evidence, impact if wrong, cheapest check, and owner. Never turn an inference into a fact because it sounds plausible.

### 3. Build the evidence ledger

For each decision-relevant claim, record:

- source or first-party observation;
- date and applicability to this context;
- conflicting evidence or known gap;
- status: `VERIFIED`, `UNVERIFIED`, or `CONFLICTING`.

Use primary sources when available. Never invent citations, measurements, confidence percentages, or regulatory conclusions. If live verification is unavailable, label the claim `UNVERIFIED` and state what would verify it.

### 4. Compare three paths

Create exactly three materially different paths when possible. Include a status-quo, null, or “do not proceed yet” path when it is a real option. For every path state:

- intent and expected upside;
- evidence that supports it;
- strongest weakness or rebuttal;
- evidence still required;
- an explicit abandon condition.

Do not collapse to one recommendation before comparing the paths against the same criteria.

### 5. Design a candidate small discriminating experiment

Choose a small safe experiment likely to distinguish the leading paths or invalidate a load-bearing assumption. Call it a **candidate small discriminating experiment**. Use “minimum,” “smallest,” “decisive,” or “optimal” only when a documented quantitative method and adequate data support that claim.

Define before execution:

- null and alternative hypotheses;
- discriminating prediction;
- baseline or control;
- variable, sample or replicate logic, and measurement method;
- primary metric and major confounders;
- `GO`, `PIVOT`, `STOP`, and `INCONCLUSIVE` thresholds;
- risk class and rationale, quantified budget and duration, qualified review, prohibited actions, safety boundary, rollback, and authorization owner.

Use `LOW` only for desk-only analysis with synthetic or non-sensitive public data, no physical operation, no external users, and no regulated, health, financial, security, legal, or personal-data consequence. Use the matching higher class when any of those signals appear. If classification is uncertain, choose the more protective class and explain the uncertainty.

Prefer an experiment that changes a decision over one that merely produces more data. Preserve negative results and deviations.

### 6. Record the human decision

The agent may recommend `GO`, `PIVOT`, `STOP`, or `NEED MORE EVIDENCE`, but the approval status starts as `PENDING`. Keep `Current decision` at `NEED MORE EVIDENCE` while approval is pending and place the agent view under `Recommended next move`. A final `GO`, `PIVOT`, or `STOP` requires a named human approver, an ISO approval date, and `Human confirmation: RECORDED`. Record dissent, override rationale, and the revisit trigger. Never impersonate approval.

### 7. Create the share draft

Write a concise English summary and a concise Chinese summary. Include the question, what was uncertain, the candidate experiment, and a link placeholder. Complete the confidentiality checklist, but leave publication approval as `PENDING` unless a named human reviewer explicitly approves it. Do not expose confidential formulas, personal data, unpublished measurements, or claims not present in the packet. Structural validation does not make a draft public-safe. Do not publish automatically.

### 8. Validate

Run:

```bash
python3 scripts/validate_brief.py <packet-folder>
```

Fix every error before presenting the packet as structurally complete. Warnings mean the packet is not execution-ready and must be disclosed, not hidden. Use `--json` when machine-readable results are useful.

## Review standard

A strong packet lets a reviewer answer all of these without guessing:

- What decision is being made, by whom, and by when?
- Which statements are facts, inferences, unknowns, or constraints?
- What evidence supports or conflicts with each important claim?
- What are the three real paths, including the cost of waiting?
- What result would make the team change direction?
- Who must authorize the experiment and final decision?

Read `references/evaluation-rubric.md` for scoring and `references/safety-boundaries.md` before work involving physical experiments, regulated products, health, finance, security, or other high-impact domains.

## Safety boundaries

This skill is a decision-support workflow, not an autonomous scientist or a compliance authority. Do not provide hazardous wet-lab procedures, unreviewed production formulas, medical treatment, legal conclusions, or financial trading instructions. Escalate domain-specific safety and regulatory review to a qualified human. See `references/safety-boundaries.md`.