SOURCE VALUES01
REVENUE_Q4$187,312,894.27
REPORT_DATE2026-09-19
ACCOUNT_IDacct_91DF83A912
Research prototype / ExactAI architecture
One agent.
Exact state.
An agent should reason with your values, not rewrite them. We’re bringing conversation, your application functions, and ExactNet into one system—with typed state that persists from the first calculation to the next question.
Typed valuesDeterministic executionTraceable output
ExactIR
TYPE
CHECK
EXECUTE DETERMINISTIC RUNTIME
EXACT RENDERING / ILLUSTRATION✓
Q4 revenue was $187,312,894.27 as of September 19, 2026.
REFFORMATTRACE
Money
Date
Identifier
01 / THE PROBLEM
A correct calculation can still
become a wrong answer.
A tool can return the right amount, then a model can rewrite it incorrectly—or calculate perfectly using the wrong invoice. Preserving values and choosing the right computation are separate problems. Our architecture addresses both, with different mechanisms and different tests.
LANGUAGE
“grew rapidly” ≈ “expanded quickly”
SEMANTICALLY EQUIVALENT
EXACT VALUE
13.42% ≠ 13.24%
MATERIALLY DIFFERENT02 / THE PLATFORM
One conversation.
Shared exact state.
Our first integration milestone connects local Qwen, application-defined functions, ExactNet, and the runtime behind one session. The application should not have to coordinate two models.
01
Values that survive the next turn
The session design keeps source values and computed results as immutable typed objects. Follow-up questions use their references, not a retelling of the last answer.
02
Your functions, typed
Register a description, argument types, return type, and trusted callable. In the first milestone, Qwen selects these functions without changing ExactNet’s checkpoint.
03
ExactNet inside the agent
Delegate a scoped exact question to the specialist. It selects supported operations and binds arguments; the runtime executes and renders the result.
04
A trace, not a claim of truth
Follow each result back to its inputs and computation. A valid execution trace explains what happened; it does not prove the model understood the request.
03 / HOW IT WORKS
Intent. Program. Execution.
A direction we can test.
Long-term research: evolve ExactNet from a fixed operation selector into a neural compiler for typed programs. Verification and calibrated abstention are research milestones, not shipped guarantees.
01⌁
UNDERSTAND
Express the exact subproblem
The conversational model supplies intent and relevant object references. Typed values remain in the runtime.
02◇
SYNTHESIZE + VERIFY
Check candidate programs
Future ExactNet proposes compositions. Type, binding, and behavioral checks help decide whether to execute, defer, or clarify.
03✓
EXECUTE
Execute and preserve
The runtime executes the accepted program, stores its results, and renders exact spans directly into the response.
04 / WORKFLOWS WE ARE EXPLORING
When “almost right”
is simply wrong.
Financial workflows
Reports, reconciliations, summaries, and calculations grounded in source values.
EXPLORE ↗Enterprise systems
Keep customer, product, account, and transaction identifiers byte-perfect.
EXPLORE ↗Time-sensitive operations
Handle dates, durations, deadlines, and derived values deterministically.
EXPLORE ↗Auditable AI
Show where exact outputs came from and which operation produced them.
EXPLORE ↗05 / PRIVATE VALUES CAN STAY LOCAL
Reason with a reference.
Keep the PII local.
A customer’s private account identifier can stay in an application-controlled registry. The model can work with an opaque reference and a minimal description; trusted local code resolves the real value only when needed.
Explore the privacy boundary →Your data, your execution boundary
The first integration targets local Qwen and ExactNet. With local functions and rendering, private values need not leave the local application to produce an answer.
Designed for data minimization
References are not automatic anonymization. Prompts and descriptors can still disclose identity, and a function or final response can expose a value. Explicitly register sensitive fields and control what crosses each boundary.
V
OUR THESIS
“Executing the wrong program perfectlyRead our core beliefs →
is still a wrong answer.”
BUILD THE TRUST LAYER
Make your AI
exactly useful.
We’re building with teams whose AI touches values that cannot drift. If that sounds like your workflow, we should talk.
Start a conversation ↗