Ethical ad filter — Taxonomy v1
- Version: 1.0
- Status: Accepted (subject to legal review via
policy-mapping-template.md)
- Last updated: 2026-05-06
- Linear: TIL-70
- Depends on: ADR-001 MVP slice (TIL-69) — v1 primary client is Browser / WebView; taxonomy tiers are defined so they behave predictably there. DNS/VPN-style clients are a future track.
This document defines high-risk ad categories (user-configurable sensitivities) and enforcement tiers. The product does not infer mental health status; optional logging is consent-gated (see evidence-bundle-v1.md, suslink-integration-spike.md).
Enforcement tiers (normative)
Tiers are ordered by strictness for merge logic: block > blur > allow_log > allow (see merge rule below). Schema enum: policy-artifact.v1.schema.json / evidence-bundle.v1.schema.json.
block
| Aspect |
Definition |
| User-visible |
Ad or slot is not rendered. Optional neutral placeholder (“Hidden by your settings”) and policy/category ID for explainability. |
| Gating |
Use when policy + signals meet high confidence and high harm thresholds and the implementation can apply the tier without breaking the host page (WebView/DOM-capable path). |
| Logging |
May emit a minimal structured event (category ID, tier, coarse URL or hashed identifier) only where the user has enabled accountability/research exports or product analytics that are disclosed in consent. Never log full page HTML by default. |
| MVP client note |
Natural fit for WebView / controlled browser (remove/hide node, swap creative). |
blur (warn / soft-block)
| Aspect |
Definition |
| User-visible |
Creative obscured (blur, collapsed card, or overlay) with friction to reveal (e.g. second tap + short “why” copy tied to taxonomy ID). |
| Gating |
Uncertain classifier confidence, urgency/shame patterns, or SusLink elevated but not critical risk—reduces false positives vs block. |
| Logging |
Same consent rules as block; optional allow_log-style watchlist if user enables “log blurred impressions.” |
| MVP client note |
Requires DOM or graphics surface control. On DNS-only clients (future track), there is no blur UX—treat blur as block at the resolver unless a companion Web surface implements blur. |
allow_log (“allow + log”)
| Aspect |
Definition |
| User-visible |
Ad shown normally; user is not interrupted. |
| Gating |
Low-confidence matches, broad keywords, “watchlist” categories, or SusLink suspicious tier when the product policy says observe-first. |
| Logging |
Structured metadata (category, tier, URL or allowed hash, timestamps) recorded only if the user opted into reporting / research / evidence bundle flows. If reporting is off, behavior matches allow from the user’s perspective. |
| MVP client note |
Primary hook for accountability exports without blocking. |
allow
| Aspect |
Definition |
| User-visible |
Ad shown; no tier-based friction. |
| Gating |
Explicit user allowlist entry, safe default when no rule matches, or engine decision to take no action. |
| Logging |
No taxonomy-driven logging unless a separate user opt-in explicitly covers it (do not silently upgrade allow to allow_log). |
Merge and overrides
- Strictest wins: For one impression, the effective tier is the most restrictive among: matched category rules, SusLink-derived bumps, and engine defaults.
- User allowlist overrides
block / blur for listed domains or advertisers where product policy and store rules permit—document exceptions (e.g. illegal content) in policy-mapping-template.md.
- SusLink elevation tables live in
suslink-integration-spike.md; they refine but do not replace taxonomy categories.
Category index (v1)
Seven high-risk categories (within the 3–7 deliverable range). IDs are stable CAT-xx keys for policy JSON.
| ID |
Short name |
Summary |
Default tier |
| CAT-01 |
Predatory lending & BNPL stress |
High-cost credit, payday-style, BNPL framed as frictionless |
blur |
| CAT-02 |
Gambling & sweepstakes urgency |
Licensed/unlicensed gambling, urgency / loss-chasing framing |
blur / block |
| CAT-03 |
High-pressure retail & scarcity |
False scarcity, manipulative countdowns, aggressive upsell |
blur |
| CAT-04 |
Weight / body image shock |
Shock imagery, miracle weight claims, risky drug ad framing |
blur |
| CAT-05 |
Alcohol & substances (sensitivity) |
Alcohol, cannabis, vaping—user opt-in sensitivity |
blur |
| CAT-06 |
Recovery / disorder-adjacent triggers |
Self-declared triggers only; never automated diagnosis |
blur |
| CAT-07 |
Deceptive lead-gen & phishing-adjacent |
Scam-like destinations; SusLink-heavy |
blur / block |
Category definitions (v1)
CAT-01 — Predatory lending & BNPL stress
- Intent: Ads that push high-cost credit, payday-style offers, or BNPL framed as “free” or “instant” without clear terms.
- Default tier:
blur; elevate to block when combined with SusLink high / critical or denylist domains.
allow_log: Broad finance keywords pending classifier tuning.- Signals (examples): Keywords (locale-specific lists), regulated financial claims, landing domains on policy denylists, partner merchant IDs.
CAT-02 — Gambling & sweepstakes urgency
- Intent: Casino / sportsbook / sweepstakes with urgency (“limited time”, “bonus expires”), loss-chasing framing, or unlicensed gambling surfaces.
- Default tier:
blur for urgency patterns; block for unlicensed / denylist domains.
allow_log: Ambiguous “sweepstakes” creative without destination resolution.- Signals: Advertiser categories (where available), destination domain lists, creative OCR for urgency phrases, SusLink impersonation / typosquat heuristics.
CAT-03 — High-pressure retail & scarcity loops
- Intent: Manipulative scarcity (“only 2 left”), false countdowns, aggressive upsell.
- Default tier:
blur.
allow_log: Known templates pending confirmation.- Signals: Creative templates, DOM hints in WebView, policy JSON bad-pattern lists.
CAT-04 — Weight loss / body image shock
- Intent: Before/after shock imagery, miracle claims, prescription drug ads without proper framing.
- Default tier:
blur.
allow_log: Borderline wellness ads for accountability sampling.- Signals: Optional on-device CV/OCR flags, advertiser category, policy keywords.
CAT-05 — Alcohol & substances (sensitivity)
- Intent: Alcohol, cannabis, vaping ads for users who opt into this sensitivity pack.
- Default tier:
blur when the user enabled the category.
allow: When category is off in user settings (no extra logging unless user opts into cross-category research).- Signals: IAB/content taxonomy, landing classification, regional rulesets.
CAT-06 — Recovery / disorder-adjacent triggers (user-declared)
- Intent: Content users flag as harmful to recovery (gambling triggers, diet culture, etc.) via self-declared tags only.
- Default tier:
blur when matched.
- Non-goals: No automated inference of diagnosis, mood state, or relapse risk.
- Signals: User-selected sub-tags only.
CAT-07 — Deceptive lead-gen & phishing-adjacent
- Intent: Destinations that are scam-like, impersonation, or phishing—overlaps SusLink.
- Default tier:
block when SusLink critical; blur when high or heuristic-only match.
allow_log: suspicious SusLink or heuristic weak signals for watchlists.- Signals: SusLink
LinkScanResult, domain trust feeds, local heuristics.
Out of scope for v1 taxonomy
- Political / issue ads (separate policy pack).
- Generic “all ads off” as the default story—positioning remains ethical filtering unless the user explicitly chooses broader blocking where legally permitted.
- TLS MITM or device-wide inspection of third-party apps (see ADR-001 non-goals).
Versioning
- Taxonomy major changes bump
taxonomyVersion in policy packs; new categories receive new CAT-xx IDs; deprecated IDs remain documented for replay of old evidence bundles.