The Numbers
Generated first-party operational snapshot from the ThumbGate runtime. This is not customer traction, install volume, revenue, or proof that a configured gate has fired.
Updated: 2026-11-12 · Version 1.40.5
Read this first: configured checks are inventory. Recorded blocks and warnings are usage evidence. This snapshot currently reports 8,602 recorded hard-block event(s) and 2 recorded warning event(s).
Gate enforcement
Configured checks
47
43 shipped defaults · 4 auto-promoted; inventory, not usage
source: gate-stats.js
Top recorded blocker
auto-surface-outbound-email-send (7 blocks)
highest recorded block count
source: gate-stats.js
Token & time savings
Estimated LLM dollars saved
$138
blended Sonnet/Opus/Haiku 80/15/5 mix from recorded blocks
source: token-savings.js
Tokens not spent
22.4M
2,000 input + 600 output per recorded block, conservative
source: token-savings.js
Methodology
Where the numbers come from. This page is regenerated from local scripts — no survey data, no hand-edited figures, no third-party attribution. Every number on this page is produced by code in the public ThumbGate repo.
- Configured checks — union of shipped default rules and the auto-promotion ledger. This is inventory, not proof that the check fired.
- Actions blocked/warned — recorded intervention counts from gate occurrence data. These are the evidence counters.
- Hours saved — conservative 15-minute/incident estimate, shown only as a function of recorded block/warn counts.
- Dollars saved — blended per-call token estimate (2k input + 600 output) × recorded blocks × 2026-04-15 Anthropic + OpenAI list prices. See
scripts/token-savings.jsfor the full price snapshot. - Bayes error rate — shown only when the feedback sample contains both safe and harmful outcomes. One-class samples are reported as n/a.
Last auto-promotion: auto-history-aware-distillation-linked-this-failure-to- on 2026-11-11. Regenerated on every release via npm run numbers:generate and on a weekly cadence.