Moderation Constitution
This document governs everything about how Jawwing works: what content is allowed, how posts are ranked in your feed, how your location data is protected, and what AI technology makes decisions. It is public, versioned, and subject to community amendment. No hidden algorithms. No secret rules.
Jawwing exists to give people a voice in their communities, free from identity-based suppression, human moderator bias, or opaque algorithms. The feed ranking, moderation rules, privacy protections, and AI technology are all documented here. These rules are public, versioned, and yours to change.
Anonymity is sacred. No action may be taken to identify, expose, or infer the identity of any user.
Speech is presumed valid. Content is allowed unless it violates a specific rule. Ambiguity resolves in favor of the poster.
Transparency is mandatory. Every moderation action is logged publicly with the rule cited and the AI agent's reasoning.
No human override. Human employees may not override AI moderation decisions. They may only propose constitutional amendments.
Open algorithm. The ranking algorithm is public. No shadow bans, no secret suppression, no hidden boosts. Every factor is documented in this constitution.
Direct threats. Content that directly threatens physical harm to a specific, identifiable person or group.
Non-consensual intimate content. Explicit content involving real people without consent.
CSAM. Any sexualized content involving minors. Zero tolerance. Immediate removal and report to NCMEC.
Doxxing. Posting private identifying information (home address, phone, workplace) without consent.
Spam. Automated or repetitive content designed to flood the feed or manipulate vote counts.
Sexually explicit content. Nudity, pornography, sexual solicitation, and graphic sexual descriptions. This platform is not for adult content.
Political speech, including criticism of governments, parties, or officials.
Satire and parody, clearly or arguably so.
Profanity and crude humor.
Unpopular opinions and minority viewpoints.
Criticism of Jawwing itself.
Automated review. Every post is reviewed by AI within seconds of submission. The AI evaluates content against all Prohibited and Restricted rules and returns a decision (approve, flag, warn, or remove) with a confidence score from 0% to 100%.
Confidence threshold. Removals require 70%+ AI confidence. Below this threshold, the post remains live and is downgraded to a flag for secondary review. The AI confidence score is displayed on every post for full transparency.
Appeals. Any user may appeal a moderation decision. Appeals are reviewed by a separate, independent AI agent instance that re-evaluates the content fresh against the same constitution.
Community override. If 500+ unique users flag a post, it triggers a mandatory re-review regardless of the original AI decision.
Image moderation. Posts with images are evaluated multimodally. The AI reviews both the text and the image together for NSFW content, violence, personally identifying information, CSAM, and other violations.
Spam detection. Duplicate content is detected via content hashing. Identical posts within a 10-minute window are flagged as spam (R-2). Rate limits cap posting at 20 posts per hour per identity.
Escalating cooldowns. When a post is removed for violating the Constitution, the poster receives an escalating posting cooldown. Strike 1: 5 minutes. Strike 2: 30 minutes. Strike 3: 2 hours. Strike 4: 12 hours. Strike 5+: 24 hours. Strikes are counted over a rolling 24-hour window. Cooldowns apply to both anonymous and authenticated users.
Rejection transparency. Users can view their rejected posts and the AI's explanation for why each post was removed. This is available at /my-posts (authenticated) or via the API. The moderation reasoning is never hidden from the poster.
Evasion detection. The AI moderator is trained to detect common evasion techniques including but not limited to: character substitution (leetspeak, unicode lookalikes), coded language, spaced letters, context tricks ('just a joke'), and obfuscation. Attempting to evade moderation is itself a violation and may result in an immediate strike.
Three sort modes. HOT (default) ranks posts by a combination of quality, engagement, and freshness. NEW shows posts in reverse chronological order. TOP ranks by total engagement volume.
HOT ranking formula. Each post's HOT score is calculated as: (Wilson Score + 0.3 * log(1 + total_votes) + 0.2 * controversy_ratio) / (age_hours + 2)^1.2. Wilson Score is a statistical lower-bound on the true upvote ratio, penalizing posts with few total votes. This means a post with 5 upvotes and 0 downvotes ranks lower than one with 100 upvotes and 10 downvotes.
Controversy bonus. Posts with high engagement on BOTH sides get a boost, not a penalty. A post with 1000 upvotes and 999 downvotes is treated as highly engaging content, not buried at score 1. The controversy ratio (min(ups, downs) / max(ups, downs)) rewards posts that spark genuine debate.
Time decay. All posts lose ranking over time via gravity factor (age_hours + 2)^1.2. A post that was hot 12 hours ago will naturally fall as newer content arrives. Posts expire from public feeds after 30 days but remain saved to the poster's account history.
Distance weighting. In HOT mode, closer posts rank higher at equal engagement. The distance boost formula is: 1 / sqrt(1 + distance_km / scale). The scale varies by scope: LOCAL uses scale=2 (strongly favors nearby), METRO uses scale=10, COUNTRY uses scale=100. A post 1km away gets roughly 1.4x the boost of a post 5km away in LOCAL scope.
Three scope levels. LOCAL (5km radius from your location) shows your immediate neighborhood. METRO shows your entire metro area. COUNTRY shows posts from everywhere. The feed auto-expands from local to metro to country until at least 10 posts are found, so you never see an empty feed.
No suppression. The algorithm has no concept of shadow bans, reach reduction, or content suppression. If a post is not removed by moderation, it competes equally in the ranking. There is no boost for accounts, no penalty for new users, no advertiser influence. The only inputs are: votes, time, distance, and engagement.
TOP ranking. TOP sort ranks posts by total engagement: upvotes + downvotes. This surfaces the most-discussed content regardless of whether sentiment is positive or negative.
GPS required to post. You can browse, vote, and comment from anywhere, but creating a post requires GPS to confirm you are physically in that area. This is the core identity of Jawwing: local voices, not remote trolls.
Coordinate fuzzing. Exact GPS coordinates are never stored. All locations are fuzzed to approximately a 1-mile radius before being saved to the database. The API returns rounded coordinates only. This prevents anyone from pinpointing where a post was made.
Distance ranges. Distances shown on posts use vague ranges (NEARBY, <5MI, <10MI) rather than exact values. This prevents triangulation of a poster's location from multiple reference points.
IP hashing. Your IP address is immediately passed through a one-way cryptographic hash (HMAC-SHA256). The raw IP is never stored. The hash cannot be reversed, even by us. It is used only for rate limiting and vote deduplication.
Anti-spoofing. Location spoofing is detected via null island checks (0,0 coordinates) and teleportation detection (impossible movement speed between consecutive posts from the same identity).
Current model: Claude Haiku 4.5 by Anthropic. Selected for speed, safety-first design, vision capability (image moderation), strong instruction following, and published safety documentation. Anthropic is an AI safety company — their models are designed to be helpful, harmless, and honest.
Model requirements. Any moderation AI must: (1) process posts within 2 seconds, (2) maintain >95% agreement with human reviewers on test sets, (3) cost under $0.001 per decision at scale, (4) have published safety documentation from its provider, (5) not create single-provider lock-in.
Model changes. Any change to the moderation model requires 7 days public notice before deployment. Old and new model test results must be published side-by-side for community review.
Audit commitment. Monthly publication of moderation accuracy statistics: false positive rate, false negative rate, and appeal overturn rate. Available on the transparency page.
Any user may propose an amendment by submitting it through the app.
Amendments require 60% approval from active users (min. 1,000 votes) over 7 days.
No amendment may reduce transparency requirements or grant humans moderation authority.
The feed algorithm (Article V) is subject to amendment like any other article. If the community votes to change the ranking formula, it changes.
Video content may only be linked from platforms with established moderation policies. Users may share links to videos on these platforms. Native video upload is not supported.
AV-1Links to video hosting platforms not on this list will be flagged for review. This list can be amended through the standard amendment process.
Want to propose an amendment?
constitution@jawwing.com