GPT-5.6 · proposal authority
Compile professional language
Interpret new document structures, identify candidate obligations, cite the relevant language, and emit a restricted typed program without executing money or external action.
The critical boundary
Let the model understand the language. Never let it own the money.
Fixed rules cannot anticipate every agreement, amendment, or closeout clause. GPT-5.6 handles the unfamiliar semantic layer. Its authority ends at a restricted proposal that deterministic software must prove.
GrantBack · execution authority
Prove or abstain
Resolve the controlling source, verify every cited span, execute exact Decimal arithmetic, create the candidate package, and fail closed when any binding changes.
High-assurance architecture
A financial agent whose refusal is part of the product.
The result stays continuously bound to evidence, calculation, package identity, human approval, and the exact action that may follow.
01 / COMPILE
Typed obligation IR
The model proposes typed arithmetic but does not execute it. The proposal contains no arbitrary code or financial authority.
02 / AUTHORITY
Controlling-source graph
Agreement and amendment precedence is resolved before any financial claim can execute.
03 / PROVENANCE
Exact source spans
Document identity, typed locator, exact quote, quote hash, and parser revision travel with each operand.
04 / MONEY
Decimal execution trace
Server-owned arithmetic reconstructs every input, intermediate step, rounding decision, and result.
05 / CONTROL
Approval-bound actions
Source root, program, amount, recipient, body, attachments, actor, nonce, and expiry form one reservation.
06 / FAILURE
One-character fail closed
A new source root blocks the changed amount, candidate package, approval reservation, and outbound action.
One MCP agent · two native surfaces
Use the same proof where professionals reason and where builders change the system.
ChatGPT turns the workflow into a natural-language accounting review. Codex turns the same MCP result into a repository-level audit and implementation path.
Professional operating surface
ChatGPT desktop app
Attach GrantBack in a conversation and run the review inside an OpenAI Apps SDK widget.
- GPT-5.6 interprets unfamiliar grant and closeout language in conversational context.
- The widget renders structured MCP output, exact evidence, accountant review state, and the negative proof.
- The professional can challenge the result, inspect the source boundary, and test a one-character change without leaving the conversation.
Builder and auditor surface
Codex desktop app
A repository-aware agent can move from the visible result to the exact implementation, tests, and release.
- Open the GrantBack workspace and invoke the same MCP hero pipeline from the project thread.
- Trace a source span through typed contracts, authority resolution, Decimal evaluation, package binding, and deployment evidence.
- Ask Codex to add a new obligation test-first, inspect the diff, run the release gates, and verify the production commit.
Operational honesty
A state name is not proof of the next state.
PROPOSALModel interpretation is not a verified obligation.
CANDIDATEA verified package is not a posted journal entry or confirmed debt.
ACCEPTEDProvider acceptance is not delivery.
DELIVEREDMail-server delivery is not recipient observation.
OBSERVEDRecipient observation is not payment.
Run the evidence
See what the model may propose, what the engine can prove, and exactly when GrantBack refuses to act.
The public case uses deidentified synthetic evidence and requires no login.