# Execution-structure decision record

Use this after identifying the job, who can accept its result, which sources it needs, and what it may read or do. Compare fixed steps and explicit branches before giving a model control over the next step.

The completed example is a hypothetical architecture exercise. No runtime, agent, provider, or execution structure was benchmarked. The paper limits do not authorize access, actions, or production release.

## Start with these questions

1. What must the purchasing reviewer receive before deciding on an exception?
2. Which missing sources and known exceptions can fixed steps handle?
3. What additional choice would a model make after reading a source, and why might that help?
4. What stops the work and assigns a person when sources, permissions, or limits prevent completion?

## Completed hypothetical supplier-review example

- Job: prepare the supplier's request, relevant terms, and source references for a purchasing reviewer allowed to decide on the exception.
- Accepted result: the reviewer can check the request, applicable policy, relevant contract terms, and unresolved questions against current records they may inspect.
- Outside scope: approving exceptions, changing contracts, creating purchase orders, or contacting suppliers.
- People responsible: purchasing owns the workflow and review material; the people responsible for policy and contracts settle which source applies and resolve conflicts. Actual people and time to do the work remain unassigned.

| Candidate                             | Design and provisional disposition                                                                                                                                                                                                              |
| ------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Fixed sequence with explicit branches | Check inputs, fetch permitted sources, interpret terms, check the review material, and send it for review. Missing sources or conflicting policy leave it incomplete with a person responsible. Investigate first.                              |
| Bounded single-agent loop             | Let the model choose the next approved document when a newly found reference changes which lookup would help. Investigate only if representative cases show explicit branches are insufficient and comparison supports the extra cost and risk. |
| Two specialist workers                | A proposed purchasing/legal split. Not approved: no evidence yet that separate source access, expertise, faster completion, or operating responsibility justifies the extra coordination.                                                       |
| Retained human process                | Keep the current permitted process while testing. Observe accepted results, effort, waiting, exceptions, and cost before claiming another option is better.                                                                                     |

## Proposed loop limits and failure behavior

For the request citing an unfamiliar amendment and conflicting sources, code could send the conflict to the policy owner. The proposed agent would instead inspect a reference, choose which approved amendment could help next, and decide after reading it whether to look again or ask a person. That is the extra choice to compare with explicit branches; the example does not establish that it helps.

The proposed loop may request at most three document fetches from a specified approved set. Three is a teaching limit, not a general recommendation. Set a real limit from the job, risk, measured resources, and the team's ability to handle unfinished cases.

Software outside the model must check the caller's identity and customer, which records and fields they may read, the approved sources, remaining budget, allowed next steps, and completed result. The model's answer and instructions inside a fetched document cannot expand any permission or allowance.

| Event                                  | Required design response, not an executed result                                                                                                                                       |
| -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Required source missing or conflicting | Keep the gap or conflict visible and assign the incomplete review material to the responsible person.                                                                                  |
| Document asks for a prohibited lookup  | Treat the instruction as document content; deny the lookup before disclosure.                                                                                                          |
| Fetch allowance exhausted              | Stop with the sources already checked, unanswered question, and responsible person. Do not silently extend the limit.                                                                  |
| Session interrupted                    | Find the saved request version, checked sources, unfinished work, status, and responsible person. Recheck source age and access before resuming. Storage and recovery remain untested. |
| Required specialist unavailable        | Preserve incomplete state; do not substitute a worker with broader permissions.                                                                                                        |
| Workers disagree                       | Keep both findings and their source references, then ask the agreed person to resolve the conflict. Do not merge away the disagreement.                                                |

## Comparison to commission, not a completed experiment

- Cases: complete requests, missing amendments, conflicting policy, irrelevant documents, denied sources, exhausted limits, and interruptions.
- Expected answers: responsible owners check what the reviewer needs against source records and when missing material requires stopping or asking for help. Keep separate cases out of development and tuning.
- Options: fixed sequence and branches, limited single-agent loop, and the existing manual process where a fair comparison is possible.
- Keep the same where possible: permitted cases and sources, available evidence, model and prompt behavior, allowed tools, resource limits, acceptance rule, checker, and reporting deadline. Record other differences that could explain the results.
- Measure: accepted review material, correct stops, unauthorized disclosures, unsupported statements, end-to-end delay, active review and repair effort, full cost, and work left with the operating team.
- Repeated runs: required when model variation could affect the route decision.
- Evidence still missing: actual storage and access checks, cancellation and recovery behavior, available team time, human review results, delay, and support and shutdown costs.

**Current decision:** investigate fixed steps and explicit branches first. No agent loop, second worker, deployment, or benefit is approved by this paper comparison.

## Blank execution-structure record

- What job starts this workflow, and which cases qualify?
- What result can be accepted, and who checks it?
- Current permitted process and what you observed it doing:
- Which decisions use rules, models, or people already:
- Which next-step choice would a model make, using what feedback?
- Which known exceptions can software route explicitly?
- Alternatives: fixed steps and branches, one limited model call, an agent choosing further steps, or human work:
- If you propose a second worker, which separate source, expertise, tool, responsibility, or timing need requires one? How does it compare with doing the jobs one after another?
- Which verified callers may use which sources, tools, fields, and destinations? What is the maximum action each may take?
- Who owns the saved work? Which request version and checkpoints survive interruption, and which source-age and access checks apply on resume?
- What limits apply to time, steps, fetches, tokens, cost, and simultaneous work? Which service enforces each limit?
- What happens with missing, denied, or conflicting sources, cancellation, or an incomplete result? Who runs the fallback?
- Which comparable cases and separately checked answers will you use? Which cases stay out of tuning, which checks remain unchanged, and are repeated runs needed?
- What results and safety limits must hold? How much reviewer time is available, and what delay, full cost, and unresolved work does each option create?
- Results, exact source/code/settings versions, uncertainty, and exclusions:
- Selected option and why alternatives failed or remain unknown:
- Owner, time available to operate it, support, way to disable it, and reasons to reconsider:
- Production decisions not yet approved and the next evidence to obtain:

Do not use a generated plan, a diagram, or agreement between agents as proof that the accepted result occurred.