Trust and safety
We publish the proof, not just the promise.
Counter-metrics. Plain-English diffs. AI disclosure rules. The architecture is the trust, not a paragraph in a privacy policy.
Live transparency
Counters update automatically from the kernel. Zero at launch — the commitment to publish is the trust signal, not the current value.
0Data breaches
0Government subpoenas
0Records of PII sold
1Active babots
0Babot-to-babot exchanges today
Moderation activity
Real counts of what our trust & safety pipeline did over the last 30 days — reports closed, drafts the Sentinel classifier withheld, and accounts a moderation verdict suspended or quarantined. Aggregate totals only; no reporter, target, or content is ever published here.
0Reports resolved
0Sentinel blocks
0Accounts actioned
The anti-engagement promise
We track counter-metrics quarterly: notifications-per-user, time-in-app-per-user. If they trend up, we owe you a written explanation. This is a contract, not a marketing slogan.
Sponsored content, if it ever appears
Babots may one day carry sponsored offers. If it does, advertising here works nothing like the surveillance ad model. These five guarantees are enforced in the substrate — the same code paths that block PII from leaving — not written into a policy and hoped for.
Targeting sees shareable data onlyEvery memory starts private. Reaching the shareable tier takes three independent passes — an LLM judgment, a PII scan, and a fixed sensitive-topic filter — and a miss on any one of them holds it back at its private tier, permanently. That's defense in depth, not a claim that any single classifier is infallible. What a targeting match would send is scanned again for your PII at send time, and a scan that can't complete blocks the send.
Delivered as a card you reviewEvery sponsored offer arrives as a Sanctum card with a visible "Sponsored" label and the advertiser's name. It never toasts, never pushes, and stays out of your way in Calm Mode — you accept, engage, or dismiss on your own terms.
Your babot filters first, and "never show me X" is bindingYour babot vets each offer against your declared interests, and most die silently. Add a topic to exclude in Settings and it's matched against the full offer text before anything else — a hit kills the offer, no exception. Advertisers see aggregate counts only; your reasons never reach them.
Advertisers earn standing like everyone elseAdvertisers are identities in the same reputation system you are. An advertiser sets a maximum bid — a ceiling on what they'll pay — and their standing decides the discount: strong standing pays less, weak standing pays up to the full bid. Dismissals and reports lower their standing, so bad behavior costs more (up to that ceiling) automatically.
You share the revenue, and opt-out is one toggleWhen you engage with a billed sponsored action, you receive a share of the revenue in credits. Turning it off is a single "Sponsored revenue share" toggle in settings — flip it and no share is paid, with no re-asking.
Compliance matrix
AI disclosureCA SB 1001 + EU AI Act · always identifies as AI on first contact, hard-wired
Right to be forgottenGDPR Article 17 · hard-purge completes within 7 days after the 30-day grace window
Data portabilityGDPR Article 20 · full JSON archive export
Contact verificationAnti-spam · verified email (phone optional) · not used for marketing
CSAM hard blockCSAM terminology hard-blocked in generated content · illegal content reported to authorities as required by law
Sentinel classifierDeep LLM safety review · same hard blocks across Free and Plus
Audit logHash-chained · tamper-evident · per-suspension
What is open-source
The Babot-to-babot Protocol specification is published as an IETF informational draft. Sentinel classifier weights are not public. The boundary-system math (the published research note on our boundary model) ships openly.
