Customer Evidence
Discovery, configuration, logs, code, cloud inventory, relationships, documents, and other customer-authorized evidence remain tenant-scoped inside Forge.
What this customer actually has
Forge Reasoner + AI-driven Knowledge
AI that knows
what it knows.
Forge combines customer-authorized evidence, curated enterprise knowledge, and a controlled reasoning layer so answers can be traced back to what was actually observed or retrieved. When the evidence is not strong enough, Forge is designed to say so instead of filling the gap.
The foundation
Three layers. One evidence boundary.
Forge separates customer facts, reusable knowledge, and reasoning so a general technology document can never silently become a claim about a customer's environment.
→
Verified Knowledge
Curated Forge knowledge and approved external references can be retrieved when relevant, with provenance, source identity, freshness, and confidence retained.
What trusted sources say
→
Forge Reasoner
The Reasoner evaluates evidence and knowledge together, labels inference, detects gaps or conflicts, and can withhold an unsupported conclusion.
What Forge concludes — and why
Grounding model
Facts do not become assumptions.
Every AI-supported answer is designed around three truth levels that stay visibly distinct.
Observed Customer Evidence
Directly supported by customer-authorized evidence. This is the strongest basis for statements about a customer's actual environment.
Retrieved Verified Knowledge
Supported by curated documentation or approved reference material. Useful for explaining technology, behavior, requirements, and known patterns.
Inference / Assumption
A reasoned conclusion rather than an observed fact. Forge should label it, explain its basis, and avoid presenting it as verified evidence.
Customer protection
Safety is part of the reasoning path.
The goal is not merely to filter a final answer. Protection is applied before retrieval, during reasoning, and before any consequential action.
Weak matches are not treated as sufficient support for customer-specific claims.
Forge tracks where supporting information came from, when it was observed, and what type of source it is.
Customer evidence, retrieval history, and generated artifacts stay bound to authenticated tenant context.
Instructions found inside retrieved documents are not automatically trusted as system instructions.
Conflicting sources are flagged instead of quietly blended into a false sense of certainty.
Age and recency can be considered when deciding whether evidence is still appropriate for a current answer.
Knowledge ingestion is designed to check rights, scope, duplication, secrets, and provenance before global reuse.
High-impact actions stay separated from analysis and require explicit human approval, scope, and audit.
Knowledge boundary
Reusable knowledge is not customer evidence.
Forge can use external retrieval infrastructure for curated reusable knowledge without turning customer-specific evidence into a global knowledge pool.
Tenant-scoped by default
Stays inside the customer / Forge evidence boundary
- Infrastructure and cloud inventory
- Customer logs and configuration
- Source code and application evidence
- Customer documents and operational data
- Discovered relationships and findings
- Tenant-specific citations and audit history
Curated reusable knowledge
Eligible for the Forge global knowledge store
- Forge-owned product documentation
- Approved architecture and design guidance
- Authorized vendor/reference material
- Reviewed release and validation information
- Curated technology concepts and operating knowledge
- Sources with explicit provenance and rights basis
Default rule: customer-specific evidence is not uploaded into the global OpenAI-backed knowledge store merely because Forge encountered it.
Knowledge lifecycle
Learn deliberately. Not indiscriminately.
Forge's AI-driven knowledge base is designed as a governed ingestion and retrieval process rather than an uncontrolled dump of documents.
1Source identified
Determine source, owner, scope, and intended use.
→
2Rights + scope checked
Confirm the material can be used and whether it is global or tenant-specific.
→
3Safety screened
Scan for secrets, sensitive data, duplicates, and obvious ingestion risks.
→
4Curated + indexed
Add approved material to the searchable knowledge layer with metadata.
→
5Retrieved with provenance
Return only relevant material and keep source identity attached.
→
6Reasoned + audited
Forge distinguishes evidence from inference and records what supported the answer.
Example
General knowledge cannot impersonate customer truth.
A simple rule illustrates the boundary.
“What database does our billing application use?”
Forge response
I don't currently have verified customer evidence identifying that application's production database. I can answer once application-to-database evidence is available.
Billing Application → Billing API → PROD-SQL-04
Forge response
Observed customer evidence shows the Billing Application depends on PROD-SQL-04, identified as Microsoft SQL Server 2022. The supporting relationship can be opened and reviewed.
Useful AI should be able to explain itself — and admit when it cannot.
Forge is being built to connect reasoning to evidence, expose uncertainty, preserve customer boundaries, and keep people in control of consequential decisions.