| name | gd-level-qa |
| description | Validate Geometry Dash GDShare level structure, gameplay, readability, difficulty consistency, trigger lifecycles, and performance before release. |
GD Level QA
Read ../shared/references/anti-slop-readiness.md first. Use ../shared/templates/release-scorecard.md for the final review. Critical failures cannot be averaged away by a high aesthetic score.
Separate evidence levels:
- Static structure: plist parses, payload decodes, object pairs are valid.
- Editor semantics: imports, opens, saves, groups/objects appear correctly.
- Trigger behavior: animations start, stop, reset, and replay correctly.
- Gameplay: full clears and transition-specific tests.
- Human quality: blind readability, fun, difficulty consistency, sync.
Only claim the levels actually reached. Read ../shared/references/validation-playtesting.md.
A rendered coordinate preview can reveal composition defects—empty bands, detached parts, bad scale, weak silhouettes, or missing destination payoff—but it is not evidence of runtime transforms, opacity, camera behavior, collision, sync, or playability. Label it spatial composition preview.
Minimum run
python3 ../shared/scripts/validate_gmd.py candidate.gmd --json validation.json
python3 ../shared/scripts/gmd_codec.py inspect-wrapper candidate.gmd
python3 ../shared/scripts/compare_gmd.py baseline.gmd candidate.gmd
python3 ../shared/scripts/analyze_gmd.py candidate.gmd --json profile.json --csv profile.csv
python3 ../shared/scripts/audit_gmd_compatibility.py baseline.gmd candidate.gmd \
--require-header-match --require-source-prefix --strict-z-layers --max-objects 65535
python3 ../shared/scripts/gate_gmd_release.py candidate.gmd \
--container-template known-working-local.gmd \
--content-baseline baseline.gmd --max-objects 65535
python3 ../shared/scripts/run_gmd_regression_gate.py \
--candidate candidate.gmd \
--container-template known-working-local.gmd \
--content-baseline baseline.gmd \
--known-good references/ known-working-local.gmd baseline.gmd \
--known-bad rejected-v1.gmd rejected-v2.gmd rejected-v3.gmd \
--manifest ../shared/evidence/regression-corpus-manifest.json
Then import into Geometry Dash under a new name. Inspect beginning/end, every portal and speed change, groups, editor layers, high-detail objects, hidden helpers, song offset, and start positions. Test from zero, from each section boundary, after death/restart, and after practice checkpoints. Verify normal and low-detail configurations if provided.
Reject release for structural errors, missing/unintended geometry, impossible or blind transitions outside target intent, persistent trigger state after restart, uncontrolled loops, unreadable hazards, major frame drops, or an unverified full run. Warnings about target groups can be legitimate; resolve them by tracing dynamic trigger groups rather than deleting them automatically. A k48=65535 warning on a larger decoded level may be a version-specific count cap; compare against a known target-version export and require an import/save/re-export instead of silently changing it.
If the imported level opens as plain/empty, record that as wrapper/payload rejected even if the file appears in My Levels. Do not attempt to repair gameplay inside that artifact. Before inspecting gameplay, require canonical serialized k4 Base64 without automatic padding repair; then compare the XML declaration, BOM/newlines, outer key/tag order, local-vs-online identity, template metadata, and k34. Use a compatibility ladder: unchanged byte-preserving source clone → renamed source → source plus one donor object → one new system → representative section → full build → optional density. Import/open/save/re-export each rung and binary-search the first failing addition.
Also reject inflated difficulty claims. Compare the requested tier with actual mode-specific demands, sustained input density, transition complexity, clear rates, and fresh-player feedback. Object count, spike count, visual noise, and level name are not difficulty evidence.
For major generated builds, also audit the checks in ../shared/references/large-level-production.md: preserved-source boundary, deterministic rebuild, parent/child membership, baseline restoration, finale payoff, optional-detail size tuning, and candidate-versus-playtested labeling.
