A short comparison, the relevant contract references and a clear note where evidence is missing.
Request lifecycle / 02
What happens between a question and a supported answer
Each request is compiled into obligations, authorized against policy, routed through only the capabilities it needs, checked against evidence and returned with an explicit outcome.
A realistic request
“Compare the renewal clauses in our current supplier contracts.”
Which contracts this user may access, where their text may go, which model may be used and what counts as enough support.
01
Decision sequence
Seven explicit control points surround model execution.
The named modules make each responsibility visible. The exact implementation may vary by deployment, but the contract and control points remain explicit.
- 01INPUT
The product sends the request
A user asks: “Compare the renewal clauses in our current supplier contracts and flag anything the documents do not prove.” The request arrives with the user's authorized tenant, workspace and data scope.
- 02CONTRACT
Mentaview defines a complete answer
The Question Contract Compiler separates the task into contracts, renewal clauses, differences and missing evidence. This makes omissions visible before an answer is written.
- 03PLAN
It chooses a bounded plan
The Cognitive Engine sees that private documents are required. It authorizes retrieval and answer checks, but does not add fresh web research because the request is about internal contracts.
- 04EVIDENCE
Only allowed knowledge enters the path
Document Retrieval searches within the permitted sources. The Confidentiality Boundary can replace sensitive values with scoped aliases before an approved external model sees the text.
- 05INFERENCE
A qualified model does the language work
The Inference Gateway sends only the approved packet to the selected provider or local model. Model choice stays separate from product policy.
- 06CHECK
The answer is checked before publication
Answer Assurance tests whether every requested part is covered and whether important statements point back to adequate evidence. Bounded repair may address a specific gap.
- 07OUTPUT
The product receives an explicit result
The caller gets a supported comparison with references and an operator trace. If the contracts do not prove a point, the result says so instead of inventing certainty.
02
Conditional routing
A different question can take a different route.
“Minimal” means no unnecessary capability. “Sufficient” means the system still has to meet the request's declared obligations.
A simple product question
Direct handling may be enough. Retrieval and research can stay off.
A question about private records
Use authorized retrieval, scoped memory and confidentiality controls where policy requires them.
A question needing current public facts
Add external research with source identity, time and evidence checks.
A disconnected or local device
Use locally available capabilities or return a clear limit when required evidence is unavailable.
03
What the product can expose
An answer is only one part of the result.
Product teams choose what end users and operators see. Mentaview provides explicit material for both audiences without publishing private internals.
The answer
A response shaped around the full request, with unsupported parts clearly marked.
The references
Named evidence that lets the product show where supported statements came from.
The route
Which capabilities were used, skipped, repaired or stopped — without raw private content in the trace.
The limits
Missing evidence, budget exhaustion, policy refusal or another reason the system could not safely complete the task.
04
Qualification boundary
A clear architecture still needs real-world qualification.
Mentaview's current labels are deliberate. A controlled pilot or canary is evidence of bounded work, not a claim of universal production readiness.
- Models can still be wrong.Planning and checks reduce some failure modes; they do not remove the need to qualify models for the actual workload.
- Evidence can still be missing or poor.A governed system should expose that problem, not manufacture support.
- Policies need owners.A team must define its permissions, provider rules, retention, budgets and acceptance criteria.
- Deployment evidence is specific.Local measurements, sealed tests and canaries must not be generalized into a production service level.
Module ownership
Each decision belongs to a defined module.
Eleven modules separate planning, evidence, privacy, memory, inference, edge operation and evaluation behind explicit contracts and promotion gates.