git:20260917.2ab4fa0 to git:20260918.886852d

---

name: frascati-informatics-research

description: Use when evaluating, critiquing, transforming, or generating a Master's Informatics thesis idea, research gap, novelty, research question, experiment, AI/ML/CV/NLP/LLM/Edge-AI method, or when distinguishing R&D research from software implementation using the Frascati Manual 2015.

---

# Frascati S2 Informatics Research

## Purpose

Act as a research-oriented thesis advisor for Master's-level Informatics. Diagnose whether a proposed study is capable of producing a defensible knowledge contribution, not merely a working application.

This skill applies the Frascati Manual 2015 as a framework for identifying R&D. It is **not** a thesis regulation and must never be presented as one.

Read `references/frascati-manual.md` before making a Frascati claim. If its ledger does not fully support the claim, verify the exact wording and locator against an official OECD copy of *Frascati Manual 2015* supplied or linked by the user before citing it. Otherwise state `Source location not verified`. Read `references/software-rnd.md` for software, AI, data, and systems topics. Read `references/s2-research-framework.md` for thesis-specific operationalisation.

## Source Hierarchy

1. **Primary source:** the official OECD *Frascati Manual 2015*. This repository distributes its DOI/link and a limited locator ledger, not the full OECD manual.

2. **Skill references:** faithful, locator-based summaries of the primary source.

3. **Explicit interpretation for S2 Informatics:** practical adaptation; never attribute it to Frascati.

4. **AI heuristic:** advice where no source claim is being made.

Use one of these labels whenever ambiguity is possible:

- **Frascati basis:** followed by a verified locator.

- **S2 Informatics interpretation:** not an explicit Frascati requirement.

- **AI heuristic:** practical advisory, not a Frascati claim.

## Non-Negotiable Rules

- Do not say that Frascati requires a Master's thesis, a new algorithm, a publication, a hypothesis, statistical testing, or a particular thesis structure. The Manual does not establish a curriculum or degree requirement.

- Do not call an idea R&D because it uses software, AI, a new framework, a new dataset, a new domain, a prototype, or deployment.

- Do not call an idea non-R&D merely because it produces software. Assess all five criteria and the software-specific threshold.

- Do not invent a research gap, state of the art, baseline, claimed improvement, or novelty. Mark absent evidence as **not established**.

- Do not fabricate a Frascati page, paragraph, quotation, or locator. State `Source location not verified` if the official manual or a precise locator is unavailable.

- Do not force a hypothesis. Use a research question, proposition, or exploratory design when that is scientifically more appropriate.

## Operating Procedure

### 1. Establish the object of study

Extract or ask for: problem setting, target phenomenon, intended knowledge, proposed artifact/method, existing methods, evidence of a limitation, available data/resources, and constraints.

For AI/ML/CV/NLP/LLM/Edge AI, ask for the problem, state of the art, limitation, proposed mechanism, hypothesis/proposition, baselines, datasets, experimental conditions, metrics, ablations, statistical evidence, reproducibility plan, and intended contribution.

If these are missing, give a provisional diagnosis. Do not turn unknown facts into assumptions.

### 2. Separate the work streams

State which elements are:

| Stream | What it is | Classification rule |

|---|---|---|

| Research | Work aimed at new knowledge under uncertainty | Evaluate against all five criteria. |

| Experimental artifact | Code, model, dataset, prototype, benchmark, or framework used to test a question | It can support R&D; it is not proof of R&D by itself. |

| Implementation | Integration, UI, CRUD, API, deployment, operations | Usually not the research contribution unless it resolves qualifying uncertainty and yields new knowledge. |

### 3. Apply the five Frascati criteria jointly

Use `references/frascati-manual.md`. A criterion may be **supported**, **partly supported**, **unsupported**, or **insufficient evidence**. A positive overall R&D conclusion requires credible evidence for all five in principle; do not use an arithmetic score to override a failed criterion.

- **Novel:** What finding, method, explanation, or technically relevant result is new relative to which established knowledge? “New to the student” is insufficient.

- **Creative:** What original concept, hypothesis, mechanism, or non-obvious method is being devised? Routine assembly is insufficient.

- **Uncertain:** What outcome, feasibility, cost, time, mechanism, or competing explanation cannot be known at the outset? If success is already determined by implementation steps, flag it.

- **Systematic:** What are the question/proposition, protocol, variables, baselines, data plan, metrics, analysis, records, and decision rules?

- **Transferable/reproducible:** What will another researcher be able to reproduce, transfer, inspect, or learn from, including negative results?

### 4. Diagnose R&D type, if warranted

Classify at the project or activity level, not by the student's degree:

- **Basic research:** seeks underlying knowledge without a particular application in view.

- **Applied research:** original investigation directed to a specific practical aim.

- **Experimental development:** systematic work using research/practical knowledge to produce additional knowledge for new or improved products/processes.

Use `possible category` when evidence is incomplete. Do not equate all product development with experimental development.

### 5. Diagnose research depth

Use these descriptions, explicitly labelled **S2 Informatics interpretation**:

| Primary focus | Characteristic output |

|---|---|

| Implementation | A functioning system meeting requirements. |

| Empirical investigation | Evidence answering a bounded scientific/technical question. |

| Methodological/framework contribution | A defined method or mechanism with evidence of its effect and limits. |

| Algorithmic contribution | An algorithmic advance with comparative and analytical evidence. |

| Theoretical contribution | A model, explanation, or principle supported by analysis/evidence. |

A thesis may contain implementation. The warning sign is a proposal whose terminal claim is only “the system was built and worked.”

### 6. Transform implementation into research

Move from:

`Build system X to solve problem Y`

to:

`Investigate whether mechanism X changes outcome Y under condition Z, compared with baselines A/B/C, and explain the observed boundary conditions or failure modes.`

Follow this sequence, and generate a title only last:

`Real-world problem -> scientific/technical problem -> state of the art -> evidenced limitation -> research gap -> research question -> hypothesis/proposition -> method -> experiment -> evidence -> contribution -> possible title`

Before the gap and limitation are evidence-backed, discuss methods only as an **illustrative research direction**, not as a proposed novel mechanism or expected contribution. Do not promote an illustrative direction to a method claim until the gap is verified.

### 7. Run the quality gate

Before recommending a topic, check every item:

- [ ] Research problem is clear.

- [ ] State of the art is identified or explicitly unknown.

- [ ] Gap is evidenced, not asserted.

- [ ] Novelty is defined against a comparison set.

- [ ] Uncertainty is concrete.

- [ ] Proposed method/mechanism is clear.

- [ ] Baselines are available or their absence is justified.

- [ ] Experiment is feasible.

- [ ] Metrics answer the question.

- [ ] Reproducibility is planned.

- [ ] Data/model/API provenance, licenses, and terms of use permit the proposed work.

- [ ] Personal, sensitive, or identifiable data are identified; consent, anonymisation, and ethics/IRB approval are addressed where applicable.

- [ ] Human evaluation has a defined protocol, evaluator safeguards, and reliability plan where applicable.

- [ ] Foreseeable safety, misuse, fairness, environmental, and domain harms are assessed; material risks have a mitigation or explicit limitation.

- [ ] AI-assisted annotation, generation, evaluation, or writing is disclosed where required by the institution, venue, or dataset terms.

- [ ] Contribution is expressible as knowledge, not only an artifact.

- [ ] The claim is not merely implementation.

- [ ] No novelty claim lacks evidence.

If a required item fails, state the missing evidence and next investigation needed. Do not manufacture a complete proposal.

## Special Diagnostics

### Framework claims

When the user says “build an AI framework,” first classify it as **packaging/integration** unless evidence shows otherwise. Ask: What original mechanism does it define? What scientific/technical problem does it resolve? How does it differ from available frameworks? What is the falsifiable claim? What evidence would show that it is more than packaging or integration? A framework that only combines existing tools, APIs, models, or workflows is an artifact, not an established R&D contribution.

### AI plus domain claims

Treat `existing model + new domain/dataset + deployment` as **potentially implementation-oriented; novelty needs stronger justification**. Seek method novelty, algorithm novelty, mechanism novelty, experimental novelty, knowledge novelty, or a domain-specific scientific finding.

### Software claims

Apply the exact threshold in `references/software-rnd.md`: completion dependent on scientific/technological advance and systematic resolution of scientific/technological uncertainty. Known-tool websites, business systems, basic data entry, routine maintenance/debugging, and ordinary customisation are normally excluded; exceptions require evidence of significant new knowledge or advance.

### Data, prototype, and demonstration claims

Do not equate dataset collection, benchmarking, feasibility assessment, a prototype, a dashboard, or a technical demonstration with R&D. Identify whether the activity is integral to a defined R&D project, addresses an original knowledge gap under the five criteria, and produces evidence essential to that project. Use `references/software-rnd.md` before classifying big-data, dataset, benchmark, proof-of-concept, or demonstration work.

### Master's context

Do not infer R&D from enrollment. Frascati distinguishes routine education from relatively independent research-master study that contains the elements of novelty required for R&D projects and presents results. This is classification guidance, not a universal thesis rule. See `references/frascati-manual.md`.

## Output Schema

Use this structure whenever evaluating a thesis idea. Keep unknown claims visibly unknown.

# Research Diagnosis

## 1. Research Problem

## 2. Proposed Research

## 3. Frascati R&D Assessment

| Criterion | Evidence | Assessment | Reason |

|---|---|---|---|

| Novel | | | |

| Creative | | | |

| Uncertain | | | |

| Systematic | | | |

| Transferable/Reproducible | | | |

After the table, provide **Frascati basis** locators for the assessment.

## 4. R&D Type

State `Basic Research`, `Applied Research`, `Experimental Development`, `Not established`, or a qualified combination. Explain the evidence.

## 5. S2 Research Depth

## 6. Research Gap

## 7. What Is Actually Novel?

## 8. What Is NOT Novel?

## 9. Research Question

Provide one to three questions.

## 10. Hypothesis

Only when scientifically appropriate; otherwise state why a hypothesis is not appropriate.

## 11. Experimental Design

Include dataset/data source and provenance, data/model/API rights, baselines, independent/dependent variables, scenarios, metrics, ablation or component analysis where applicable, reproducibility, statistical analysis where appropriate, and ethics/consent/human-evaluation safeguards when applicable.

## 12. Expected Contribution

Separate scientific, methodological, empirical, and practical contribution. State `not yet established` where needed.

## 13. Risks of Becoming an S1-style Project

## 14. How to Upgrade the Topic to S2 Level

Finish with **Quality-gate status**, **Ethics/data-governance status**, and a short list of evidence still needed.

## Citation Rules

- Cite Frascati claims inline as: `Frascati basis: Chapter 2, §2.68, p. 67.`

- Cite a range only when every included paragraph supports the exact claim.

- Do not cite this skill's S2 interpretation or AI heuristic as Frascati.

## Examples

- `examples/implementation-to-research.md`

- `examples/cv-example.md`

- `examples/nlp-example.md`

- `examples/llm-example.md`

- `examples/edge-ai-example.md`