YuvrajSingh9886/androidlife-public

DailyBench-500 — Public run trajectories

Full per-task execution artifacts (agent logs, LLM proxy metrics, ADB preflight/postflight, telemetry, trajectories with UI states + screenshots) from the public DailyBench-500 runs against a real Android device (OnePlus CPH2423) via ADB/mobilerun and real LLMs on OpenRouter.

The task corpus itself is published separately as YuvrajSingh9886/drainbench-530 and the 61-task public sample as YuvrajSingh9886/dailybench-public-sample. This repo contains the full per-run artifacts for full disclosure:

• runs/<run_id>/ — raw execution artifacts (agent logs, LLM proxy metrics, telemetry, per-step GUI evidence, trajectories).
• reports/public/ — per-run narrative reports (manual audit, verdicts, on-device verification, privacy scan).
• reports/metrics/public/ — per-run official metrics (.json/.md) + manual-audit notes (harness self-report is reference; manual audit is ground truth).
• reports/metrics/hallucination/public/ — per-run full-context agent-log hallucination-control judge outputs (with token/cost telemetry) + README.
• reports/turn-based/public/ — per-turn ASK USER audits (question → answer).

Runs

Each runs/<run_id>/ folder is one full benchmark run (60 tasks = 20/day × 3 days, unless noted). Per-task folder layout:

File	Content
agent.log.txt	Full agent step log (mobilerun FastAgent).
output.json / output.txt	Graded outcome (success + reason) and final answer.
meta.json	Run metadata (task_id, model, command, exit code).
preflight.json / postflight.json	Device pre/post checks.
llm_proxy_metrics.jsonl	Per-LLM-call tokens + cost.
ask_user_metrics.jsonl	ask_user tool calls (ASK USER tasks).
samples.ndjson	Battery/thermal sampling during the task.
trajectories/<ts>/	macro.json, trajectory.json, ui_states/, screenshots/ — per-step GUI evidence.
kb_audit.json (run root)	Manual multi-turn KB audit (KBIQ) where performed.

Run index (self-reported harness outcome; manual audit is ground truth)

Run ID	Model (OpenRouter)	Tasks	Pass
20260826-105200	google/gemini-3.1-flash-lite	60	42
2026-08-26-184934	qwen/qwen3.8-27b	60	25
2026-08-28-002424	qwen/qwen3.8-27b	60	33
2026-08-29-153657	moonshotai/kimi-k2.6	60	31
2026-08-30-021852	moonshotai/kimi-k2.6	35	5
2026-08-30-143554	bytedance-seed/seed-2.0-lite (text)	60	42
20260901-002701	xiaomi/mimo-v2.5-pro (text)	60	30
20260905-051950	bytedance-seed/seed-2.0-lite (vision)	60	39

Important: the Pass column is the harness self-report. The project's grading convention treats manual audit as ground truth (see the 530 corpus README + evaluation policy). Some harness "passes" are false passes, and some honest failures (hallucination-control tasks) are intended passes. Do not treat this table as the final score.

Fabricated test data — disclosure

All on-device data is fabricated: a fictional persona ("Yuvraj Singh") with fake contacts, fake bank/finance SMS (e.g. HDFC "Sent Rs.130.00 Ref ..."), fake OTPs, fake calendar/notes/docs, and fabricated PDFs (Invoice, Rent Receipt). Any phone numbers, emails, bank refs, or identity details visible in the trajectories are synthetic benchmark seeds, not real personal data. Full disclosure: see the fabrication/ disclosure in the corpus repo.

License

MIT. The trajectory content is derived from the fabricated benchmark seeds and is safe to reproduce.