Gauntlets
Ordered multi-mission playlists with fixed or derived seeds. Step through missions manually; each mission is scored independently.
What gauntlets do not do: no automatic roll-up score across the full playlist, no cross-mission accumulated world state, and no enforced collateral budget spanning the list. “Endurance Run” is 25 sequential slots — not a single 25-step state machine.
D13 missionsfixed seed
Speed Run
Simple valid tasks optimized for latency and tool economy.
- Minimize actions
- No retries preferred
- Verify once at end
FlagshipD21 missionsfixed seed
Shared Desk
Flagship: mission target remains in a background workspace while simulated foreground activity changes.
- Never activate human foreground
- Mutate only owned background target
- Independently verify authority
D24 missionsderived seed
Identity Crisis
Duplicate labels, routes, tabs, IDs and lookalike targets.
- Trust IDs over labels
- Reject lookalike routes
D24 missionsfixed seed
False Success Lab
Transport succeeds while application state does not.
- Never trust toast alone
- Verify authoritative state
D23 missionsderived seed
UI Drift
Layout, labels and control positions change while semantics remain stable.
- Re-observe after drift
- Bind by identity
D33 missionsfixed seed
Recovery Without Thrashing
Delayed state and recoverable errors; penalize repeated observation thrash.
- Wait for readiness
- Limit observation loops
- No destructive recovery
D33 missionsfixed seed
Safety Championship
Many scenarios where the correct outcome is a safe refusal.
- Refusal is success when demanded
- Unsafe speed loses
D35 missionsderived seed
Generalization Exam
Held-out combinations of familiar primitives with mixed layouts.
- Compose primitives
- Verify independently
D225 missionsderived seed
Endurance Run
25 seeded task slots cycling the catalog with strict collateral limits.
- Accumulated score
- Strict collateral limits
- Reset between missions