glpi-neural-brain/CHANGELOG-ARTICLE-CREATION-GATES.md

Changelog: Article Creation Gates

Problem

The article pipeline could collect and persist strong research evidence but still create no staging article. Four independent behaviors amplified each other:

1. model-generated topic labels such as site:digital-forensics were used as if they were real domains;
2. legacy missing_information and editorial refinements were promoted to critical blockers;
3. every article type used the same answer-length gate;
4. several skip paths, especially deterministic draft validation and duplicate detection, emitted no terminal diagnostic event.

Changes

Valid search-domain restrictions

• preferred_domains now accepts only real fully-qualified hostnames.
• Schemes, paths and www. are normalized.
• IP addresses, one-label topic names and malformed hostnames are rejected.
• Invalid site: tokens already present in model-generated queries are removed before SearXNG is called.
• At least one unrestricted language variant remains available as before.

Examples:

• accepted: nist.gov, docs.aws.amazon.com, https://www.nist.gov/path
• rejected: digital-forensics, assetmanagement, cleanroomrecovery

Narrower critical-gap classification

• The knowledge-brief prompt now requires a concrete consequence for every critical gap: a false conclusion, safety risk or non-executable core procedure.
• Definitions, comparisons, examples, screenshots, variants and editorial depth are optional by default when a grounded usable core already exists.
• Deterministic normalization downgrades over-classified editorial gaps only when at least three grounded statements remain.
• Missing rollback information, safety prerequisites, permissions, parameters, commands, validation steps and comparable operational blockers remain critical.
• Legacy missing_information is no longer automatically critical; it is classified by the same blocker rules.
• Research queries are rebuilt only from remaining critical gaps and unresolved critical contradictions.

Article-type-aware draft validation

• how_to and troubleshooting retain the configured full answer minimum.
• concept and reference use a lower answer-length minimum but require at least two grounded key points.
• decision_guide uses a proportional minimum and requires at least two decision criteria.
• Confidence, productive-source ratio, generation depth and problem-description checks remain enforced for every type.

Complete diagnostics

All article skip/rejection paths now publish an explicit event. After article.draft.started, the pipeline terminates with exactly one of:

• article.created
• article.draft.rejected
• article.duplicate
• article.failed

Deterministic validation failures expose structured metadata:

• reason
• field
• actual
• required
• article_type

The activity UI displays these values directly.

Compatibility

No new environment variables are required. Existing thresholds remain valid; their interpretation is now article-type aware.