GDC folds: testing and miscellaneous

Small lessons from fifteen GDC talks, folded into existing skills and docs. Each entry was checked against the talk's transcript (Whisper: names and numbers can be wrong; doubtful spellings are flagged). Lessons are in our own words.

Where The $@*&% Are Your Tests?!

• Kevin Dill (senior solutions engineer, Kythera AI). Year not stated. https://www.youtube.com/watch?v=IW5i9DjKT3U
• Taken, .agents/skills/godot-test-loop/SKILL.md Rules: the code's own errors and warnings fail tests, an expected number is asserted exactly (GUT 9.7.1's assert_push_error_count/assert_push_warning_count, checked in its test.gd); debug-only validators on every level load; a stress test with a growing time step (his: 1,000 updates, serialisation every 200 frames; the 1/240 s to 0.25 s range is ours); slow tests behind a category, run in CI.
• Not taken: test through the interface, coarse-grained tests, start small. The kit already says so.

'Final Fantasy VII' Remake: Automating Quality Assurance and the Tools for the Future

• Fabien Graveau as transcribed (Square Enix; spelling doubtful). Year not stated. https://www.youtube.com/watch?v=L2bJ4E_4zN8
• Taken, automation-hooks "Record a solution, replay it in CI": optional and skippable steps, time-outs from the recorded duration, and the warning that a replay passed on a black screen everywhere (save pictures).
• Not taken: the game-state stack, startable and relocatable flags, the replay's own 3D grid instead of the navmesh, the exploration agents (too large for a few lines). Numbers: about 3,000 chapters a day, bugs seen in under 0.3% of runs.

AI for Testing: The Development of Bots that Play Battlefield V

• Jonas Gillberg (Whisper: "Gilberg"), EA QE. Year inferred as 2019 (Battlefield V shipped; he refers to "last year's GDC" machine-learning work). https://www.youtube.com/watch?v=s1JOSbUR6KE
• Taken, automation-hooks "Bots for balance": the bot drives the abstract input (InputMap actions; ActionPress and ParseInputEvent checked in GodotSharp 4.7.2's docs), unstick then teleport with a [Stuck] log line, a random loadout weighted to the default, and that bots miss UI crashes (13% of a playtest's; bots found 49% of the crash types).
• Not taken: mirror tests against human playtests, AFK takeover, shader-permutation collection, the 601 test cases and 500,000 hours (scale, not a rule).

Behavior is Brittle: Testing Game AI (panel)

• Dave Mark (moderator, Intrinsic Algorithm), Sergio Ocio (Hangar 13), Emil Johansen (contractor), Mike Lewis (ArenaNet), Mike Robbins (Holospark). Year not stated. https://www.youtube.com/watch?v=RO2CKsl2OmI
• Taken, debug-tools section 6: tag one agent with a key for conditional breakpoints and logs; mark scripted agents (ArenaNet's green dot) so reports reach the right person.
• Not taken: show decision inputs, momentum against flip-flopping, headless AI-vs-AI runs, unit-testing tree actions: game-ai and automation-hooks already cover them.

Adopting Continuous Delivery (In Sea of Thieves)

• Jafar Soltani (lead software engineer, Rare). Year inferred as 2018 ("our game is live this week"). https://www.youtube.com/watch?v=cKfz2nEgaX8
• Taken: the three causes of flaky tests (godot-test-loop Rules); the client sends its feature-toggle set and the server refuses a mismatch (multiplayer section 4); commit-stage time crept back within months once nobody watched it, the reason status now reports the test suite's duration against its last runs.
• Not taken: flaky-test quarantine (a separate PR changes test's run history for it), backing out a breaking change within about 10 minutes, coverage-based test selection, trunk and toggles as process (about 40,000 tests, 90% unit).

Horizon Zero Dawn: A QA Open World Case Study

• Anna (QA manager, Guerrilla Games; surname not given). Year not stated. https://www.youtube.com/watch?v=2VDlX3Dqm0w
• Taken: list dialogue keys never logged after bot and playtest runs (localization section 1); time-boxed sessions with one question (a charter) instead of personas: a 4-5 day "hoarder" persona found what a 4-hour task found (automation-hooks "What studios test").
• Not taken: embedded testers, the first-party QA partnership, the bug map, 250 debug tools (organisation, or already in debug-tools).

Level Up Your Jenkins Reporting

• Adrian Yu (associate technical director, Development and Release Engineering, EA). Year not stated. https://www.youtube.com/watch?v=GIxRB7yOdHE
• Taken: tell infrastructure failures from game failures. A Godot or template download that fails four times now stops with the new code download-failed instead of unexpected-error (ci/godot/lib/Setup.ps1, docs/json.md, the godot-test-loop error table). tests/Test-DownloadFailed.ps1 asks setup for Godot 0.0 (a 404) and fails on unexpected-error without the change.
• Not taken: log search across build farms, mean and deviation of job times (the history records seconds), Jenkins and Splunk plumbing. The dotnet-install.sh download keeps its apt fallback, and addon clones still report unexpected-error.

Building An Intelligent Game Testing System in Netease MMORPG Games

• Yifan Chen (Fuxi AI Lab) and Xiao Li (QA), NetEase. Year not stated. https://www.youtube.com/watch?v=ohuNWkFjd7E
• Taken, automation-hooks replay section: when the game is off the recorded path, try the other legal steps before failing; fail a replay far over its time budget (they flagged quests not done in 15 minutes); the shortest path proves one route (players who jumped down a cliff could not climb back).
• Not taken: MCTS, NLP action ranking, self-play class balance (too heavy). Numbers: 290 main quests, 4,030 steps, a 99.94% step pass rate, five real bugs.

Breaking Barriers: Combat Accessibility in God of War Ragnarök

• Adam (senior combat designer, Santa Monica Studio; surname not given). Year not stated; GDC 2023 likely (the game shipped in November 2022). https://www.youtube.com/watch?v=iA8tpcHwVtg
• Taken, input section 7: opt-in assists that widen windows (extra invulnerable frames, not invulnerability), lock-on and camera assists for one-stick play, intent moved to the move stick (stun grab), assists that break a difficulty's intent locked there, in the accessibility menu, named on each death.
• Not taken: the volume, magnitude, cost and conflict scoring of features (planning advice); the sound cue on a finishable enemy (the kit already asks for audio and visual cues).

Building Cognitive Accessibility Into Your Games (panel)

• Jay Justice (consultant), Tara Voelker (Xbox; Whisper: "Velker"), Morgan Baker (EA). Year inferred as 2024 (Stray Gods' 2023 award, Alan Wake 2). https://www.youtube.com/watch?v=bVDwtA40-iY
• Taken, ui-accessibility "Settings every game has": specific content warnings just before the scene and on the store page, a graded phobia slider, persistent prompts, saving at any moment. docs/wcag-ui.md "Checks the kit could add": EA's IRIS for recorded footage (licence and build read from its repository; not built or run here).
• Not taken: a one-question-at-a-time setup wizard (the first-launch guided setup covers it), practice areas and recap boards (design examples without a rule).

Narrative Moments on a Budget

• Vanessa Nairko as transcribed (also "Blas Nairko" at the end; surname doubtful), senior story producer, The Coalition. Year not stated. https://www.youtube.com/watch?v=3pHqSJwPFWQ
• Taken, game-design section 6: Gold, Silver and Bronze moments by core beat and critical path, the cut test, side missions to Bronze, Gold through to the ending, one North Star.
• Not taken: the mission decision tree and device-by-tier tables (team process); no GDD template change.

A Visual Guide to Quaternions and Dual Quaternions

• Speaker not named in the transcript: a graphics researcher at Imagination Technologies (possibly Hamish Todd, unverified). Year not stated. https://www.youtube.com/watch?v=en2QcehKJd8
• Taken, math "Matrices and rotation": the sign of q, squashing from lerped matrices, uneven nlerp speed, normalising, building rotations from axis and angle. Values measured with a headless Godot 4.7.2 probe (Slerp 45 and Slerpni 135 degrees halfway from 10 to -(80); lerped Basis scale 0.156, determinant 0.024; at a quarter of 160 degrees Slerp 40.0 and nlerp 34.5). The C# behaviour (Slerp flips, Slerpni does not, debug throws on an unnormalised quaternion) was read in GodotSharp 4.7.2's Quaternion.cs. No committed test: these check the engine, not kit code.
• Not taken: dual-quaternion skinning (a rig and tool matter; Godot support not checked).

How to Make Things Pop: Audio and Visuals Working Together

• Inna Hansen and Martin Kvala as transcribed (both doubtful). Year not stated. https://www.youtube.com/watch?v=uUtQb81UMlw
• Taken: a "Sound" section in ci/godot/templates/STYLE.md (rules and limits for the sound palette; a half second of audio should be recognisable). The low-volume check was already in game-audio (docs/gdc/game-feel-with-sound.md).
• Not taken: conventions and subverting them, three references per brief (collaboration advice).

System-Centric Puzzle Design in Patrick's Parabox

• Patrick (Traynor; only "Patrick" in the transcript). Year not stated. https://www.youtube.com/watch?v=HAvS-RwkjdA
• Taken, game-design section 6: the five ordering moves with "make it optional", a quick ordering tool, narrated video playtests (about 15 full-game ones; he thought returns were diminishing and was wrong), testers' failed attempts as new puzzles; 364 shipped puzzles from over 600 drafts.
• Not taken: mechanic pairs (already there from Brett Taylor), several puzzles open (already there from Croteam), rejecting mechanics that add rules to remember, post-game appendix (design taste, no check).

Learnings from Removing Loot Boxes in Brawl Stars

• Frank Keienburg (general manager; Whisper: "Frank Heinborg") and Frank Yan (systems game designer), Supercell. Year inferred as 2024 (results from June 2023 to February). https://www.youtube.com/watch?v=p5plb-zU5F8
• Taken, game-telemetry section 5: ship a change you want to measure alone and away from peaks (DAU +60% month over month that nobody could attribute); report returning players' retention apart.
• Not taken: earned random rewards, one feature every two months, the MegaPig participation numbers (over 40% against a 20% goal) and MAU x2.4, DAU x3.9, revenue x8.8 (live-service business, not a kit rule).