THE DISTINCTION / STATE, EXECUTION, AND EVIDENCE
Giving an agent custom functions does not make it a different kind of system. The meaningful questions are where exact state lives, how results reach the answer, and who checks the computation.
01 / TWO DIFFERENT FAILURE MODES
A tool computes an invoice total of $1,283.47. If the model rewrites it as $1,238.47, the calculation was correct but the answer was corrupted. A registry reference rendered directly closes that particular path.
A model selects last month’s invoice instead of this month’s. The tool can execute perfectly and still answer the wrong question. Exact rendering cannot fix that. Binding checks, behavioral evidence, and abstention address a separate problem.
02 / COMPARE THE CONTRACTS
| Concern | A basic tool loop | First ExactAI milestone | Research target |
|---|---|---|---|
| Custom functions | Model selects a callable. | Qwen selects typed application functions. | ExactNet composes dynamically supplied typed operators. |
| Intermediate state | Tool results return to model context. | Typed registry objects persist across turns; context uses descriptors and references. | Intermediate program values remain in runtime state. |
| Multi-step work | Model chooses successive tool calls. | One session coordinates Qwen, functions, and built-in ExactNet operations. | One semantic delegation can produce an entire exact program. |
| Final answer | Model may write values into prose. | ExactIR references and formatting materialize exact spans. | The same rendering boundary is retained. |
| Checks | Depend on the application and tools. | References, signatures, result types, execution errors, and provenance. | Add semantic types, behavioral checks, and calibrated execute/defer decisions. |
| New functions | Expose descriptions to the model. | No ExactNet retraining; Qwen owns selection. | Test unseen-function composition by ExactNet without new output classes. |
| Private values | Exposure depends on tool arguments, returned data, and deployment. | Raw PII can stay in a local registry; use opaque IDs and minimal descriptors. | Preserve that boundary through program synthesis and execution. |
These are architectural choices, not capabilities forbidden to tool-calling agents. A well-designed tool system can preserve typed state, validate arguments, and render results directly. Our hypothesis is that packaging these boundaries with a specialist for exact programs can make them easier to adopt and more effective. That needs comparative evidence.
Privacy is also a deployment choice, not an exclusive advantage over tools. Local models and local functions can keep private information on your system. References help minimize disclosure, but identifying prompts, descriptors, logs, or outputs can still leak it. See the local-data design →
03 / FROM CALL SELECTION TO PROGRAM SELECTION
LLM → invoice_total(subtotal, tax)
← result value
LLM → subtract(current, previous)
← result value
LLM → compose answerEach return is another opportunity for the application to preserve or lose the exact-value contract.
LLM → “Calculate and compare invoices”
ExactNet → candidate typed program
Verifier → accept / clarify / defer
Runtime → execute accepted program
LLM ← result references
Renderer → materialize exact spansMoving an exact subproblem into one program could reduce repeated large-model orchestration. Latency and cost improvements remain hypotheses to measure.
04 / WHAT WE ARE NOT CLAIMING
Calculators, typed representations, references, program synthesis, and external execution are building blocks. Custom functions alone are not a novelty claim.
A type-correct program can be semantically wrong. Behavioral contracts cover specified properties and cases; they do not automatically recover an ambiguous user’s intention.
A trace explains the source and computation of an exact span. It does not certify every sentence around it. Literal detection is an incomplete guardrail.
Verified synthesis, RLBC, calibrated abstention, and unseen-function composition belong to the research roadmap. The first milestone is the integrated conversational wrapper.