Glossary

One-line definitions for every technical term used in the tutorial, with deep links to the deep dive that explains it in context. If you hit an unfamiliar name in a chapter, this is the fastest way to recover.

Entries are organized alphabetically. Most are followed by a short link of the form → Chapter X §Y that points to the deep dive, sidebar, or appendix where the term is properly introduced.

A

• Action-effect verification — After an entity-affecting action (a build/placement or a camera move), the agent re-detects and confirms the expected change, recording CONFIRMED … or the exact phrase no visible change; repeated misses feed the stuck-loop detector. → Chapter 2 §2.1 Step 11.
• Anchor box — A YOLO-style detector predicts each object’s bounding box as an offset from a pre-defined anchor of a given aspect ratio, rather than predicting raw coordinates. → Appendix A — YOLO and object detection.
• Asyncio — Python’s cooperative-multitasking concurrency primitive. The agent’s game loop is built on a single asyncio event loop so the strategist can think in the background while the executor is acting. → Chapter 1 §Async-first architecture.

B

• Bradley-Terry — A statistical model that turns pairwise win/loss outcomes into a single skill rating per competitor by maximum likelihood. → Appendix C — Bradley-Terry MLE and bootstrap CIs.
• Backpressure — When a producer is faster than its consumer, backpressure is the mechanism that slows the producer (or drops events) so memory doesn’t blow up. → Chapter 15 §Backpressure semantics.
• Bootstrap (statistical) — A resampling technique that estimates the uncertainty of a statistic by repeatedly sampling with replacement from the observed data. We use it to put confidence intervals on Bradley-Terry skill estimates. → Appendix C.

C

• Cache hit / cache miss (LLM) — Whether the provider was able to reuse its internal KV-cache for a prefix of your request. Visible in usage.cache_read_input_tokens. → Chapter 5 — Prompt caching callout.
• cache_control — Anthropic’s marker for “this block of the prompt can be cached.” → Chapter 5 — Prompt caching callout.
• COCO format — A JSON-based bounding-box / segmentation label format with explicit category IDs. The default format CVAT exports. → Chapter 9 — COCO vs YOLO callout.
• Columnar storage — Storing tables one column at a time instead of one row at a time. Massively faster for analytical scans because you only read the columns you query. The basis of DuckDB, Parquet, and ClickHouse. → Chapter 16 — DuckDB internals.
• Composite tool — In our agent, a tool that internally chains several primitive actions (press, click, rescan) into one API roundtrip. → Chapter 5 §5.3.
• Confidence interval (CI) — A range estimate around a statistic that contains the true value with a stated probability under repeated sampling. → Appendix C.

D

• DuckDB — An embedded analytical (OLAP) database. Like SQLite for analytics: zero-config, file-based, columnar, vectorized. We use it as the cold-path event log. → Chapter 16.
• DXT1 / BC1 — A GPU-friendly block-compression texture format that encodes every 4×4 pixel block into 8 bytes by storing two endpoint colors plus a 2-bit index per pixel. AoE2’s SLD sprite format uses it. → Chapter 11 — DXT1 deep dive.
• Determinism (LLM context) — Whether the same prompt yields the same output. Surprisingly hard with LLMs: even temperature=0 plus a fixed seed isn’t fully deterministic across hardware due to BF16 floating-point order. → Chapter 18 — Determinism deep dive.

E

• Event broker — A piece of infrastructure that decouples event producers from consumers via pub/sub, queues, or streams. → Appendix B — Event brokers and Redis Streams.

H

• Hungarian algorithm — A polynomial-time algorithm for assigning N predicted tracks to N detected entities optimally given a cost matrix (typically IoU or pixel distance). The standard pairing step in multi-object trackers. → Chapter 2 — Kalman + Hungarian deep dive.

I

• IoU (Intersection over Union) — A similarity score for two boxes: area(A ∩ B) / area(A ∪ B). Used everywhere in detection — for NMS, for matching tracks to detections, for measuring mAP. → Appendix A.

K

• Kalman filter — A recursive estimator that maintains a state distribution (mean + covariance) over time, predicting forward each step and correcting against noisy observations. We use a constant-velocity Kalman per tracked entity. → Chapter 2 — Kalman deep dive.

M

• mAP / mAP50 / mAP50-95 — Mean Average Precision: the standard detection metric. mAP50 averages precision across classes at IoU ≥ 0.5; mAP50-95 averages across 10 IoU thresholds from 0.5 to 0.95 in 0.05 steps. → Appendix A.
• MLE (Maximum Likelihood Estimation) — Choosing parameter values that make the observed data most probable under the model. Bradley-Terry skills are fit by MLE. → Appendix C.
• MM algorithm (Minorization–Maximization) — The iterative fitting algorithm we use for Bradley-Terry. Each step bounds the log-likelihood from below by a simpler surrogate, then maximizes the surrogate. Hunter (2004). → Appendix C.

N

• NMS (Non-Maximum Suppression) — The post-processing step that takes a detector’s raw box predictions, sorts by confidence, and greedily drops any box whose IoU with a higher-confidence box exceeds a threshold. → Appendix A.

O

• OLAP vs OLTP — OLAP (analytical) workloads read lots of rows and few columns to compute aggregates; OLTP (transactional) workloads read/write a few rows touching many columns. DuckDB is OLAP; Postgres/SQLite are primarily OLTP. → Chapter 16.
• OPRO — Optimization by PROmpting: a paper from DeepMind where one LLM proposes new prompts and another LLM scores them, iteratively converging on better prompts. Related to our autoresearch loop. → Chapter 22 — Prompt optimization landscape.

P

• Pareto frontier — The set of candidates that are non-dominated: none is beaten on every score component at once. The autoresearch tournament keeps a frontier over the five score components so a candidate strong on a single axis isn’t culled for a weak composite. → Chapter 22 — How a tournament runs.
• Prompt caching — Provider-side reuse of the KV-cache for a stable prefix of your prompt. → Chapter 5 — Prompt caching callout.

R

• Reactive tier — A deterministic, no-LLM layer that runs routine economy upkeep every turn (queue villagers below the age cap, reassign idle villagers to the nearest resource); it cedes combat to the LLM by returning nothing on alarm. → Chapter 2 §2.1 Step 9.
• Redis Streams — A Redis data structure (XADD / XREAD) that behaves like a durable, replayable, consumer-group-aware log. Our cross-process broker backend. → Appendix B.
• Reflective prompt mutation — Proposing prompt edits by reasoning over full game traces (turn-by-turn reasoning, actions, verification) plus the per-component score breakdown, rather than over summary metrics alone. → Chapter 23 — The mutator.
• RTC (turn pipelining) — Request-to-completion overlap: a routine turn launches the next executor plan as a background task and executes the previous turn’s committed head while it computes, hiding the LLM roundtrip. → Chapter 2 §2.1 Step 9.

S

• SAHI (Slicing Aided Hyper Inference) — A technique for detecting small objects in large images by tiling the input, running the detector on each tile, and stitching results. Implemented (including an adaptive variant that only tiles likely regions) but disabled in v6 (adaptive_sahi=False): tiling a retina screenshot into 640 crops shows the model objects ~2.4× larger than its training scale, which lowers real F1, so the agent runs a single pass at imgsz=640 instead. → Chapter 7 §7.4.
• SSE (Server-Sent Events) — A one-way HTTP streaming protocol where the server pushes text/event-stream chunks to the client. Simpler than WebSocket for fan-out telemetry. → Chapter 19 — SSE callout.
• Structured output — Constraining an LLM to emit a parseable, schema-validated response (typically JSON). → Chapter 5 — Structured output callout.

T

• Tool use / function calling — An API pattern where the LLM is given a set of “tools” with JSON-schema arguments; the model emits a tool call and the host runs it and feeds back the result. → Chapter 4 — Agentic tool loops deep dive.

V

• Vectorized execution — Processing a batch (vector) of rows at a time through each operator, instead of one row at a time. The reason DuckDB and modern OLAP engines are 10–100× faster than row-at-a-time engines on analytical queries. → Chapter 16.

Y

• YOLO — You Only Look Once: a family of single-shot object detectors that predict all boxes in one forward pass. We use YOLO26n (NMS-free). → Appendix A.