📖 PUMA Vault — Master Glossary
Overview
All technical terms used in the PUMA project and this vault, organised alphabetically. Total terms: 65+. All definitions in English.
A
Agent OS — An operating system-like framework for orchestrating multiple AI agents, managing their context, tool access, memory, and coordination. In PUMA’s Stage 5 vision, LangGraph + OPA forms the Agent OS layer. See Smart-PMO-Vision.
AI-Driven Concept Mapping — Automatically generating concept maps from text or notes using AI semantic analysis. Used in this vault via Obsidian’s graph view combined with Smart Connections plugin.
AMI (Autodiagnóstico y Mejora Iterativa) — A prompting methodology where the researcher presents work to an AI critic, receives structured diagnosis, and iterates — retaining full authorship. Core technique in CDD. See PN-AMI-DRCA-IIPR-Frameworks.
Agentic AI — AI agents capable of executing complex, multi-step tasks autonomously with access to tools, memory, and decision-making capability. PUMA’s Stage 4–5 implements agentic patterns. Contrast with single-turn LLM inference.
AnythingLLM — A local RAG platform running privately on hardware, supporting Ollama models and vector stores (ChromaDB, Qdrant). Used in PUMA’s AI research phase for private document Q&A.
B
Backlog — The ordered list of all work items (user stories, bugs, features) for a software product, managed in tools like Jira or Linear. PUMA Stage 3 evaluates LLM backlog prioritisation.
BDD (Behaviour-Driven Development) — A development approach specifying desired behaviour in Given/When/Then scenarios before implementation. Related to SDD. See PN-SDD-Framework.
Benchmark — A standardised evaluation procedure with defined datasets, metrics, and reproducible conditions enabling comparison across systems and time. PUMA is a benchmark for PM+LLM evaluation. See MOC-LLM-Benchmarks-PM-AI.
BMAD (Breakthrough Method for Agile AI-Driven Development) — A multi-agent framework simulating a full agile team: Analyst, Product Manager, Architect, Developer, QA, Sustainability Officer. Each agent is defined in Markdown/YAML. See BMAD-Agent-Roster.
C
CDD (Context-Driven Development) — A development approach where all AI interactions and tool usage are informed by rich, explicit context about the project, constraints, and goals. Every PUMA prompt uses CDD principles via RCOIF framing.
Chain-of-Thought (CoT) — A prompting technique instructing LLMs to reason step-by-step before producing a final answer. One of PUMA’s four prompting strategies (zero-shot, few-shot-3, few-shot-6, CoT). See PN-CoT-FewShot-Prompting.
ChromaDB — An open-source, embeddings-focused vector database used for local RAG systems. Simple to set up; used in PUMA Stage 4 (optional RAG agent).
CodeCarbon — A Python library tracking and estimating CO₂ emissions from computational workloads. Integrated in PUMA as a first-class experimental variable. GitHub: mlco2/codecarbon. See Carbon-Tracking-Log.
CoGEE (Code and GPT for Effort Estimation) — (Tawosi et al., 2024). State-of-the-art story point estimation using LLMs, achieving MAE ~1.9 SP with GPT-4. Key PUMA baseline for Stage 2.
Cognitive Offloading — Delegating critical thinking to AI tools, risking atrophied reasoning skills. The Keshav Three-Pass Method and DRCA explicitly guard against this. See PN-MIT-Student-Method.
Contextual Anchoring — A prompting technique where core constraints and context are restated at the end of a long prompt to prevent model drift from key requirements. Used in all PUMA triage prompts.
CPM (Critical Path Method) — An algorithm identifying tasks where any delay would compromise a project milestone. Relevant to PUMA’s Smart PMO Stage 5.
D
DRCA (Deconstrucción y Reconstrucción Conceptual Avanzada) — Four-step cognitive procedure for deeply processing complex concepts: Deconstruct → Reconstruct → Critical → Advance. Combined with Keshav Pass-3 for deep paper analysis. See PN-AMI-DRCA-IIPR-Frameworks.
DSR (Design Science Research) — A research paradigm requiring the production and evaluation of useful artefacts. PUMA follows DSR: the benchmark framework is the artefact, evaluated by utility, novelty, and rigour. See PN-DSR-SLR-Methods.
Dataview — An Obsidian plugin that queries the vault as a database. Used throughout this vault for dynamic dashboards, literature tracking, and task views.
E
EBSE (Evidence-Based Software Engineering) — Applying systematic evidence synthesis (SLR, meta-analysis) to SE research and practice decisions. The epistemological foundation of PUMA’s SLR.
Effect size (r) — A standardised measure of the magnitude of a statistical effect, independent of sample size. In PUMA: r = Z / √N from Wilcoxon test. r ≥ 0.1 = small, ≥ 0.3 = medium, ≥ 0.5 = large.
EGI (Exploración Guiada Interactiva) — A multi-turn AI interaction pattern for mapping unfamiliar domains through structured panoramic → drilling → gap identification sequences. Used during Keshav Pass-1 and literature exploration. See PN-EGI-Framework.
Embeddings — Dense vector representations of text capturing semantic meaning. Similar texts produce similar vectors. Used in RAG systems and semantic note-taking. See PN-RAG-Embeddings-VectorDB.
F
F1-macro — The mean F1 score calculated independently for each class, then averaged. Treats all classes equally regardless of frequency. Primary metric for PUMA Stage 1 (triage). Minimum threshold: ≥ 0.55.
Few-Shot Prompting — A prompting strategy providing k labelled examples within the prompt context, enabling in-context learning without fine-tuning. PUMA tests few-shot-3 and few-shot-6 variants. See PN-CoT-FewShot-Prompting.
FINER criteria — Framework for evaluating research question quality: Feasible, Interesting, Novel, Ethical, Relevant. PUMA’s research question satisfies all five. See PN-DSR-SLR-Methods.
Fine-Tuning — Adjusting a pre-trained LLM on domain-specific data to improve performance on a target task. Excluded from PUMA MVP scope (cost, reproducibility risks).
Fibonacci scale — The non-linear scale used for story point estimation: {1, 2, 3, 5, 8, 13, 21}. Non-linearity reflects increasing uncertainty for larger work items.
G
gCO₂eq — Grams of CO₂ equivalent. The unit used by CodeCarbon for measuring computational carbon footprint. PUMA targets the first PM+LLM study to report this per task.
GTD (Getting Things Done) — David Allen’s productivity system: capture all commitments, clarify next actions, conduct regular reviews. Implemented in 90 - GTD/ folders. See PN-PARA-GTD-Zettelkasten.
H
Human-in-the-Loop (HITL) — A design principle where AI systems produce recommendations requiring explicit human validation before action. All PUMA agent outputs are HITL by design. Encoded in SP-PUMA-Constitution.
I
IIPR (Ingeniería Inversa de Prompts y Respuestas) — Reverse prompt engineering: diagnosing why a prompt produced poor output and systematically redesigning it. Used in PUMA prompt optimisation. See PN-AMI-DRCA-IIPR-Frameworks.
Issue Triage — Reviewing and classifying work items (priority, severity, component) in PM tools. Stage 1 of the PUMA benchmark. Dataset: Jira SR.
Iterative Refinement Loop — A cycle where AI improves a text or idea through multiple guided iterations. Core technique in AMI methodology.
ITIL 4 — IT Infrastructure Library v4. A framework for IT service management referenced in PUMA’s PM context.
J
Jira SR (Jira Social Repository) — Public dataset of 50,000+ real Jira issues from Apache Software Foundation projects with manually assigned priority labels. Primary dataset for PUMA Stage 1. DOI: 10.5281/zenodo.5901893.
Johnny Decimal — A filing system using two-digit area codes (10–90) and decimal sub-codes giving every note a unique, memorable address. Used for vault folder organisation. See JD-Master-Index.
K
Keshav Three-Pass Method — A verified academic paper-reading framework by S. Keshav (University of Waterloo, 2007). Three passes of increasing depth: bird’s-eye (5–10 min), content grasp (≤1h), virtual re-implementation (4–5h). Used as the PUMA literature review protocol. See PN-MIT-Student-Method.
L
LangGraph — A library from LangChain for building stateful, multi-actor AI agent workflows as directed graphs with cycles. Core orchestration technology for PUMA Stage 4–5.
LlamaIndex — A data framework connecting LLMs to external data sources, enabling RAG and structured querying. Used in PUMA Stage 4 for semantic indexing of historical issues.
LLM (Large Language Model) — A neural network trained on vast text corpora capable of generating, summarising, classifying, and reasoning over text. PUMA evaluates Llama 3.2 8B and Mistral 7B.
LLM-Based Literature Review — Using AI to analyse large volumes of papers and extract patterns, gaps, and evidence tables. Used in PUMA SLR phase with Perplexity, Consensus, Elicit.
LM Studio — A desktop application for discovering, downloading, and running open-source LLMs locally.
M
MAE (Mean Absolute Error) — Primary metric for PUMA Stage 2: MAE = mean(|predicted_SP − actual_SP|). Lower is better. Minimum threshold: ≤ 3.0 SP. Baseline: mean historical SP per project.
Marco RCOIF — See RCOIF.
MIT Student Method — See Keshav Three-Pass Method. Note: “MIT Student Method” as a formal 3-question methodology is not a verified published framework. The vault uses Keshav’s Three-Pass as the verified equivalent.
MOC (Map of Content) — A note serving as a structured index and navigation hub for a cluster of related notes. Higher-level than tags, more flexible than folders. See 80 - MOC/.
MTTD (Mean Time To Detection) — A PM metric measuring average time between issue occurrence and detection. Related to triage efficiency. PUMA Stage 5 target: 30% MTTD reduction vs. manual.
Multi-Agent Collaboration — Using multiple specialised AI agents that work together on a project. PUMA’s Smart PMO vision (Stage 5) and BMAD framework both implement this pattern.
O
Obsidian — A Markdown-based note-taking application creating interconnected knowledge networks via bidirectional links, tags, and plugins. This vault runs on Obsidian.
OPA (Open Policy Agent) — A general-purpose policy engine enforcing governance rules on AI agent actions. Used in PUMA Stage 5 for agent action validation.
Ollama — An open-source tool for running LLMs locally via a REST API. Primary inference engine for PUMA experiments. See LN-Tools-Ollama-ClaudeCode-OpenCode-BrowserOS.
One-Shot and Few-Shot Prompting — See Few-Shot Prompting.
OpenSpec — A YAML/Markdown-based specification format for SDD, maintaining a single evolving specification document. Used in PUMA for brownfield/iterative spec management. See PN-SDD-Framework.
P
PARA (Projects, Areas, Resources, Archive) — A four-category organisation system by Tiago Forte, integrated into this vault’s folder structure (10-→70-). See PN-PARA-GTD-Zettelkasten.
PMBOK 7 — Project Management Body of Knowledge, 7th edition (PMI, 2021). Value-delivery focused framework referenced in PUMA’s PM context.
PRISMA 2020 — Preferred Reporting Items for Systematic Reviews and Meta-Analyses, 2020 update. Protocol used for PUMA SLR. See PRISMA-Log.
PRISMA-trAIce — An extension of PRISMA documenting the “AI trace” — which tasks were AI-assisted, with which models, with what validation. Used in PUMA’s AI use declaration.
Prompt Engineering — The practice of designing, refining, and optimising prompts to elicit desired behaviour from LLMs. PUMA’s independent variable: four strategies (zero-shot, few-shot-3, few-shot-6, CoT).
PUMA (Platform for Understanding and Management with Agents) — The PUMA Project this vault supports. Full title: “Can language models manage ICT projects? Empirical benchmark of local LLM agents for issue triage and effort estimation.” See PR-PUMA-Ch1-Introduction.
Q
Quantization (4-bit, Q4_K_M) — Reducing LLM memory requirements by representing weights in lower precision. Enables ~8B parameter models to run in ~5GB RAM without significant quality loss.
Qdrant — A high-performance vector database for similarity search, suitable for local RAG systems. Used in PUMA Stage 4 (optional).
R
RAG (Retrieval-Augmented Generation) — Combining LLM generation with retrieval from an external knowledge base. Used in PUMA Stage 4 (optional): agent retrieves similar historical issues before classifying. See PN-RAG-Embeddings-VectorDB.
RCOIF (Role, Context, Objective, Instructions, Format) — A structured prompting framework for AI-assisted research. Every significant PUMA AI interaction uses this structure. See PN-RCOIF-Framework.
RCF (Reference Class Forecasting) — A technique using distributions from comparable past projects to correct optimism bias in estimates. Flyvbjerg’s antidote to the Uniqueness Trap. Relevant to PUMA’s estimation module.
Red Teaming — Actively searching for evidence against your own conclusions to test their robustness. Core component of the PUMA QA Agent role (Keshav Pass-3 equivalent). See PN-KeyConcepts-Agents-Reproducibility-RedTeam.
RMSE (Root Mean Square Error) — Secondary metric for effort estimation: sqrt(mean((predicted−actual)²)). More sensitive to large errors than MAE.
S
SDD (Spec-Driven Development) — An engineering approach where executable specifications are written before code; specs are the source of truth. See PN-SDD-Framework.
Semantic Note-Taking — Taking notes connected via conceptual relationships detected by AI. Implemented via Smart Connections plugin + Zettelkasten permanent notes.
SLR (Systematic Literature Review) — A reproducible, explicit methodology for synthesising research literature, following a pre-registered protocol with defined inclusion/exclusion criteria. See WF-SLR-Pipeline.
Smart PMO — A Project Management Office augmented by AI agents performing automated triage, estimation, risk detection, and reporting under human governance. PUMA’s Stage 5 vision. See Smart-PMO-Vision.
Spec Kit — GitHub’s open-source toolkit formalising SDD with a 4-phase gated workflow: Specify → Plan → Tasks → Implement. Each phase requires explicit human approval. See PN-SDD-Framework.
Sprint — A time-boxed Agile iteration (typically 2 weeks) where a defined set of work items is completed. PUMA uses sprints aligned to PEC deliverables.
Story Points (SP) — A dimensionless unit estimating relative effort, complexity, and risk in Agile. Follows Fibonacci scale: {1, 2, 3, 5, 8, 13, 21}. PUMA Stage 2 evaluates LLM story point estimation.
T
TAWOS (The Agile Workflow Optimisation Suite) — Dataset of 23,000+ real Agile user stories with story point estimates (Tawosi et al., 2022). Primary dataset for PUMA Stage 2. See LN-Datasets-JiraSR-TAWOS.
Templater — An Obsidian plugin for advanced templates with JavaScript, date variables, and user prompts. Used for all PUMA note templates.
temperature=0 — An LLM parameter setting making output deterministic (greedy decoding). Used throughout PUMA for reproducibility.
Token Tax — The computational and financial cost of tokens consumed per task. PUMA tracks this as a practical efficiency metric (target: < 0 for local).
Triage — See Issue Triage.
U
Uniqueness Trap — The documented tendency of project managers to treat each project as absolutely unique, preventing statistical learning from historical patterns (Flyvbjerg & Gardner, 2023). One of PUMA’s three motivating problems.
V
Vector Databases — Databases storing embeddings for intelligent similarity search. Enable RAG retrieval. Examples: ChromaDB (simple), Qdrant (production), Weaviate. See PN-RAG-Embeddings-VectorDB.
W
Wilcoxon signed-rank test — A non-parametric statistical test comparing two paired samples. Used in PUMA to test H1 and H2 significance (α = 0.05, two-sided). See PN-Wilcoxon-FINER-Cornell-PRISMA.
Z
Zettelkasten — A knowledge management method using atomic, permanently located notes linked by explicit connections. Implemented in this vault’s 30 - Permanent/ folder (flat structure, declarative titles). See PN-PARA-GTD-Zettelkasten.
Zero-Shot CoT — Chain-of-thought prompting without any examples, using a trigger phrase like “Think step by step:” to elicit reasoning. One of PUMA’s four prompting strategies (Strategy 1).
*Glossary v2.0 · 65 terms · Updated April 2026