public-content-scrubber

What this catches

The privacy posture is Stripe Press / Linear / Anthropic — quality of artifact in public, financial detail private. This scrubber is the structural enforcement of that posture across every public surface (newsletters, ship logs, sales pages, founder letters, Reddit/LinkedIn drafts, brief issues, site copy).

It blocks four categories of accidental leak:

1. Dollar amounts — $5,000, $1.2M, $500/mo (unless the trailing context names a known pricing/methodology word like max position, daily exposure, starter, pro, tier)
2. Financial terms — MRR, ARR, revenue, P&L, profit, commission, bankroll, fills, win rate, edge realized, drawdown, Sharpe (unless context contains target, spec, threshold, min, max, goal, floor, cap — these signal methodology, not actual performance)
3. Customer counts — exact figures: "47 paying users", "200 subscribers"
4. Trading profile URLs — kalshi.com/profile/<user> (track record stays NDA-only)

When to use

Wired into every public-output skill BEFORE publish:

• newsletter-weekly-publish
• ship-log-publish
• founder-letter-publish
• sales-page-edit
• product-page-edit
• reddit-outreach-paced
• linkedin-outreach-paced
• brief-publish
• site/index.html or any site/* edit

Pattern: drafter generates → scrubber scans → on hit, drafter rewrites → scrubber scans again → on clean, voice-gate runs (advisory) → publish.

Procedure

python3 ${REPO_ROOT}/tools/voice/scrubber.py "$DRAFT_PATH"
EXIT=$?
if [ $EXIT -eq 0 ]; then
  echo "STATUS=clean"
elif [ $EXIT -eq 1 ]; then
  echo "STATUS=blocked"
  echo "$(cat /dev/stdin)" | python3 ${REPO_ROOT}/tools/voice/scrubber.py - > /tmp/scrubber-hits.json
  # Surface hits to drafter for revision
fi

Exit 0 = clean, exit 1 = hits found (block publish), exit 2 = bad input.

Hard rules

1. Never bypass. Every public output goes through. False positives are acceptable; false negatives are unrecoverable.
2. Cooper override only. If the drafter believes a hit is a false positive, the workflow opens an approval gate OVERRIDE_SCRUBBER_<reason>. Cooper YES → publish proceeds with the hit on record. NO → revise.
3. Keep updating allowlists conservatively. New legit pricing tiers / methodology values get added to tools/voice/scrubber.py PRICING_CONTEXT regex. Each addition is a commit Cooper reviews.
4. Pair with voice-gate. The scrubber blocks; voice-gate advises. Both run; both findings surface.

Calibration evidence (regression tests)

Verified by 7 calibration tests in tools/voice/scrubber.py:

• Path E parameters + product pricing tiers → clean
• "win rate target ≥75%" methodology language → clean
• "$5,000 MRR" → blocked (dollar + financial term)
• "47 paying users" → blocked (customer count)
• "kalshi.com/profile/cooper" → blocked (trading URL)
• "Last week trading P&L grew" → blocked (financial term)
• Plain prose with no figures → clean

The companion tools/voice/score.py runs in parallel as advisory voice rubric.