Add optional LISA expert-based advanced ad detection #8

New issue

Open

opened 2026-09-11 05:17:22 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-09-11 05:17:22 +00:00

(Migrated from github.com)

Parent: #1
Depends on: #4, #6

Goal

Add an optional expert-system layer using LISA for cases where deterministic filter lists are insufficient, while keeping classic rule-based blocking authoritative and inspectable.

Scope

Feed page/request observations into a LISA fact model: DOM/layout signals, resource URLs, element attributes, iframe/context information, script/resource behavior, known ad patterns, and other bounded evidence available from Nyxt.
Define experts/rules that classify likely ads, trackers, overlays, sponsored widgets, and related unwanted content.
Combine LISA results with the Star-Lang decision pipeline through an explicit policy boundary rather than hidden side effects.
Support confidence/severity thresholds and a dry-run/log-only mode.
Keep expert decisions explainable: record matched expert/rule and supporting facts.
Allow the expert layer to be enabled/disabled independently of ordinary filter-list blocking.
Expose configuration hooks that the settings page can surface later.

Safety/performance constraints

Deterministic filter-list matches should not require running expensive expert analysis.
Expert evaluation must have bounded execution and must not stall browsing.
Low-confidence detections should be configurable as report-only rather than automatically blocked.
Do not train or silently mutate expert rules from browsing data; rule updates must be explicit/versioned.

Acceptance criteria

LISA can classify a fixture set containing obvious ads plus difficult cases not matched by static lists.
Every expert block has an explanation and provenance.
The feature can run in log-only mode for evaluation.
It can be switched off without affecting normal Star-Lang/filter-list behavior.
Benchmarks establish acceptable page/request overhead before it becomes enabled by default anywhere.

Parent: #1 Depends on: #4, #6 ## Goal Add an optional expert-system layer using LISA for cases where deterministic filter lists are insufficient, while keeping classic rule-based blocking authoritative and inspectable. ## Scope - Feed page/request observations into a LISA fact model: DOM/layout signals, resource URLs, element attributes, iframe/context information, script/resource behavior, known ad patterns, and other bounded evidence available from Nyxt. - Define experts/rules that classify likely ads, trackers, overlays, sponsored widgets, and related unwanted content. - Combine LISA results with the Star-Lang decision pipeline through an explicit policy boundary rather than hidden side effects. - Support confidence/severity thresholds and a dry-run/log-only mode. - Keep expert decisions explainable: record matched expert/rule and supporting facts. - Allow the expert layer to be enabled/disabled independently of ordinary filter-list blocking. - Expose configuration hooks that the settings page can surface later. ## Safety/performance constraints - Deterministic filter-list matches should not require running expensive expert analysis. - Expert evaluation must have bounded execution and must not stall browsing. - Low-confidence detections should be configurable as report-only rather than automatically blocked. - Do not train or silently mutate expert rules from browsing data; rule updates must be explicit/versioned. ## Acceptance criteria - LISA can classify a fixture set containing obvious ads plus difficult cases not matched by static lists. - Every expert block has an explanation and provenance. - The feature can run in log-only mode for evaluation. - It can be switched off without affecting normal Star-Lang/filter-list behavior. - Benchmarks establish acceptable page/request overhead before it becomes enabled by default anywhere.