ValueTypeAI

FIRST MILESTONE / ILLUSTRATED CONVERSATION

One conversation.
Values that stay.

Follow a custom invoice function and an ExactNet follow-up through the same session. This is an illustrative target flow, not a live model execution or benchmark.

INTERACTIVE FLOW

Invoice calculation → follow-up comparison
01 ASK + REGISTER
“Explain this invoice and calculate its total.”
Subtotal / local registry$1,200.00
Tax / local registry$83.47
Previous invoice / local registry$1,100.00
→
02 CUSTOM FUNCTION
$MoneySUBTOTALinput
$MoneyTAXinput
$MoneyCURRENTresult

Qwen selects invoice_total(SUBTOTAL, TAX). The runtime stores CURRENT. The renderer shows $1,283.47; session history retains its reference.

→
03 EXACTNET FOLLOW-UP
“How much higher is that than the previous invoice?”
delegate comparison
candidates CURRENT, PREVIOUS
compute SUBTRACT
          [CURRENT, PREVIOUS]
store DIFFERENCE
✓ types✓ arguments✓ format
→
04 RENDER + RETAIN
This invoice totals $1,283.47, which is $183.47 higher than the previous invoice.
✓Rendered from referencesValues persist in the session

05 / CONCRETE COMPARISON

Same request.
Different value paths.

This separate revenue example illustrates rendering choices. Tool calling can also preserve typed values; the distinction depends on the surrounding architecture.

INPUT

“Compare Q4 revenue of $187,312,894.27 with Q3 revenue of $164,821,304.11.”

01
COMMON APPROACH

LLM only

Prompt → generated response

POSSIBLE OUTPUT

“Q4 revenue was $187,312,849.27, up 13.6%.”

! Transposed digits can look plausible
  • Value handlingToken regeneration
  • ArithmeticProbabilistic
  • Failure modeMay be silent
  • Output traceNot inherent
02
EXISTING PATTERN

RAG + tools

Retrieve → tool call → generated response

TOOL RESULT

percent_change(...) → 13.6%

~ Exact calculation, free-form assembly
  • Value handlingIntegration-specific
  • ArithmeticExact when delegated
  • Failure modeVaries by tool chain
  • Output traceApplication-defined
03
TYPED EXACTNESS LAYER

ValueTypeAI

END TO END

Register → ExactIR → validate → render

ILLUSTRATIVE RENDERED OUTPUT

“Q4 revenue was $187,312,894.27, up 13.6%.”

✓ Source value referenced, not rewritten
  • Value handlingTyped references
  • ArithmeticDeterministic runtime
  • Remaining riskValid but wrong selection
  • Output traceBuilt into the response

Deterministic execution does not establish semantic correctness. Read the full tool-calling comparison →

THE MODEL DECIDES

What should
be said.

  • Understand the request
  • Compose natural language
  • Select exact objects and operations

THE RUNTIME GUARANTEES

What must
stay exact.

  • Preserve source values
  • Validate and execute operations
  • Format results with provenance

YOUR WORKFLOW, MADE EXACT

Bring us the values
that can’t drift.

We’re working with teams building AI for data-heavy, high-consequence workflows.

Start a conversation ↗