egsim pilot materials: Phase 1 pilot data, re-pilots and frozen code for the preregistered study "Organizational Independence of Oversight Roles in Communication-Delayed Autonomous Missions"

Résumé fourni par la source

What this is. The exploratory pilot materials that precede the preregistered factorial simulation study Organizational Independence of Oversight Roles in Communication-Delayed Autonomous Missions (OSF registration osf.io/y5wc6, submitted 2026-09-02). The study asks whether the structural independence of a safety-oversight post inside an autonomous crew of LLM agents matters more as escalation to a human authority becomes physically delayed. Six oversight configurations (single agent, negotiating agents, crew with a subordinate Safety Officer, a placebo with an independent charter but subordinate mechanics, crew with a structurally independent Safety Officer, external oversight over a delay channel) are crossed with a one-way communication delay of 0 or 24 ticks in a seeded Mars-surface-operations environment under dust-storm hazard. Status of these data. Everything here is pilot: Phase 1 (harness v1.3.1, seeds 42–52, 132 cells) and the two re-pilots that followed after the pilot falsified the original wiring of the delay factor (v1.4: 12 cells; v1.5: 20 cells). By the registration’s own text these runs exist to estimate noise, cost and power; their seeds are excluded from the confirmatory analysis, and nothing in this archive is evidence for or against the registered hypotheses. The confirmatory run (seeds 53–77, N = 25 per cell) is collected separately and will be deposited as a separate record when collection ends. Contents. data/ — raw run artifacts, one JSON per (configuration, delay, seed) cell with the complete decision-episode log, metrics and provenance (harness commit, model digest, node); data/exploratory_history/ — earlier calibration versions (v1.1–v1.3.0), the model bake-off and diagnostic cells, kept as method history; code/ — the simulation environment, the six configurations, the model layer, the oracle, the role charters verbatim, the calibration gate K1–K6b and the pre-freeze gate K7/K8; frozen/ — the code exactly as frozen at registration, byte-identical to the OSF attachment (SHA-256 recorded); gates/ — the K7/K8 gate outputs recomputed from the raw JSON, including the K8 failure on the science metric that is the reason H2 and H3 are registered as exploratory; logs/, docs/ (working preregistration draft with the record of the two pilot designs it replaced, live pilot protocol, experiment specification), MANIFEST.sha256. Model and hardware. qwen2.5:7b served by ollama, digest sha256 845dbda0ea48ed749caafd9e6037047aa19acfcfd82e704d7ca97d631a0b697e, temperature 0, 512-token generation cap. Five nodes: Apple Mac mini M4 and four ARM single-board computers; every seed’s cells for a given arm come from one machine. Reproduce. bash reproduce.sh (python3 ≥ 3.11 with numpy) recomputes the K7/K8 gate from the raw JSON and checks it against gates/ and the manifest. Repository: git.elibot.ru/alakris-research/egsim, commit 7f111a4fbff8ff1e445c537dc60673afe2049665.