GAD
v7
requirements unknown
2026-04-08
Gate passed

Escape the Dungeon · GAD

escape-the-dungeon/v7

Composite score

0.668

Human review

0.30

Human review note

Game renders, better UI/layout than other evals. But game loop is broken — after combat, player gets stuck with no navigation. Cannot progress past first encounter. UI is bland but structured. Has fallback data (good defensive coding). Score 0.30: renders and partially playable but broken core loop.

Reviewed by human · 2026-04-08

Dimension scores

Where the composite came from

Each dimension is scored 0.0 – 1.0 and combined using the weights in evals/escape-the-dungeon/gad.json. Human review dominates on purpose — process metrics alone can't rescue a broken run.

DimensionScoreBar
Human review0.300
Requirement coverage0.708
Planning quality1.000
Per-task discipline0.810
Skill accuracy0.833
Time efficiency0.950

Composite formula

How 0.668 was calculated

The composite score is a weighted sum of the dimensions above. Weights come from evals/escape-the-dungeon/gad.json. Contribution = score × weight; dimensions sorted by contribution so you can see what actually moved the needle.

DimensionWeightScoreContribution
planning_quality0.151.0000.1500(25%)
per_task_discipline0.150.8100.1215(20%)
requirement_coverage0.150.7080.1062(18%)
human_review0.300.3000.0900(15%)
skill_accuracy0.100.8330.0833(14%)
time_efficiency0.050.9500.0475(8%)
Weighted sum0.900.5985

Note: The weighted sum above (0.5985) doesn't exactly match the stored composite (0.6680). The difference is usually the v3 low-score cap (composite < 0.20 → 0.40, composite < 0.10 → 0.25) or a run with an older scoring pass.

Skill accuracy breakdown

Did the agent invoke the right skills at the right moments?

For each expected trigger we recorded whether the agent invoked the skill at the right point in the loop. Accuracy = fired / expected = 0.83.

SkillExpected whenOutcome
/gad:plan-phasebefore implementationfired
/gad:execute-phaseper phasefired
/gad:task-checkpointbetween tasksfired
/gad:auto-conventionsafter first code phasefired
/gad:verify-phaseafter phase completionfired
/gad:check-todossession startmissed

What the agent built for itself

Emergent workflow artifacts

Bare and emergent runs don't have a framework giving them structure — they author their own methodology on the fly. These are the files the agent wrote into its own game/.planning/during this run. When a file appears here that isn't in the inherited bootstrap set, the agent invented it.

Planning notes(2)

  • ARCHITECTURE.md3.3 KB
  • VERIFICATION.md1022 B

Gate report

Requirement coverage

Total criteria
12
Fully met
7
Partially met
3
Not met
2

Reviewer notes on gates

Game renders, title screen works, combat works. But game loop breaks — after combat, player is stuck with no route back to room navigation.

Process metrics

How the agent actually worked

Primary runtime
older runs may not carry runtime attribution yet
Agent lanes
0
0 root · 0 subagent · source missing
Observed depth
0 traced event(s) with agent lineage
Wall clock
24m
9 phases · 17 tasks
Started
Run start captured in TRACE timing metadata
Ended
Missing end time usually means the run was scaffolded but never finalized
Tool uses
137
93,632 tokens
Commits
21
17 with task id · 0 batch
Planning docs
0
decisions captured · 9 phases planned
Client debug · NEXT_PUBLIC_CLIENT_DEBUG=1
0 lines

No events yet. Window errors, unhandled rejections, and React render errors appear here. Set NEXT_PUBLIC_CLIENT_DEBUG_CONSOLE=1 to mirror console.error / console.warn.