35 terms, 199 links to the work carrying them. The number on a term is how many pieces of work it points to.
Discipline
What the work is about. 13 terms.
AI evaluation 14
Deciding whether a model or agent actually did what it claimed.
Also called model evaluation, agent evaluation, LLM evaluation, LLM-as-judge.
AI security 16
Attacks on what an agent reads, stores, and later trusts.
Also called red teaming, AI red teaming, adversarial testing, prompt injection, responsible disclosure, vulnerability research.
Benchmarking 5
Measuring a system under conditions stated well enough to reproduce.
Also called performance testing, performance engineering.
Data modeling 20
Representing a domain faithfully, so the structure survives contact with the data.
Also called schema design, normalization, dimensional modeling.
Data quality 9
Finding what is wrong in source data, and proving it is wrong.
Also called data validation, data profiling.
Entity resolution 3
Deciding whether two records describe the same thing, when nothing says so.
Also called record linkage, deduplication, fuzzy matching.
Human agency 8
What people give up to a system, and what it costs them.
Metadata architecture 3
Structures that describe other data — catalogs, authority files, controlled vocabularies — so a fact about a source is recorded once and can be checked.
Also called authority control, master data management, taxonomy, data catalog.
Migration 3
Moving a working system to a new platform without stopping its users.
Also called replatforming, platform migration.
RLHF 2
Measuring how RLHF alignment and status bias shift agent decision routing under pressure.
Storytelling 10
An analogy from lived experience carrying a technical idea, so the reader arrives at the mechanism through something they already know.
Teaching 17
Course material or explanatory writing built so a learner arrives at the answer rather than watching it appear.
Team leadership 4
Running the people doing the work, and the mechanism that checks it.
Also called technical leadership, mentoring.
Method
How a claim was established, rather than what it was about. 9 terms.
Ablation study 3
Running the same goals with one thing changed, to isolate what it costs.
Cold-cache measurement 4
Clearing every cache before each run, so the engine is measured and not the buffer pool.
Content addressing 4
Fingerprinting input configurations and query state via SHA-256 cryptographic hashes.
Data contracts 4
Enforcing reproducible experimental parameters and data schemas via immutable YAML specifications.
Also called schema enforcement, data validation contracts.
Deterministic grading 6
Scoring outcomes by rule, not by judgement, so a result can be re-derived.
Also called rule-based scoring, automated grading.
Evaluation sets 5
A labelled set of cases, built so a system's accuracy is measurable.
Also called ground truth, labelled data, test sets.
Per-turn tracing 3
Recording every step an agent took, rather than asking it afterwards.
Also called execution tracing, agent traces.
Pre-commit gates 5
Enforcing automated hooks, claim verifiers, and privacy rules before code commits.
Also called security automation, policy as code, guardrails, shift-left.
Process telemetry 3
Deriving quantitative behavioral metrics directly from execution logs without judgment.
Also called observability, behavioural metrics.
Technology
What it was built with. 13 terms.