AI should answer from facts you can point to.
MLNavigator researches and builds local AI systems that trace answers back to source records, preserve receipts, reduce repeated review work, and remain useful without outside services.
Inside your environmentNo document egress
- MSA v4 §8.2
- Addendum p.3
- Policy §11
run 7f2a · policy legal-review@3 · model pinned llama-3.1-8b · offline
Read the research →
Local AI that answers from your records, shows the source, and runs without outside services.
Four things we require before a local answer is useful for review
Not a brand statement — working constraints. Each one has to hold for an operator to defend the run later.
adapterOS came out of four constraints we kept hitting
An answer had to stay attached to its source. The source had to remain inspectable. The work had to survive without an outside service. And restarting the same review from an empty transcript made less sense every time. adapterOS is machinery for satisfying those at once — local runs against approved records, with a trace a reviewer can follow.
The answer is only one artifact
A Controlled Evaluation asks whether a run can preserve useful context for later human review: the sources involved, what ran, and what a reviewer decided.
Source trace
The exact records and passages an answer drew from.
Run receipt
What ran — model, policy, inputs, outputs — as a signed line a reviewer can check.
Review packet
What a human accepted, rejected, or marked uncertain.
run local-0142
sources approved set (3)
model llama-3.1-8b (pinned)
policy legal-review@3
network offline
review held for legal approvalUseful without outside services
Sensitive document work runs locally. No outbound network calls, no telemetry, no routine document egress. Models are verified by hash before use, and updates are explicit and verified.
Customer environment
Approved records
Local model
Questions
Answers
Receipts
Review state
No routine document egress
Less repeated effort
When a team has already reviewed the source trail, the next run should reuse that work and focus attention on the delta — fewer repeated full-context runs on the same material.
First runfull source pass
Later runreviewed context + delta
Local operationno re-shipping the same documents
Measured as compute per useful answer — an engineering direction, not a marketing number.
Define one workflow before evaluating
Start by specifying one source-bound document workflow your reviewers already own: its source boundary, reviewer route, prerequisites, and acceptance questions. A controlled evaluation is a separate decision under a separate agreement.
01
Name the candidate workflow
Choose the review, reporting, or compliance task whose records and decision path need definition.
02
Map the source boundary
Identify the approved sources, reviewer route, operating constraints, and unresolved prerequisites.
03
Write the acceptance plan
Define what a later evaluation would need to demonstrate without treating the plan as an implemented system.
04
Decide whether to evaluate
Use the definition packet to decide whether a separately agreed controlled evaluation is appropriate.
You keep: the definition packet
- Great Plains Regional I-Corps
- ACCEL-KS Grant
- Company history and research inputs — not customer validation
Start with one answer your team can trace.
Bring one sensitive workflow, the records it must stay inside, and the review standard it has to meet. We define the source boundary, reviewer route, prerequisites, and acceptance questions before deciding what follows.