glpi-neural-brain/AUTONOMOUS-RESEARCH-ROUTING-V7.md

Autonomous Research Routing v7

v7 hardens the three routing points exposed by the v6 real-world run without changing the autonomous-research interval or vector thresholds.

1. Facet-aware primary-source exploration

Broad comparison/integration questions can contain more than one technical entity. A single primary source is no longer required to cover the whole compound question before it may be fetched.

Example:

• facet A: OWASP SAMM
• facet B: MITRE ATT&CK
• integration question: how both can be combined

For a low-relevance search snippet, v7 may fetch at most one primary/authoritative result per recognized entity facet, with at most two such facet probes per research query. The candidate must contain the facet's technical entity terms. Generic official pages do not qualify.

This is only a prefetch exception. The fetched full text still passes the normal model/deterministic evidence assessment, minimum quality threshold, source filter and later article claim review.

New candidate telemetry includes authoritative_facet on the per-result decision. Existing authoritative_exploration_eligible_count and authoritative_exploration_selected_count now count bounded entity-facet probes.

2. Autonomous opportunity nodes are primary article seeds

When an article attempt was requested by the autonomous scanner, production Knowledge nodes carried by the opportunity are treated as required article seeds if they are present in the bounded source pool.

They:

• survive the later topic-coherence filter,
• are prepended to the planner's selected source IDs,
• remain auditable through article.sources.autonomous_seeds.

ANN/cluster search still adds supporting sources, but it no longer replaces the graph nodes that originally caused the knowledge-gap decision.

This rule is scoped to the autonomous synthesis trigger. Other article entry paths keep their previous source-selection behavior.

3. Compute-agent startup grace

The first vector rebuild may happen before a newly started Agent has registered its vector_graph capability. During the initial bootstrap only, v7 now uses the existing BRAIN_VECTOR_GRAPH_AGENT_WAIT as a registration grace period.

Flow:

1. perform the normal immediate capability check;
2. if no compute Agent is visible and bootstrap is still running, publish vector.graph.agent.waiting;
3. poll for the capability until BRAIN_VECTOR_GRAPH_AGENT_WAIT expires;
4. submit the normal compute job when the Agent appears;
5. only then use the existing local fallback when Agent offload is optional.

Steady-state/incremental vector rebuilds do not add this registration wait and retain the fast fallback behavior.

Unchanged by v7

• BRAIN_AUTONOMOUS_RESEARCH_INTERVAL is untouched. A test value such as 10m remains valid.
• Autonomous Research stays independent from the Thinking switch.
• Vector similarity/affinity thresholds are unchanged.
• Orphan-cluster scoring and v6 cohesion rules are unchanged.
• Full-text evidence and article quality gates are not weakened.