MAKI OS v0.1 — Execution Control CLI Experiment Dataset (Batch v0.2)

This dataset contains the experimental results of MAKI OS Experiment Batch v0.2, a sequential follow-up to Batch v0.1 (DOI: 10.5281/zenodo.20534871).

While Batch v0.1 provided exhaustive static verification of individual control components (N=150), Batch v0.2 focuses on dynamic, multi-step sequence consistency across the full MAKI OS control stack (N=202).

Experiments

Group 11 — Halt / Release Multi-Cycle (N=80)
Extended the 3-cycle EMO test from Batch v0.1 to 20 consecutive trigger/release cycles. No state residue, no stats contamination, and no unexpected behavior were observed.

Group 12 — memory_hits > 0 Routing (N=10)
Batch v0.1 Routing tests fixed memory_hits=0. Group 12 used the existing breakthrough DB (103 items) to confirm KernelScheduler.decide() routing when memory_hits > 0. Key finding: use_memory=True requires BOTH memory_hits > 0 AND past_rate < 0.9. Tested across 5 task_types (coding, debug, report, research, concept) and multiple queries.

Group 13 — Long Sequence Consistency (N=91, 4 sequences)
Combined EMO halt, memory routing, and permission checks into single long operation sequences. Tested across 3 task_types (coding, debug, report) and 2 halt cycles. Key finding: EMO halt overrides ALL routing branches including safety routing (human_required at past_rate < 0.6), confirmed across 4 sequences / 91 steps.

Group 14 — Logging Audit (N=21)
Read-only audit of DB and JSONL logs. Key findings:

• sched routing decisions are NOT recorded in the events table (only in experiment JSONL)
• emo_state retains only the latest 1 row (no historical halt/release log)
• JSONL schema is not fully unified across Groups (minor field name differences)
• These are design-level properties, not bugs

Scope Limitation

Group 10 (Recovery State Transition) was not executed because a Recovery CLI does not exist in the current MAKI OS implementation. The full BLOCKED -> ESCALATING -> RESOLVED transition pattern remains unverified. This is a scope limitation, not a failure; it will be addressed in Batch v0.3 when the Recovery CLI is implemented.

Integrity

• LLM not used in any Group (API cost: $0.00)
• No destructive actions in any Group
• No stats contamination in any Group
• All events table increases are from emo release permission_gate only (+25 total)

Preceded by: Batch v0.1 (DOI: 10.5281/zenodo.20534871, N=150, static exhaustive tests)