External Research activates only after the system identifies a current-information gap, local evidence is insufficient and policy explicitly permits network discovery and source retrieval.
External Research
Search only when authorized local evidence is insufficient.
Plans bounded discovery, validates targets, recovers canonical public documentation, performs safe source collection, preserves provenance and queues accepted evidence for retrieval.
A bounded collection of acquired public evidence with validated targets, source provenance, research receipts and a clear account of rejected, unavailable or unverified material.
Why it exists
The problem this module is designed to solve.
Many questions depend on current public information, but unrestricted browsing introduces egress, security, latency, provenance and cost risks. Search results alone are not evidence, and a hidden search-vendor dependency undermines architectural freedom.
02 / How it works
A bounded path from need to accountable result.
The public model below describes responsibilities and decisions, not sensitive implementation details, provider secrets or customer data.
- 01
Describe the missing evidence
Turn the unresolved answer facet into a bounded research request instead of issuing an open-ended browsing instruction.
- 02
Authorize budget and targets
Apply egress, time, result-count and target-safety constraints before any external request is attempted.
- 03
Discover and retrieve
Use replaceable discovery and retrieval adapters, rank owner-bound official sources, follow only corroborated query-aligned documentation or PDF links, and keep failed or unsafe targets outside the evidence set.
- 04
Normalize and hand off
Preserve public-source provenance, verify that each answer facet is supported locally, record what happened and pass accepted evidence into the same retrieval and assurance path as local sources.
02B / Qualified production profile
Public evidence, recovered under a hard boundary.
The release profile expands source coverage without granting open-ended browsing. Discovery, canonical-document refinement, scraping and answer publication each remain separately bounded and observable.
Execution envelope
More reach, with every extra step still earned.
The exact candidate admits up to 20 ranked results, 10 scrapes and six same-origin link refinements. A two-source success quorum, 500 ms tail grace, 10-second per-scrape timeout and 12-second global research wall budget cap the work.
Canonical pages before incidental matches
Owner-bound ranking and independently corroborated, query-aligned documentation or PDF links can recover primary material buried behind a portal page.
Evidence must support the requested relation
Strict grounding checks subject and relation locally, preserves exact source references and abstains instead of promoting an unsupported facet.
Exact lists and multi-facet requests stay finite
Adjacent canonical chunks can close a declared inventory; strict recovery inspects a bounded ranked window and never turns into an unbounded crawl.
Unsafe or unavailable is not evidence
Connect-time DNS/IP checks, redirect validation, robots policy, byte limits, deadlines and explicit partial outcomes keep rejected material outside the packet.
03 / Customer and operator value
Adds current public evidence without turning unrestricted browsing or a specific search vendor into a core dependency.
Egress and 12-second research budget
Connect-time DNS and IP-scope enforcement
Canonical documentation and PDF recovery
Ten-language Wikimedia search and resolution transport
Strict source-first evidence handoff
04 / Where it creates value
Concrete situations, not generic feature claims.
These are representative product situations. Every deployment still requires its own policy, data boundary and acceptance criteria.
Regulatory and market horizon
Collect current public material under explicit scope and show which sources support each reported change.
Technical and supplier verification
Validate current product, standards or supplier information without letting one search provider become part of the cognitive core.
Current-fact verification
Escalate from local knowledge to bounded Web evidence when freshness is necessary and abstain when reliable source retrieval fails.
05 / Role in the cognitive system
A clear responsibility creates a trustworthy boundary.
No module is allowed to become an invisible monolith. It owns a narrow contract, composes with named capabilities and refuses responsibilities that belong elsewhere.
What it owns
- Research request
- External search and scrape budget
- Target safety
- Network transport
- Public evidence provenance
What it composes with
What it refuses to own
Boundary before convenience.
It does not publish answers, write memory or require a mandatory third-party search service.
06 / Vision and mission
Fresh evidence under explicit authority
This module extends accountable action beyond local knowledge while keeping the network optional, the provider replaceable and every external step visible in the decision path.
AI systems that remain useful, inspectable and sovereign across models, providers and deployment boundaries.
Build the cognitive layer that chooses the smallest sufficient path and turns evidence into accountable action.
Capture the value of advanced AI without surrendering data control, architectural freedom or intellectual honesty.
07 / Evidence and maturity
What the current label means — and what remains open.
Passed with moderate confidence
A bounded, disabled-by-default network research path is integrated and locally exercised with adversarial transport checks; live provider operations and staged traffic remain gated.
- The prospective 500-case release run has not been completed.
- Provider terms, inspected egress, monitoring and staged live traffic remain open.
- The transport cannot attest every connected-socket and TLS-handshake property.
What exists today
The frozen 2026-09-04 production-candidate holdout passed 36/36 answer contracts: semantic, provenance and citation checks each passed 36/36, with zero orphan citations, budget failures or research wall-budget exhaustion. Research p95 was 6.891 seconds and maximum 7.645 seconds under the 12-second budget. A separate explicit ten-language Wikimedia flow supplies bounded candidate evidence to Memory review without granting Memory ownership of egress or provider search. A separate 500-case campaign remains the broader relevance baseline.
Canary
A restricted test profile and supporting evidence exist. Results apply only to the stated profile and must not be generalized to other workloads or deployments.
What must earn promotion
Complete the prospective 500-case release run on this exact profile, approve provider terms and inspected egress, attach dashboards and alerts, then earn promotion through staged live traffic. Until then, network research stays disabled by default.
Public truth boundary: M4− is a non-standard Mentaview engineering label for internal laboratory validation. It is not an official TRL decision and does not assert production qualification, customer acceptance, independent assurance, certification or universal performance. Inspect the complete assessment record.