VisionClaw: Supplementary Materials

Methodology

Reproducible analysis pipeline for the autobiographical deployment study (N=4, 555 interactions).

Note: Raw session logs (raw-logs/) are not publicly released due to privacy and confidentiality concerns. Anonymized sample data is provided in the Example Data tab instead.

Reproduction

Step 1: Log Extraction

1cd scripts/2bash extract.sh /path/to/output/

Extracts raw session logs from the VisionClaw agent runtime. Queries sessions.json for all glass-initiated sessions, uses jq to extract structured fields (timestamp, role, text, tool calls, image flags, token usage), and produces per-participant JSONL files. Output files are renamed to anonymized identifiers (p1–p4) and placed in raw-logs/.

Step 2: Quantitative Analysis

1cd scripts/2python3 analyze.py ../raw-logs/

Input: 4 participant JSONL logs in raw-logs/ Output: 555 interactions, 118 sessions, 55 active days, 25.8 hours (see results/analysis-output.txt)

Step 3: Use-Case Classification

555 interactions were classified independently by 4 LLMs (GPT-5.4, GLM-5, Claude Opus-4.6, MiniMax-M2.7). Each model read every interaction and assigned one of 6 categories defined in classification/category-definitions.txt. Results: classification/{gpt,glm,opus,minimax}-classifications.jsonl

Step 4: Reconciliation

Final category percentages were determined by cross-analyst median across the 4 models, validated against two human coders' independent labels.

File Map

1raw-logs/                                # Per-participant interaction logs2  p1-glass-sessions.jsonl                # 139 raw -> 119 dedup, 17 active days3  p2-glass-sessions.jsonl                # 282 raw -> 173 dedup, 19 active days4  p3-glass-sessions.jsonl                # 110 raw ->  89 dedup,  5 active days5  p4-glass-sessions.jsonl                # 258 raw -> 174 dedup, 14 active days6 7scripts/8  extract.sh                             # Raw log extraction from VisionClaw system logs9  analyze.py                             # Main analysis script10 12classification/13  interactions.jsonl                     # 555 deduplicated interactions (input to classifiers)14  category-definitions.txt               # 6 category definitions from the paper15  gpt-classifications.jsonl              # GPT-5.4 classification (555 lines)16  glm-classifications.jsonl              # GLM-5 classification (555 lines)17  opus-classifications.jsonl             # Claude Opus-4.6 classification (555 lines)18  minimax-classifications.jsonl          # MiniMax-M2.7 classification (555 lines)19 20results/21  analysis-output.txt                    # Full output of analyze.py11  classification-output.txt              # LLM classification results summary

Key Numbers

1Aggregate (N=4)2  Total interactions          5553  Total sessions              1184  Active participant-days      555  Avg active days/participant  13.8 (range 5--19)6  Total hours                  25.87  Avg interactions/day         10.18  Max interactions/day         699 10Per-participant11  P1       119 ix, 17 days,  4.9h, 35 sessions  [2026-02-06..2026-03-31]12  P2       173 ix, 19 days, 10.4h, 47 sessions  [2026-03-09..2026-03-31]13  P3        89 ix,  5 days,  4.8h,  7 sessions  [2026-03-27..2026-03-31]14  P4       174 ix, 14 days,  5.6h, 29 sessions  [2026-03-15..2026-03-31]15 16Tools (N=1767 calls, avg 3.2/interaction)17  exec               557 (32%)18  browser            541 (31%)19  read               209 (12%)20  web_search         122  (7%)21  web_fetch           92  (5%)22  process             89  (5%)23  memory_search       49  (3%)24  write               22  (1%)25  image               15  (1%)26  edit                13  (1%)27  Depth: 0=21% | 1=27% | 2-3=23% | 4+=29% (max 27)28  Multi-source (2+):  29%29 30Latency (median)31  All                 12.2s32  Non-browser         13.4s33  Browser             15.5s34  Voice-only           8.4s35 36Use-case categories (cross-analyst median, 4 LLMs)37  Retrieve             30%38  Shop                 19%39  Save                 16%40  Communicate          14%41  Recall               12%42  Control               9%43 44Other45  Camera usage          39%46  No response            2%47  Time of day:  morning 26% | afternoon 44% | evening 27% | night 3%48  Session median:       16.0 min

Analysis Pipeline

System log extraction: extract.sh queried the VisionClaw agent runtime’s session store (sessions.json) to find all sessions initiated from the smart glasses (session keys matching agent:main:glass*). For each session, the script used jq to extract structured fields from every message: timestamp, role (user/assistant/toolResult), text content, tool call names, has_image flags, image URL references, and token usage. This produced two outputs per participant: (1) a clean structured JSONL with only the analysis-relevant fields, and (2) a full raw dump including tool results and model thinking traces. Files were sorted by timestamp, then renamed to anonymized identifiers (p1–p4) and placed in raw-logs/.
Interaction extraction & dedup: analyze.py parsed both glass-format ([original_instruction]) and text-chat records. Filters removed noise (misrecognition, system messages, short acknowledgements). Deduplication used a 120-second window on normalized query text.
Metrics computation: Tool-chain analysis scanned 50 subsequent events per interaction. Sessions defined by 30-minute inactivity gap. Latency capped at 600s.
Use-case classification: All 555 interactions classified by 4 LLMs in parallel. Each model read the full interaction text and assigned one of 6 categories. Results compared across models and against 2 human coders.
Validation: Final values cross-checked against independent analyses by two other team members. Disagreements resolved through reconciliation and manual review.
Visualization: The Visualization tab provides an interactive D3.js chart generated from the 555-interaction dataset. The chart shows interaction volume over time, colored by use-case category, with per-participant breakdowns.

Scripts

Core scripts used to extract and analyze the study data.

Extracts VisionClaw session records into structured and raw JSONL files.

1#!/usr/bin/env bash2# extract.sh — Extract VisionClaw session logs3# Usage: ./extract.sh [output-dir]4#5# Produces two files:6#   1. visionclaw-all-glass-sessions.jsonl       — Clean structured data (timestamp, role, text, tools, usage, image flags)7#   2. visionclaw-all-glass-sessions-full-raw.jsonl — Complete raw session data (everything including tool results, thinking)8#9# Data source: ~/.openclaw/agents/main/sessions/sessions.json10#   - Finds all session keys matching "agent:main:glass*"11#   - Reads each referenced .jsonl session file12#   - Merges and sorts by timestamp1314set -euo pipefail1516SESSIONS_DIR="${HOME}/.openclaw/agents/main/sessions"17SESSIONS_JSON="${SESSIONS_DIR}/sessions.json"18OUTPUT_DIR="${1:-/tmp/openclaw}"1920mkdir -p "$OUTPUT_DIR"2122if [ ! -f "$SESSIONS_JSON" ]; then23  echo "Error: sessions.json not found at $SESSIONS_JSON" >&224  exit 125fi2627echo "Finding glass session files..."2829# Get all glass session file paths from sessions.json30GLASS_FILES=$(jq -r '31  to_entries[] |32  select(.key | startswith("agent:main:glass")) |33  .value.sessionFile34' "$SESSIONS_JSON" 2>/dev/null | sort -u)3536FILE_COUNT=$(echo "$GLASS_FILES" | grep -c '.' || true)37echo "Found $FILE_COUNT glass session references"3839# --- File 1: Structured/clean extraction ---40echo "Extracting structured data..."41echo "$GLASS_FILES" | while read -r f; do42  if [ -f "$f" ]; then43    jq -c '44      select(.type=="message") |45      {46        timestamp: .timestamp,47        role: .message.role,48        content_types: [.message.content[]? | .type],49        text: ([.message.content[]? | select(.type=="text") | .text] | join("\n")),50        tools: ([.message.content[]? | select(.type=="toolCall") | .name] | if length > 0 then . else null end),51        has_image: ([.message.content[]? | .type] | any(. == "image")),52        has_image_url: (([.message.content[]? | select(.type=="text") | .text] | join("\n")) | test("tool_call_image_url")),53        usage: (if .message.usage then {54          input: .message.usage.inputTokens,55          output: .message.usage.outputTokens,56          cost: .message.usage.cost.total57        } else null end)58      }59    ' "$f" 2>/dev/null60  fi61done | sort -t'"' -k4 > "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl"6263# --- File 2: Full raw data ---64echo "Extracting full raw data..."65echo "$GLASS_FILES" | while read -r f; do66  if [ -f "$f" ]; then67    cat "$f"68  fi69done > "${OUTPUT_DIR}/visionclaw-all-glass-sessions-full-raw.jsonl"7071# --- Stats ---72STRUCTURED_LINES=$(wc -l < "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl")73RAW_LINES=$(wc -l < "${OUTPUT_DIR}/visionclaw-all-glass-sessions-full-raw.jsonl")74STRUCTURED_SIZE=$(ls -lh "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | awk '{print $5}')75RAW_SIZE=$(ls -lh "${OUTPUT_DIR}/visionclaw-all-glass-sessions-full-raw.jsonl" | awk '{print $5}')7677FIRST_TS=$(jq -r '.timestamp' "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | head -1)78LAST_TS=$(jq -r '.timestamp' "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | tail -1)7980USER_COUNT=$(jq -r 'select(.role=="user")' "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | wc -l)81ASSISTANT_COUNT=$(jq -r 'select(.role=="assistant")' "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | wc -l)82IMAGE_COUNT=$(jq -r 'select(.has_image_url==true)' "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | wc -l)8384echo ""85echo "=== Done ==="86echo "Structured: ${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl (${STRUCTURED_LINES} lines, ${STRUCTURED_SIZE})"87echo "Full raw:   ${OUTPUT_DIR}/visionclaw-all-glass-sessions-full-raw.jsonl (${RAW_LINES} lines, ${RAW_SIZE})"88echo ""89echo "Date range: ${FIRST_TS} → ${LAST_TS}"90echo "Messages:   user=${USER_COUNT} assistant=${ASSISTANT_COUNT}"91echo "With camera images: ${IMAGE_COUNT}"92echo ""93echo "Per-day breakdown:"94jq -r '.timestamp[0:10]' "${OUTPUT_DIR}/visionclaw-all-glass-sessions.jsonl" | sort | uniq -c

Quantitative analysis pipeline producing the summary values used in the paper.

1#!/usr/bin/env python32"""3VisionClaw Deployment Study — Analysis Script4===============================================5Merged analysis pipeline integrating multiple independently developed variants.67Includes Thai-noise filtering, system-message filtering, 120-second deduplication, and session-based hour estimates.8Includes timezone conversion and multi-source counting.9Includes a scan window of 50 events, tool usage counting, and per-participant session summaries.1011Usage:12  python3 analyze.py <logs-dir>13"""1415import json, re, os, sys, glob, statistics16from collections import Counter, defaultdict17from datetime import datetime1819# - Configuration -20SESSION_GAP_SEC = 1800      # 30 min21DEDUP_WINDOW_SEC = 120      # 2 min22LAT_CAP_SEC = 600           # 10 min23TIMEZONE_OFFSET = -6        # -6 hours24SCAN_WINDOW = 50            # scan window for tool-chain tracing2526SKIP_SHORT = {"done", "ok", "yes", "no", "ok installed", "hi", "ok do it"}2728# System message patterns to skip29SYS_PATTERNS = [30    r'^A new session was started via /new or /reset',31    r'Summarize this naturally for the user',32    r'just completed successfully\.',33    r'You are running as a subagent',34]353637# - Helpers -3839def load_jsonl(path):40    data = []41    with open(path) as f:42        for line in f:43            try:44                data.append(json.loads(line))45            except:46                continue47    return data484950def parse_ts(ts):51    if not ts:52        return None53    try:54        return datetime.fromisoformat(ts.replace("Z", "+00:00"))55    except:56        return None575859def is_noise(text):60    """Detect Thai-character noise caused by speech-recognition errors."""61    return len(re.findall(r'[\u0E00-\u0E7F]', text or '')) >= 1626364def is_system_message(text):65    """Detect system or metadata messages."""66    return any(re.search(p, text, re.IGNORECASE) for p in SYS_PATTERNS)676869def clean_text_chat(text):70    """Clean messages in text-chat format."""71    t = (text or '').strip()72    if not t:73        return None74    if is_system_message(t):75        return None76    # Handle the [Chat messages since...] wrapper77    if t.startswith('[Chat messages'):78        m = re.search(r'User: (.+?)(?:\nAssistant:|\[Current|\Z)', t, re.S)79        if not m:80            return None81        t = m.group(1).strip()82    # Strip message IDs83    t = re.sub(r'\[message_id:.*?\]', '', t).strip()84    t = re.sub(r'^\[[^\]]+\]\s*', '', t).strip()85    if len(t) < 3:86        return None87    if t.lower().rstrip('!.') in SKIP_SHORT:88        return None89    if is_noise(t):90        return None91    return t929394# - Extraction -9596def extract_interactions(data):97    """Per-message parser for both glass-format and text-chat records."""98    ixs = []99    for i, e in enumerate(data):100        if e.get("role") != "user":101            continue102        text = e.get("text", "")103        ts = e.get("timestamp", "")104        if not ts:105            continue106107        # --- Parse interaction text ---108        if "[original_instruction]" in text:109            # Glass format (supports both multi-line and single-line records)110            om = re.search(111                r'\[original_instruction\]\s*\n?(.*?)(?:\[gemini_rewritten|\[tool_call|\[message_id|\Z)',112                text, re.DOTALL)113            rm = re.search(114                r'\[gemini_rewritten_instruction\]\s*\n?(.*?)(?:\[tool_call|\[message_id|\Z)',115                text, re.DOTALL)116            orig = om.group(1).strip() if om else ""117            rewritten = rm.group(1).strip() if rm else orig118            if orig.strip() in ["<noise>", "", ".", ".."]:119                continue120            if is_noise(orig) or is_noise(rewritten):121                continue122            if not rewritten or len(rewritten) < 3:123                continue124            query = rewritten[:300]125        else:126            # Text-chat format127            query = clean_text_chat(text)128            if not query:129                continue130            query = query[:300]131132        # --- Extract tools and response over the scan window ---133        tools = []134        resp_text = ""135        resp_ts = ""136        for j in range(i + 1, min(i + SCAN_WINDOW, len(data))):137            nxt = data[j]138            role = nxt.get("role")139            if role == "assistant":140                if nxt.get("tools"):141                    tools.extend(nxt["tools"])142                nt = (nxt.get("text") or "").strip()143                if nt and not resp_text:144                    resp_text = nt[:500]145                    resp_ts = nxt.get("timestamp", "")146            elif role == "toolResult":147                continue  # skip, don't break — trace full chain148            elif role == "user":149                break150151        # Latency152        lat = None153        t1, t2 = parse_ts(ts), parse_ts(resp_ts)154        if t1 and t2:155            lat = (t2 - t1).total_seconds()156            if lat <= 0 or lat > LAT_CAP_SEC:157                lat = None158159        # Source types for multi-source counting160        source_types = set()161        for t in tools:162            if t in ("exec",):163                source_types.add("shell")164            elif t in ("web_fetch", "web_search"):165                source_types.add("web")166            elif t in ("browser",):167                source_types.add("browser")168            elif t in ("memory_search", "memory_get"):169                source_types.add("memory")170            elif t in ("read", "write", "edit"):171                source_types.add("files")172173        ixs.append({174            "timestamp": ts,175            "date": ts[:10],176            "query": query,177            "tools": tools,178            "tool_count": len(tools),179            "unique_tools": len(set(tools)),180            "has_browser": "browser" in tools,181            "latency": lat,182            "has_response": bool(resp_text),183            "source_count": len(source_types),184        })185    return ixs186187188# - Deduplication (120s normalized window) -189190def dedup_key(s):191    s = (s or "").lower()192    s = re.sub(r'\s+', ' ', s).strip()193    s = re.sub(r'[^a-z0-9 ]', '', s)194    return s[:120]195196197def dedup(rows):198    """Merge-style deduplication that keeps the best tool and latency data."""199    out = []200    seen = {}  # key → index in out201    for r in rows:202        k = dedup_key(r["query"])203        t = parse_ts(r["timestamp"])204        if k in seen:205            idx = seen[k]206            prev = out[idx]207            prev_t = parse_ts(prev["timestamp"])208            if prev_t and t and abs((t - prev_t).total_seconds()) < DEDUP_WINDOW_SEC:209                # Merge duplicate records while preserving the most informative fields210                for tool in r["tools"]:211                    if tool not in prev["tools"]:212                        prev["tools"].append(tool)213                prev["tool_count"] = len(prev["tools"])214                prev["unique_tools"] = len(set(prev["tools"]))215                if r["latency"] and (not prev["latency"] or r["latency"] < prev["latency"]):216                    prev["latency"] = r["latency"]217                if r["has_response"]:218                    prev["has_response"] = True219                # Recompute source types after merging220                src = set()221                for tool in prev["tools"]:222                    if tool in ("exec",): src.add("shell")223                    elif tool in ("web_fetch", "web_search"): src.add("web")224                    elif tool in ("browser",): src.add("browser")225                    elif tool in ("memory_search", "memory_get"): src.add("memory")226                    elif tool in ("read", "write", "edit"): src.add("files")227                prev["source_count"] = len(src)228                continue229        if t:230            seen[k] = len(out)231        out.append(r)232    return out233234235# - Sessions & Hours (session-based) -236237def sessionize(rows):238    if not rows:239        return []240    srows = sorted(rows, key=lambda x: x["timestamp"])241    sessions = [[srows[0]]]242    for r in srows[1:]:243        t1 = parse_ts(sessions[-1][-1]["timestamp"])244        t2 = parse_ts(r["timestamp"])245        if t1 and t2 and (t2 - t1).total_seconds() > SESSION_GAP_SEC:246            sessions.append([r])247        else:248            sessions[-1].append(r)249    return sessions250251252def session_hours(sessions):253    total = 0254    for s in sessions:255        if len(s) >= 2:256            t1 = parse_ts(s[0]["timestamp"])257            t2 = parse_ts(s[-1]["timestamp"])258            if t1 and t2:259                d = (t2 - t1).total_seconds() / 3600260                if d >= 0:261                    total += d262    return total263264265def session_durations_min(sessions):266    durs = []267    for s in sessions:268        if len(s) >= 2:269            t1 = parse_ts(s[0]["timestamp"])270            t2 = parse_ts(s[-1]["timestamp"])271            if t1 and t2:272                d = (t2 - t1).total_seconds() / 60273                if d >= 0:274                    durs.append(d)275    return durs276277278def med(lst):279    return statistics.median(lst) if lst else 0280281282def pct(a, b):283    return (100 * a / b) if b else 0284285286# - Main -287288def main():289    if len(sys.argv) < 2:290        print("Usage: python3 analyze.py <logs-dir>")291        sys.exit(1)292293    logdir = sys.argv[1]294    files = sorted(glob.glob(os.path.join(logdir, "p*-glass-sessions.jsonl")))295    if not files:296        print(f"No matching files in {logdir}")297        sys.exit(1)298299    # --- Extract per participant ---300    per_p = {}301    all_ixs = []302303    for fpath in files:304        name = os.path.basename(fpath).replace("-glass-sessions.jsonl", "")305        pid = name.split("-")[0]306        data = load_jsonl(fpath)307        raw = extract_interactions(data)308        dd = dedup(raw)309        sessions = sessionize(dd)310        hours = session_hours(sessions)311        dates = sorted(set(x["date"] for x in dd if x["date"]))312313        per_p[pid] = {314            "raw": len(raw), "dedup": len(dd),315            "days": len(dates), "hours": hours,316            "sessions": len(sessions),317            "range": f"{dates[0]}..{dates[-1]}" if dates else "",318        }319        all_ixs.extend(dd)320        print(f"{pid}: {len(raw)} raw -> {len(dd)} dedup, {len(dates)} days, {hours:.1f}h")321322    # --- Aggregate ---323    N = len(all_ixs)324    total_days = sum(p["days"] for p in per_p.values())325    total_hours = sum(p["hours"] for p in per_p.values())326    days_list = [p["days"] for p in per_p.values()]327    total_sessions = sum(p["sessions"] for p in per_p.values())328329    # Tools330    all_tools = []331    for ix in all_ixs:332        all_tools.extend(ix["tools"])333    tc = Counter(all_tools)334    tool_total = sum(tc.values())335336    d0 = sum(1 for x in all_ixs if x["tool_count"] == 0)337    d1 = sum(1 for x in all_ixs if x["tool_count"] == 1)338    d23 = sum(1 for x in all_ixs if 2 <= x["tool_count"] <= 3)339    d4p = sum(1 for x in all_ixs if x["tool_count"] >= 4)340    max_depth = max((x["tool_count"] for x in all_ixs), default=0)341342    # Latency343    l_all = sorted([x["latency"] for x in all_ixs if x["latency"]])344    l_nb = sorted([x["latency"] for x in all_ixs if x["latency"] and not x["has_browser"] and x["tool_count"] > 0])345    l_br = sorted([x["latency"] for x in all_ixs if x["latency"] and x["has_browser"]])346    l_vo = sorted([x["latency"] for x in all_ixs if x["latency"] and x["tool_count"] == 0])347348    # Response, multi-source349    no_resp = sum(1 for x in all_ixs if not x["has_response"])350    multi_src = sum(1 for x in all_ixs if x["source_count"] >= 2)351352    # Time of day (MDT conversion)353    tod = Counter()354    for ix in all_ixs:355        t = parse_ts(ix["timestamp"])356        if t:357            h = (t.hour + TIMEZONE_OFFSET) % 24358            if 6 <= h < 12:359                tod["morning"] += 1360            elif 12 <= h < 18:361                tod["afternoon"] += 1362            elif 18 <= h < 24:363                tod["evening"] += 1364            else:365                tod["night"] += 1366367    # Sessions (merged for aggregate)368    all_sessions = sessionize(all_ixs)369    sess_durs = session_durations_min(all_sessions)370371    # Daily372    daily = Counter(x["date"] for x in all_ixs)373    max_daily = max(daily.values()) if daily else 0374375    # --- Output ---376    print(f"\n{'='*70}")377    print(f"  Results")378    print(f"{'='*70}")379380    print(f"\nPer-participant:")381    for name, p in sorted(per_p.items()):382        print(f"  {name}: {p['dedup']} ix, {p['days']} days, {p['hours']:.1f}h, {p['sessions']} sessions  [{p['range']}]")383384    print(f"\nAggregate (N={len(per_p)}):")385    print(f"  Total interactions:     {N}")386    print(f"  Total sessions:         {total_sessions}")387    print(f"  Total active days:      {total_days}")388    print(f"  Total hours:            {total_hours:.1f}")389    print(f"  Avg days/participant:   {total_days/len(per_p):.1f} ({min(days_list)}--{max(days_list)})")390    print(f"  Avg ix/active day:      {N/total_days:.1f}")391    print(f"  Max ix in one day:      {max_daily}")392393    print(f"\nTools:")394    print(f"  Avg tool calls/ix:      {tool_total/N:.1f}")395    print(f"  Breakdown (N={tool_total}):")396    for t, c in tc.most_common(10):397        print(f"    {t:<15} {c:>4} ({pct(c, tool_total):.0f}%)")398    print(f"  Depth: 0={pct(d0,N):.0f}% | 1={pct(d1,N):.0f}% | 2-3={pct(d23,N):.0f}% | 4+={pct(d4p,N):.0f}% (max {max_depth})")399    print(f"  Multi-source (2+):      {pct(multi_src, N):.0f}%")400401    print(f"\nLatency (median):")402    print(f"  All:           {med(l_all):.1f}s")403    print(f"  Non-browser:   {med(l_nb):.1f}s")404    print(f"  Browser:       {med(l_br):.1f}s")405    print(f"  Voice-only:    {med(l_vo):.1f}s")406407    print(f"\nOther:")408    print(f"  No response:   {no_resp}/{N} ({pct(no_resp, N):.0f}%)")409    print(f"  Time of day:   morning {pct(tod['morning'],N):.0f}% | afternoon {pct(tod['afternoon'],N):.0f}% | evening {pct(tod['evening'],N):.0f}% | night {pct(tod['night'],N):.0f}%")410    print(f"  Session median: {med(sess_durs):.1f} min")411412    print(f"\nMethodology:")413    print(f"  Format: per-message mixed parser (glass + text-chat)")414    print(f"  Filters: <noise>, Thai noise, system msgs, short ack")415    print(f"  Dedup: {DEDUP_WINDOW_SEC}s window, normalized text key (120 chars)")416    print(f"  Tool scan: {SCAN_WINDOW} entries, skip toolResult")417    print(f"  Session gap: {SESSION_GAP_SEC//60} min")418    print(f"  Hours: session-based (interaction timestamps only)")419    print(f"  Timezone: MDT (UTC{TIMEZONE_OFFSET:+d})")420    print(f"  Latency cap: {LAT_CAP_SEC}s")421422423if __name__ == "__main__":424    main()

Example Data

Anonymized sample from the raw glass session logs (JSONL format). Each line is one message record as logged by the VisionClaw system.

Analysis

Quantitative analysis of all 555 interactions using scripts/analyze.py.

How We Defined and Computed Each Metric

Below describes how each metric reported in the paper was operationalized:

Interaction: A single user-initiated voice command or text input that triggers an agent response. Each raw log entry contains both the original speech transcription and a Gemini-rewritten instruction; the rewritten version is used as the canonical query. Entries marked as <noise>, containing noise artifacts from speech recognition errors, system-generated messages, or short acknowledgements (e.g., "ok", "yes") are excluded.
Deduplication: When the same query appears multiple times within a 120-second window (common with voice input retries), entries are merged into a single interaction. The merge preserves the most complete tool list and the fastest latency.
Session: A sequence of interactions with no gap exceeding 30 minutes. Total usage hours are computed as the sum of first-to-last interaction timestamps within each session.
Tool depth: The number of tool calls the agent made to fulfill a single interaction. The script traces up to 50 subsequent log events to capture the full tool chain (e.g., web_search → browser → exec).
Latency: Wall-clock time from the user message timestamp to the first assistant text response. Capped at 600 seconds to exclude stale sessions. Reported as median, broken down by: all interactions, non-browser tool use, browser-involved, and voice-only (no tools).
Multi-source: An interaction is multi-source if the agent used 2 or more distinct tool categories: shell (exec), web (web_search/web_fetch), browser, memory (memory_search/memory_get), or files (read/write/edit).
Time of day: UTC timestamps are converted to local time and bucketed: morning (6–12), afternoon (12–18), evening (18–24), night (0–6).

Run: python3 scripts/analyze.py raw-logs/

Output

Full reproduced output:

1p1: 139 raw -> 119 dedup, 17 days, 4.9h2p2: 282 raw -> 173 dedup, 19 days, 10.4h3p3: 110 raw -> 89 dedup, 5 days, 4.8h4p4: 258 raw -> 174 dedup, 14 days, 5.6h56======================================================================7  Results8======================================================================910Per-participant:11  p1: 119 ix, 17 days, 4.9h, 35 sessions  [2026-02-06..2026-03-31]12  p2: 173 ix, 19 days, 10.4h, 47 sessions  [2026-03-09..2026-03-31]13  p3: 89 ix, 5 days, 4.8h, 7 sessions  [2026-03-27..2026-03-31]14  p4: 174 ix, 14 days, 5.6h, 29 sessions  [2026-03-15..2026-03-31]1516Aggregate (N=4):17  Total interactions:     55518  Total sessions:         11819  Total active days:      5520  Total hours:            25.821  Avg days/participant:   13.8 (5--19)22  Avg ix/active day:      10.123  Max ix in one day:      692425Tools:26  Avg tool calls/ix:      3.227  Breakdown (N=1767):28    exec             557 (32%)29    browser          541 (31%)30    read             209 (12%)31    web_search       122 (7%)32    web_fetch         92 (5%)33    process           89 (5%)34    memory_search     49 (3%)35    write             22 (1%)36    image             15 (1%)37    edit              13 (1%)38  Depth: 0=21% | 1=27% | 2-3=23% | 4+=29% (max 27)39  Multi-source (2+):      29%4041Latency (median):42  All:           12.2s43  Non-browser:   13.4s44  Browser:       15.5s45  Voice-only:    8.4s4647Other:48  No response:   9/555 (2%)49  Time of day:   morning 26% | afternoon 44% | evening 27% | night 3%50  Session median: 16.0 min5152Methodology:53  Format: per-message mixed parser (glass + text-chat)54  Filters: <noise>, Thai noise, system msgs, short ack55  Dedup: 120s window, normalized text key (120 chars)56  Tool scan: 50 entries, skip toolResult57  Session gap: 30 min58  Hours: session-based (interaction timestamps only)59  Timezone: MDT (UTC-6)60  Latency cap: 600s

Classification

Use-case classification process, from open coding to final LLM-assisted labeling of all 555 interactions.

Phase 1: Open Coding

Three authors independently reviewed the first 200--300 interactions of deployment data and identified 10--15 initial use-case categories through LLM-assisted open coding. Categories were then consolidated via thematic analysis: overlapping codes were merged and unique categories retained, producing the final 6 categories.

1Category      Description2--------------------------------------------------------------------3Communicate   Hands-free email and messaging (inbox triage, compose, Slack, iMessage)4Retrieve      Situated information lookup (weather, research, camera-grounded queries)5Save          Hands-free capture (meal photos, receipts, posters, voice notes)6Recall        Context-triggered memory recall (past meals, meetings, decisions)7Shop          Camera-assisted product discovery (Amazon cart, price comparison)8Control       Task automation and device control (IoT, GitHub, cron, LaTeX)

Full definitions and representative examples for each category are in classification/category-definitions.txt.

Phase 2: LLM Classification

Using the category definitions from Phase 1, Author A classified all 555 interactions with 4 LLMs in parallel (GPT-5.4, GLM-5, Opus-4.6, MiniMax-M2.7). Each model read the full interaction text and tool calls, then assigned one of the 6 categories. Final labels were determined by majority vote across the 4 models.

1Category       GPT-5.4  GLM-5  Opus-4.6  MiniMax  Majority2----------------------------------------------------------------3Communicate     100(18%)  96(17%)  88(16%)  89(16%)   94(17%)4Retrieve        129(23%) 143(26%) 162(29%) 158(28%)  146(26%)5Save             94(17%)  91(16%)  95(17%)  97(17%)   96(17%)6Recall           84(15%)  62(11%)  72(13%)  52( 9%)   63(11%)7Shop             91(16%) 111(20%)  84(15%) 103(19%)  103(19%)8Control          57(10%)  52( 9%)  54(10%)  56(10%)   53(10%)910Agreement: all 4 agree 403/555 (73%), 3-of-4 115 (21%), 2-2 split 37 (7%)

Phase 3: Cross-Analyst Validation

Author B independently classified all 555 interactions using GPT-5.4, manually reviewing each label. Results were compared against Author A's 4-LLM majority vote.

1                 Author A      Author B2Category        4-LLM maj.    (GPT-5.4)3----------------------------------------------4Communicate       94(17%)       59(11%)5Retrieve         146(26%)      183(33%)6Save              96(17%)       91(16%)7Recall            63(11%)       64(12%)8Shop             103(19%)      105(19%)9Control           53(10%)       53(10%)

Phase 4: Final Values

Final category percentages were determined by taking the cross-analyst median of the remaining non-outlier results. Four categories showed strong consensus; the remaining two were resolved by median. Raw medians summed to 103%, so values were normalized to 100% (e.g., Control rounded from 10% to 9%).

1Final use-case distribution (N=555)2  Communicate      14%3  Retrieve         30%4  Save             16%5  Recall           12%6  Shop             19%7  Control           9%

Camera Detection

Camera usage was estimated through two independent methods: (1) automated detection via has_image flags and tool_call_image_url patterns in the session logs (~25%), and (2) LLM-assisted detection during classification, which identified additional camera interactions not captured by log flags (deictic references like "what is this", "add this to cart"). The final estimate of 39% reflects the combined signal.

Anonymized Representative Examples

Interactions where all 4 LLMs unanimously agreed on the category (403/555 = 73%):

1Communicate (75 unanimous / 94 total)2  "Send an email to [email] with the subject 'hi' and no body text"3  "Archive the newsletter emails"4  "Check the latest message in the Slack channel"5  "Send message to [person name] with 'On my way'"6  "Draft a collaboration email to the first author of this paper of [the captured paper]"78Retrieve (101 unanimous / 146 total)9  "Identify the type of sushi in the current view"10  "What time is sunset in [location] today?"11  "Search for the related research paper 'Genie: Generative Interactive Environments'"12  "Compare ingredients of [country name] and [country name] Sun Chips to see if they are the same"13  "Search for over-the-counter allergy meds similar to this prescription"14  "Give me my daily briefing for today"15  "Troubleshoot why the rice cooker lid is stuck"1617Save (84 unanimous / 96 total)18  "Save the current meal to the food log"19  "Create a memo in Notion from the Whole Foods receipt"20  "Create a Google Calendar event for [schedule name] on Saturday at 8 PM"21  "Take a photo and save it to the 'photos' folder in Google Drive"22  "Save the visible hotel information document to memory"23  "Capture the research poster and save it to Notion"2425Recall (49 unanimous / 63 total)26  "Recall what I ate yesterday"27  "Search memory for what I have eaten this week"28  "Check previous conversation about the Barcelona flight"29  "Search notes for gift ideas for [person name]"30  "What was I doing 5 years ago today?"31  "Check key logger for lost keys"3233Shop (66 unanimous / 103 total)34  "Find sunglasses with a design similar to [this] black classic-style sunglasses"35  "Search Amazon for a cheaper price on the Hydro Flask bottle"36  "Search Amazon for Pampers baby wipes price: single pack vs multi-pack"37  "Search Amazon for SIHOO Doro C300 Pro office chair and add to cart"38  "Search for cheaper alternatives for AirPods Pro and text the results via Telegram"3940Control (28 unanimous / 53 total)40  "Turn off the living room light"41  "Change the smart light bulb color to red"42  "Open the PDF with the URL [url] on the Mac"43  "Debug why it is not logging any TSP file in the VisionClaw skill"

Anonymized Ambiguous Cases (2-2 split)

37 interactions (7%) where models split evenly between two categories. These were not individually adjudicated; final category percentages were determined at the aggregate level via cross-analyst median (Phase 4).

Save vs Shop (boundary: "add to list" vs "add to cart")
1"Search Amazon for box of Diet Coke and add to shopping list"2 GPT=Save, GLM=Save, Opus=Shop, MiniMax=Shop => Final: Shop
Retrieve vs Shop (boundary: product research vs purchase intent)
1"Check the delivery date for the SIHOO chair Amazon order"2 GPT=Retrieve, GLM=Shop, Opus=Retrieve, MiniMax=Shop => Final: Retrieve34"Find thin dental floss similar to previous one"4 GPT=Shop, GLM=Shop, Opus=Retrieve, MiniMax=Retrieve => Final: Shop
Retrieve vs Communicate (boundary: status check vs messaging action)
1"Check the shipping status of the visa application passport package"2 GPT=Retrieve, GLM=Communicate, Opus=Communicate, MiniMax=Retrieve => Final: Communicate34"Check the status of domain [domain name] renewal on Porkbun"4 GPT=Communicate, GLM=Communicate, Opus=Retrieve, MiniMax=Retrieve => Final: Retrieve56"Check email inbox"6 GPT=Communicate, GLM=Retrieve, Opus=Communicate, MiniMax=Retrieve => Final: Communicate
Control vs Save (boundary: automation setup vs persistent capture)
1"Set up recurring task to monitor for keys, take a photo when seen"2 GPT=Control, GLM=Control, Opus=Save, MiniMax=Save => Final: Control
Retrieve vs Recall (boundary: web search vs memory search)
1"Search memory for recent Carhart purchases and draft a memo"2 GPT=Recall, GLM=Retrieve, Opus=Retrieve, MiniMax=Recall => Final: Retrieve
Control vs Shop (boundary: UI navigation vs shopping flow)
1"Click the first search result"2 GPT=Control, GLM=Control, Opus=Shop, MiniMax=Shop => Final: Control

Visualization

Session-level dataset used by the bubble chart above. Each entry represents one session: anonymized participant, deployment day, start hour, interaction count, duration in minutes, dominant category, per-category breakdown, and action types. 118 sessions total.

Schema

1{2  "pid": "P1",                // Anonymized participant (P1-P4)3  "day": 1,                  // Deployment day (1-54)4  "hour": 13.15,             // Session start (decimal hours, local time)5  "interactions": 25,        // Number of interactions in session6  "duration": 67.9,          // Session duration (minutes)7  "category": "Shop",        // Dominant category (plurality)8  "categories": {"Save": 7, "Shop": 18},  // Per-category counts9  "actions": ["browsed/added to cart", ...]  // Action types observed10}

Data (118 sessions)

1{"pid": "P1", "day": 1, "hour": 5.22, "interactions": 6, "duration": 5.2, "category": "Control", "categories": {"Control": 6}, "actions": ["opened webpage", "general query"]}2{"pid": "P1", "day": 1, "hour": 13.15, "interactions": 25, "duration": 67.9, "category": "Shop", "categories": {"Save": 7, "Shop": 18}, "actions": ["browsed/added to cart", "general query", "opened webpage"]}3{"pid": "P1", "day": 5, "hour": 15.9, "interactions": 12, "duration": 5.1, "category": "Communicate", "categories": {"Save": 3, "Shop": 4, "Communicate": 5}, "actions": ["browsed/added to cart", "checked/sent email", "general query", "opened webpage"]}4{"pid": "P1", "day": 6, "hour": 20.12, "interactions": 5, "duration": 15.8, "category": "Shop", "categories": {"Recall": 2, "Shop": 3}, "actions": ["browsed/added to cart", "general query"]}5{"pid": "P1", "day": 6, "hour": 22.43, "interactions": 13, "duration": 48.9, "category": "Shop", "categories": {"Retrieve": 3, "Shop": 9, "Control": 1}, "actions": ["searched web", "browsed/added to cart", "opened webpage"]}6{"pid": "P1", "day": 6, "hour": 13.83, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}7{"pid": "P1", "day": 7, "hour": 19.6, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}8{"pid": "P1", "day": 9, "hour": 10.97, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}9{"pid": "P1", "day": 10, "hour": 23.67, "interactions": 1, "duration": 1, "category": "Save", "categories": {"Save": 1}, "actions": ["saved/noted item"]}10{"pid": "P1", "day": 13, "hour": 19.48, "interactions": 3, "duration": 3.3, "category": "Shop", "categories": {"Shop": 3}, "actions": ["browsed/added to cart"]}11{"pid": "P1", "day": 13, "hour": 20.15, "interactions": 3, "duration": 1.9, "category": "Shop", "categories": {"Shop": 3}, "actions": ["browsed/added to cart"]}12{"pid": "P1", "day": 13, "hour": 2.38, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}13{"pid": "P1", "day": 26, "hour": 12.42, "interactions": 3, "duration": 20.4, "category": "Shop", "categories": {"Shop": 3}, "actions": ["browsed/added to cart", "searched web"]}14{"pid": "P1", "day": 30, "hour": 23.65, "interactions": 4, "duration": 32.7, "category": "Shop", "categories": {"Shop": 4}, "actions": ["browsed/added to cart"]}15{"pid": "P1", "day": 35, "hour": 11.8, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}16{"pid": "P1", "day": 38, "hour": 22.57, "interactions": 4, "duration": 7.7, "category": "Shop", "categories": {"Shop": 2, "Retrieve": 1, "Control": 1}, "actions": ["browsed/added to cart", "searched web", "general query"]}17{"pid": "P1", "day": 38, "hour": 23.3, "interactions": 8, "duration": 28.9, "category": "Control", "categories": {"Control": 4, "Shop": 4}, "actions": ["browsed/added to cart", "opened webpage"]}18{"pid": "P1", "day": 49, "hour": 15.28, "interactions": 2, "duration": 1.1, "category": "Communicate", "categories": {"Communicate": 2}, "actions": ["checked/sent email"]}19{"pid": "P1", "day": 49, "hour": 15.87, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["opened webpage"]}20{"pid": "P1", "day": 49, "hour": 16.55, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["opened webpage"]}21{"pid": "P1", "day": 49, "hour": 17.13, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["general query"]}22{"pid": "P1", "day": 50, "hour": 10.37, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["general query"]}23{"pid": "P1", "day": 50, "hour": 11.07, "interactions": 1, "duration": 1, "category": "Communicate", "categories": {"Communicate": 1}, "actions": ["checked/sent email"]}24{"pid": "P1", "day": 50, "hour": 13.85, "interactions": 3, "duration": 1.5, "category": "Shop", "categories": {"Shop": 3}, "actions": ["browsed/added to cart", "general query"]}25{"pid": "P1", "day": 50, "hour": 17.45, "interactions": 1, "duration": 1, "category": "Control", "categories": {"Control": 1}, "actions": ["general query"]}26{"pid": "P1", "day": 51, "hour": 10.78, "interactions": 1, "duration": 1, "category": "Communicate", "categories": {"Communicate": 1}, "actions": ["general query"]}27{"pid": "P1", "day": 51, "hour": 13.63, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}28{"pid": "P1", "day": 52, "hour": 18.22, "interactions": 3, "duration": 27.9, "category": "Retrieve", "categories": {"Retrieve": 2, "Save": 1}, "actions": ["saved/noted item", "general query"]}29{"pid": "P1", "day": 52, "hour": 13.65, "interactions": 3, "duration": 1.0, "category": "Communicate", "categories": {"Communicate": 2, "Retrieve": 1}, "actions": ["opened webpage", "checked schedule", "general query"]}30{"pid": "P1", "day": 52, "hour": 15.03, "interactions": 2, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 2}, "actions": ["searched web", "general query"]}31{"pid": "P1", "day": 53, "hour": 9.62, "interactions": 2, "duration": 26.6, "category": "Recall", "categories": {"Recall": 1, "Communicate": 1}, "actions": ["checked/sent email", "recalled past info"]}32{"pid": "P1", "day": 53, "hour": 10.68, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["opened webpage"]}33{"pid": "P1", "day": 53, "hour": 12.32, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["opened webpage"]}34{"pid": "P1", "day": 53, "hour": 15.23, "interactions": 1, "duration": 1, "category": "Control", "categories": {"Control": 1}, "actions": ["general query"]}35{"pid": "P1", "day": 54, "hour": 13.73, "interactions": 1, "duration": 1, "category": "Communicate", "categories": {"Communicate": 1}, "actions": ["general query"]}36{"pid": "P2", "day": 32, "hour": 14.47, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}37{"pid": "P2", "day": 34, "hour": 11.68, "interactions": 3, "duration": 26.6, "category": "Save", "categories": {"Save": 2, "Communicate": 1}, "actions": ["searched web", "checked/sent email", "saved/noted item", "opened webpage"]}38{"pid": "P2", "day": 34, "hour": 14.53, "interactions": 3, "duration": 4.7, "category": "Shop", "categories": {"Save": 1, "Shop": 2}, "actions": ["browsed/added to cart", "searched web", "saved/noted item"]}39{"pid": "P2", "day": 35, "hour": 14.03, "interactions": 1, "duration": 1, "category": "Shop", "categories": {"Shop": 1}, "actions": ["browsed/added to cart"]}40{"pid": "P2", "day": 36, "hour": 10.2, "interactions": 3, "duration": 3.5, "category": "Shop", "categories": {"Shop": 3}, "actions": ["browsed/added to cart", "searched web", "general query"]}41{"pid": "P2", "day": 36, "hour": 11.9, "interactions": 1, "duration": 1, "category": "Save", "categories": {"Save": 1}, "actions": ["searched web", "opened webpage"]}42{"pid": "P2", "day": 36, "hour": 14.55, "interactions": 1, "duration": 1, "category": "Save", "categories": {"Save": 1}, "actions": ["saved/noted item"]}43{"pid": "P2", "day": 37, "hour": 12.08, "interactions": 3, "duration": 3.8, "category": "Save", "categories": {"Save": 2, "Recall": 1}, "actions": ["saved/noted item", "opened webpage"]}44{"pid": "P2", "day": 38, "hour": 23.55, "interactions": 2, "duration": 2.2, "category": "Shop", "categories": {"Shop": 2}, "actions": ["browsed/added to cart", "searched web"]}45{"pid": "P2", "day": 38, "hour": 10.18, "interactions": 2, "duration": 10.3, "category": "Control", "categories": {"Control": 2}, "actions": ["general query"]}46{"pid": "P2", "day": 39, "hour": 18.28, "interactions": 10, "duration": 80.9, "category": "Recall", "categories": {"Control": 3, "Recall": 4, "Shop": 3}, "actions": ["checked tasks", "recalled past info", "opened webpage", "general query", "browsed/added to cart", "saved/noted item"]}47{"pid": "P2", "day": 39, "hour": 11.23, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["searched web"]}48{"pid": "P2", "day": 39, "hour": 11.75, "interactions": 9, "duration": 35.6, "category": "Communicate", "categories": {"Control": 3, "Communicate": 4, "Recall": 2}, "actions": ["checked tasks", "recalled past info", "sent message", "general query", "saved/noted item"]}49{"pid": "P2", "day": 39, "hour": 13.07, "interactions": 3, "duration": 39.8, "category": "Save", "categories": {"Save": 3}, "actions": ["saved/noted item", "general query"]}50{"pid": "P2", "day": 39, "hour": 15.4, "interactions": 2, "duration": 4.7, "category": "Shop", "categories": {"Shop": 2}, "actions": ["browsed/added to cart"]}51{"pid": "P2", "day": 40, "hour": 19.8, "interactions": 9, "duration": 57.2, "category": "Save", "categories": {"Save": 4, "Recall": 1, "Communicate": 2, "Control": 2}, "actions": ["recalled past info", "saved/noted item", "general query", "sent message"]}52{"pid": "P2", "day": 40, "hour": 10.9, "interactions": 12, "duration": 30.9, "category": "Recall", "categories": {"Recall": 6, "Save": 1, "Retrieve": 5}, "actions": ["searched web", "checked tasks", "checked schedule", "sent message", "opened webpage", "general query", "saved/noted item"]}53{"pid": "P2", "day": 41, "hour": 13.08, "interactions": 5, "duration": 11.1, "category": "Retrieve", "categories": {"Retrieve": 5}, "actions": ["searched web"]}54{"pid": "P2", "day": 41, "hour": 13.87, "interactions": 10, "duration": 27.2, "category": "Retrieve", "categories": {"Retrieve": 7, "Recall": 2, "Control": 1}, "actions": ["searched web", "general query", "opened webpage"]}55{"pid": "P2", "day": 41, "hour": 15.7, "interactions": 5, "duration": 29.7, "category": "Control", "categories": {"Control": 5}, "actions": ["general query"]}56{"pid": "P2", "day": 42, "hour": 12.4, "interactions": 1, "duration": 1, "category": "Control", "categories": {"Control": 1}, "actions": ["general query"]}57{"pid": "P2", "day": 42, "hour": 16.07, "interactions": 2, "duration": 3.4, "category": "Save", "categories": {"Save": 2}, "actions": ["general query"]}58{"pid": "P2", "day": 42, "hour": 17.93, "interactions": 1, "duration": 1, "category": "Save", "categories": {"Save": 1}, "actions": ["general query"]}59{"pid": "P2", "day": 43, "hour": 18.93, "interactions": 5, "duration": 15.4, "category": "Save", "categories": {"Save": 2, "Communicate": 1, "Shop": 1, "Control": 1}, "actions": ["searched web", "checked schedule", "checked/sent email", "general query", "browsed/added to cart"]}60{"pid": "P2", "day": 44, "hour": 12.95, "interactions": 5, "duration": 8.7, "category": "Retrieve", "categories": {"Save": 1, "Retrieve": 2, "Recall": 2}, "actions": ["searched web", "recalled past info", "general query"]}61{"pid": "P2", "day": 44, "hour": 14.77, "interactions": 1, "duration": 1, "category": "Save", "categories": {"Save": 1}, "actions": ["checked schedule"]}62{"pid": "P2", "day": 44, "hour": 15.43, "interactions": 4, "duration": 8.9, "category": "Save", "categories": {"Save": 1, "Communicate": 1, "Shop": 1, "Control": 1}, "actions": ["browsed/added to cart", "checked/sent email", "general query"]}63{"pid": "P2", "day": 44, "hour": 17.63, "interactions": 5, "duration": 23.7, "category": "Communicate", "categories": {"Communicate": 3, "Save": 2}, "actions": ["checked/sent email", "checked schedule"]}64{"pid": "P2", "day": 45, "hour": 13.32, "interactions": 2, "duration": 4.8, "category": "Retrieve", "categories": {"Retrieve": 1, "Shop": 1}, "actions": ["browsed/added to cart", "general query"]}65{"pid": "P2", "day": 45, "hour": 16.73, "interactions": 5, "duration": 49.0, "category": "Recall", "categories": {"Save": 1, "Recall": 2, "Control": 1, "Retrieve": 1}, "actions": ["saved/noted item", "general query"]}66{"pid": "P2", "day": 46, "hour": 11.12, "interactions": 1, "duration": 1, "category": "Control", "categories": {"Control": 1}, "actions": ["opened webpage"]}67{"pid": "P2", "day": 47, "hour": 14.9, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["searched web", "opened webpage"]}68{"pid": "P2", "day": 50, "hour": 22.35, "interactions": 4, "duration": 5.2, "category": "Save", "categories": {"Save": 4}, "actions": ["searched web", "checked schedule"]}69{"pid": "P2", "day": 50, "hour": 23.98, "interactions": 6, "duration": 14.0, "category": "Shop", "categories": {"Shop": 4, "Recall": 2}, "actions": ["searched web", "opened webpage", "general query", "browsed/added to cart", "saved/noted item"]}70{"pid": "P2", "day": 51, "hour": 13.1, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["searched web"]}71{"pid": "P2", "day": 51, "hour": 14.88, "interactions": 3, "duration": 8.5, "category": "Retrieve", "categories": {"Recall": 1, "Retrieve": 2}, "actions": ["searched web", "recalled past info", "general query", "opened webpage"]}72{"pid": "P2", "day": 52, "hour": 22.28, "interactions": 6, "duration": 30.5, "category": "Retrieve", "categories": {"Retrieve": 6}, "actions": ["searched web", "general query", "opened webpage"]}73{"pid": "P2", "day": 52, "hour": 10.95, "interactions": 3, "duration": 16.1, "category": "Save", "categories": {"Save": 2, "Retrieve": 1}, "actions": ["searched web", "general query", "opened webpage"]}74{"pid": "P2", "day": 52, "hour": 12.1, "interactions": 2, "duration": 1, "category": "Shop", "categories": {"Shop": 2}, "actions": ["browsed/added to cart"]}75{"pid": "P2", "day": 52, "hour": 13.3, "interactions": 6, "duration": 31.2, "category": "Recall", "categories": {"Recall": 6}, "actions": ["recalled past info", "general query"]}76{"pid": "P2", "day": 52, "hour": 16.35, "interactions": 4, "duration": 3.6, "category": "Shop", "categories": {"Shop": 4}, "actions": ["browsed/added to cart", "searched web", "general query"]}77{"pid": "P2", "day": 54, "hour": 20.5, "interactions": 4, "duration": 18.7, "category": "Retrieve", "categories": {"Retrieve": 3, "Save": 1}, "actions": ["searched web", "checked schedule", "opened webpage"]}78{"pid": "P2", "day": 54, "hour": 23.2, "interactions": 7, "duration": 6.9, "category": "Retrieve", "categories": {"Save": 1, "Retrieve": 6}, "actions": ["searched web", "recalled past info", "general query", "opened webpage"]}79{"pid": "P2", "day": 54, "hour": 0.0, "interactions": 2, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 2}, "actions": ["general query"]}80{"pid": "P2", "day": 54, "hour": 10.03, "interactions": 3, "duration": 3.2, "category": "Shop", "categories": {"Shop": 3}, "actions": ["searched web", "browsed/added to cart"]}81{"pid": "P2", "day": 54, "hour": 10.8, "interactions": 2, "duration": 3.5, "category": "Shop", "categories": {"Shop": 2}, "actions": ["browsed/added to cart", "searched web"]}82{"pid": "P2", "day": 54, "hour": 12.73, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["searched web"]}83{"pid": "P3", "day": 50, "hour": 14.15, "interactions": 31, "duration": 145.6, "category": "Save", "categories": {"Save": 26, "Communicate": 3, "Control": 2}, "actions": ["recalled past info", "checked/sent email", "saved/noted item", "general query"]}84{"pid": "P3", "day": 51, "hour": 18.85, "interactions": 3, "duration": 8.2, "category": "Save", "categories": {"Communicate": 1, "Save": 2}, "actions": ["saved/noted item", "general query"]}85{"pid": "P3", "day": 52, "hour": 16.5, "interactions": 10, "duration": 21.8, "category": "Retrieve", "categories": {"Retrieve": 6, "Communicate": 1, "Recall": 2, "Save": 1}, "actions": ["searched web", "recalled past info", "checked/sent email", "general query", "saved/noted item"]}86{"pid": "P3", "day": 53, "hour": 18.35, "interactions": 6, "duration": 26.9, "category": "Save", "categories": {"Save": 3, "Control": 2, "Recall": 1}, "actions": ["checked tasks", "saved/noted item"]}87{"pid": "P3", "day": 53, "hour": 19.43, "interactions": 19, "duration": 48.9, "category": "Save", "categories": {"Recall": 4, "Shop": 4, "Save": 7, "Retrieve": 2, "Communicate": 2}, "actions": ["searched web", "recalled past info", "opened webpage", "sent message", "checked/sent email", "general query", "saved/noted item"]}88{"pid": "P3", "day": 53, "hour": 10.38, "interactions": 5, "duration": 8.4, "category": "Retrieve", "categories": {"Retrieve": 5}, "actions": ["searched web", "checked schedule", "general query"]}89{"pid": "P3", "day": 54, "hour": 19.15, "interactions": 15, "duration": 25.7, "category": "Retrieve", "categories": {"Retrieve": 8, "Save": 6, "Communicate": 1}, "actions": ["searched web", "recalled past info", "checked/sent email", "general query", "saved/noted item"]}90{"pid": "P4", "day": 38, "hour": 23.38, "interactions": 10, "duration": 38.5, "category": "Shop", "categories": {"Control": 3, "Shop": 5, "Communicate": 2}, "actions": ["searched web", "checked tasks", "opened webpage", "checked/sent email", "general query", "browsed/added to cart", "saved/noted item"]}91{"pid": "P4", "day": 38, "hour": 8.78, "interactions": 6, "duration": 25.3, "category": "Retrieve", "categories": {"Communicate": 2, "Retrieve": 3, "Recall": 1}, "actions": ["searched web", "recalled past info", "checked schedule", "checked/sent email", "general query", "saved/noted item"]}92{"pid": "P4", "day": 38, "hour": 10.03, "interactions": 1, "duration": 1, "category": "Control", "categories": {"Control": 1}, "actions": ["general query"]}93{"pid": "P4", "day": 39, "hour": 8.8, "interactions": 5, "duration": 29.1, "category": "Communicate", "categories": {"Communicate": 3, "Control": 1, "Save": 1}, "actions": ["checked/sent email", "checked tasks", "recalled past info", "saved/noted item"]}94{"pid": "P4", "day": 39, "hour": 9.93, "interactions": 1, "duration": 1, "category": "Communicate", "categories": {"Communicate": 1}, "actions": ["sent message"]}95{"pid": "P4", "day": 39, "hour": 10.85, "interactions": 2, "duration": 3.2, "category": "Control", "categories": {"Control": 1, "Retrieve": 1}, "actions": ["searched web", "checked tasks"]}96{"pid": "P4", "day": 40, "hour": 20.7, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["searched web"]}97{"pid": "P4", "day": 40, "hour": 21.37, "interactions": 4, "duration": 2.4, "category": "Communicate", "categories": {"Communicate": 3, "Retrieve": 1}, "actions": ["checked/sent email"]}98{"pid": "P4", "day": 40, "hour": 9.17, "interactions": 11, "duration": 10.4, "category": "Communicate", "categories": {"Communicate": 11}, "actions": ["checked/sent email"]}99{"pid": "P4", "day": 41, "hour": 14.2, "interactions": 3, "duration": 22.5, "category": "Communicate", "categories": {"Communicate": 2, "Retrieve": 1}, "actions": ["checked/sent email", "checked schedule"]}100{"pid": "P4", "day": 41, "hour": 15.92, "interactions": 3, "duration": 12.6, "category": "Retrieve", "categories": {"Save": 1, "Retrieve": 2}, "actions": ["checked schedule", "recalled past info", "saved/noted item"]}101{"pid": "P4", "day": 41, "hour": 17.12, "interactions": 5, "duration": 4.5, "category": "Communicate", "categories": {"Communicate": 5}, "actions": ["checked/sent email"]}102{"pid": "P4", "day": 41, "hour": 17.95, "interactions": 1, "duration": 1, "category": "Communicate", "categories": {"Communicate": 1}, "actions": ["checked/sent email"]}103{"pid": "P4", "day": 44, "hour": 9.1, "interactions": 2, "duration": 1, "category": "Communicate", "categories": {"Communicate": 2}, "actions": ["checked/sent email"]}104{"pid": "P4", "day": 46, "hour": 8.58, "interactions": 8, "duration": 11.3, "category": "Retrieve", "categories": {"Retrieve": 6, "Recall": 2}, "actions": ["searched web", "general query", "opened webpage"]}105{"pid": "P4", "day": 47, "hour": 13.22, "interactions": 1, "duration": 1, "category": "Control", "categories": {"Control": 1}, "actions": ["opened webpage"]}106{"pid": "P4", "day": 48, "hour": 9.27, "interactions": 19, "duration": 43.6, "category": "Communicate", "categories": {"Communicate": 9, "Retrieve": 6, "Recall": 4}, "actions": ["searched web", "checked schedule", "sent message", "opened webpage", "checked/sent email", "general query"]}107{"pid": "P4", "day": 48, "hour": 14.7, "interactions": 3, "duration": 3.3, "category": "Retrieve", "categories": {"Retrieve": 3}, "actions": ["searched web", "general query"]}108{"pid": "P4", "day": 48, "hour": 16.35, "interactions": 1, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 1}, "actions": ["searched web"]}109{"pid": "P4", "day": 49, "hour": 8.47, "interactions": 17, "duration": 34.2, "category": "Retrieve", "categories": {"Retrieve": 11, "Recall": 3, "Communicate": 3}, "actions": ["searched web", "sent message", "checked/sent email", "general query", "saved/noted item"]}110{"pid": "P4", "day": 50, "hour": 8.48, "interactions": 8, "duration": 9.8, "category": "Control", "categories": {"Retrieve": 2, "Control": 3, "Communicate": 3}, "actions": ["searched web", "checked/sent email", "general query"]}111{"pid": "P4", "day": 50, "hour": 12.03, "interactions": 6, "duration": 14.1, "category": "Control", "categories": {"Control": 4, "Recall": 2}, "actions": ["searched web", "general query"]}112{"pid": "P4", "day": 50, "hour": 13.97, "interactions": 8, "duration": 16.4, "category": "Retrieve", "categories": {"Recall": 2, "Retrieve": 4, "Communicate": 2}, "actions": ["opened webpage", "saved/noted item", "general query"]}113{"pid": "P4", "day": 51, "hour": 12.8, "interactions": 2, "duration": 1.1, "category": "Retrieve", "categories": {"Retrieve": 2}, "actions": ["searched web"]}114{"pid": "P4", "day": 52, "hour": 15.05, "interactions": 22, "duration": 19.1, "category": "Retrieve", "categories": {"Retrieve": 14, "Communicate": 8}, "actions": ["checked/sent email", "checked schedule", "general query", "sent message"]}115{"pid": "P4", "day": 53, "hour": 10.42, "interactions": 6, "duration": 3.3, "category": "Retrieve", "categories": {"Communicate": 1, "Retrieve": 4, "Recall": 1}, "actions": ["searched web", "checked/sent email", "checked schedule", "general query"]}116{"pid": "P4", "day": 53, "hour": 15.87, "interactions": 3, "duration": 10.1, "category": "Recall", "categories": {"Recall": 3}, "actions": ["searched web", "general query"]}117{"pid": "P4", "day": 54, "hour": 10.82, "interactions": 12, "duration": 22.7, "category": "Recall", "categories": {"Recall": 6, "Save": 3, "Communicate": 3}, "actions": ["saved/noted item", "recalled past info", "general query", "sent message"]}118{"pid": "P4", "day": 54, "hour": 13.72, "interactions": 3, "duration": 1, "category": "Retrieve", "categories": {"Retrieve": 3}, "actions": ["general query"]}