Claude Code Sycophancy Meter

sycophancy-meter.py

#!/usr/bin/env python3
"""sycophancy-meter: how often does Claude say "you're right"?

Run from a machine with Claude Code conversation transcripts.
Outputs a histogram by day + day-of-week analysis, normalized by message count.

Usage: python3 sycophancy-meter.py [--exclude-today]
"""

import os
import re
import sys
from datetime import datetime, date
from pathlib import Path
from collections import defaultdict

PROJECTS_DIR = Path(os.environ.get("CLAUDE_PROJECTS_DIR", Path.home() / ".claude" / "projects"))
EXCLUDE_TODAY = "--exclude-today" in sys.argv

print(f"Scanning transcripts in {PROJECTS_DIR} ...")

right = defaultdict(int)
absolutely = defaultdict(int)
msgs = defaultdict(int)
today_str = date.today().isoformat()

for jsonl in PROJECTS_DIR.rglob("*.jsonl"):
    for line in open(jsonl, errors="ignore"):
        if '"role":"assistant"' not in line and '"role": "assistant"' not in line:
            continue
        m = re.search(r'"timestamp":\s*"(\d{4}-\d{2}-\d{2})', line)
        if not m:
            continue
        d = m.group(1)
        if EXCLUDE_TODAY and d == today_str:
            continue
        msgs[d] += 1
        if re.search(r"you're right", line, re.IGNORECASE):
            right[d] += 1
        if re.search(r"you're absolutely right", line, re.IGNORECASE):
            absolutely[d] += 1

dates = sorted(right.keys())
days = ["Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun"]

if not dates:
    print("No \"you're right\" found. Your AI is refreshingly honest.")
    sys.exit(0)

max_rate = max(right[d] / msgs[d] * 1000 for d in dates)
scale = 40 / max_rate

print(f"\n{'Date':<12} {'Day':<4} {'#':>3} {'msgs':>5} {'rate':>6}  per 1000 msgs")
print(f"{'':_<12} {'':_<4} {'':_>3} {'':_>5} {'':_>6}  {'':_<40}")
for d in dates:
    total = right[d]
    ab = absolutely[d]
    dt = datetime.strptime(d, "%Y-%m-%d")
    day = days[dt.weekday()]
    m = msgs[d]
    rate = total / m * 1000
    bar_len = round(rate * scale)
    ab_len = round(ab / m * 1000 * scale)
    n_len = max(bar_len - ab_len, 0)
    print(f"{d:<12} {day:<4} {total:>3} {m:>5} {rate:>5.1f}‰  {'░' * n_len}{'█' * ab_len}")

# Day of week analysis
dow_right = defaultdict(int)
dow_msgs = defaultdict(int)
dow_abs = defaultdict(int)
for d in dates:
    w = datetime.strptime(d, "%Y-%m-%d").weekday()
    dow_right[w] += right[d]
    dow_msgs[w] += msgs[d]
    dow_abs[w] += absolutely[d]

active_days = sorted(dow_right.keys())
max_dow = max(dow_right[w] / dow_msgs[w] * 1000 for w in active_days)
ds = 30 / max_dow

print(f"\n{'Day':<4} {'right':>6} {'msgs':>6} {'rate':>6}  per 1000 msgs")
print(f"{'':_<4} {'':_>6} {'':_>6} {'':_>6}  {'':_<30}")
for w in range(7):
    if w not in dow_right:
        continue
    r = dow_right[w]; m = dow_msgs[w]; a = dow_abs[w]
    rate = r / m * 1000
    bl = round(rate * ds); al = round(a / m * 1000 * ds)
    print(f"{days[w]:<4} {r:>6} {m:>6} {rate:>5.1f}‰  {'░' * max(bl - al, 0)}{'█' * al}")

tr = sum(right.values()); ta = sum(absolutely.values())
tm = sum(msgs[d] for d in dates)
print(f"\n░░ you're right: {tr - ta}")
print(f"██ you're ABSOLUTELY right: {ta}")
print(f"   total: {tr} across {tm} messages ({tr / tm * 1000:.1f}‰)")