Original prompt: full use-case database, full production JSON, full extracted source-document text, full previous-proposal payload. Highest cost, used as comparison baseline.

Slim use-case fields, stripped production JSON (drops the role catalogue and audit fields), summary-only for long source documents, and compact previous-proposal bullets. ~60% fewer input tokens vs. Variant A.

Production rendered as markdown — overview table, description prose, then workflow grouped by stage and subsection with each step's fields in a table. Aims to reduce process-step-ID hallucination by removing the JSON nesting and giving each ID its own code-styled table cell.

No use-case catalogue in the prompt. The model discovers use cases on demand by calling a `search_use_cases` tool (exact filters, text search, vector search). Defaults to excluding use cases already linked into this production's workflow.

Like Variant C but the catalogue's `Already Mapped` column is removed and replaced with a top-of-prompt 'Forbidden Pairs' block that explicitly lists every (useCaseId, processStepId) currently on the production. Reduces the model's tendency to misread the column as a recommendation about which step a use case should target.

Like Variant E, but raw source-document text is replaced by a cached production-level summary tailored for AI use case identification. The summary is regenerated only when the active source-document set changes. Falls back to Variant E's per-doc summaries if summary generation fails.

Generator behaviour matches Variant F (cached document summary, forbidden pairs, full catalogue). The reviewer pass is refocused: dead validation summaries and the verbose proposal-history dump are removed; the use case catalogue is added so the reviewer can flag pairs where a clearly better use case exists for the proposed step. Scoring rubric narrows to step fit, application specificity, and better-fit-available.

Original prompt: full role catalog JSON in the system prompt, full production JSON, raw extracted source-document text, full known-members section, full JSON dump of the 3 most recent completed proposals. Highest cost, used as comparison baseline.

Compact markdown role catalogue plus markdown-rendered production. Raw source-document text is replaced by a cached production-level summary tailored for role assignment (named people, departments, vendors, scale signals). The summary is regenerated only when the active source-document set changes. Falls back to raw source-document text if summary generation fails.

Like Variant B, but the prompt no longer instructs the model to produce a `previousItemsRationale` array. Pending and rejected roles are listed informationally; the AI decides which (if any) to repropose. Novelty is enforced by code: any role already accepted on the production is dropped before review.

Original prompt: full production document, raw extracted source-document text, JSON dump of the 3 most recent completed proposals, full action-rule restatement in the user prompt. Highest cost, used as comparison baseline.

Production rendered as markdown (overview + per-stage workflow tables); previous proposals as compact bullets, accepted steps filtered out; raw source-document text replaced by a cached production-level summary tailored for workflow planning. Tools narrowed to ontology search (get_organization dropped — org info inlined). Current Team section removed (Role agent's domain). AI review runs against the same production markdown, org info, source-document summary, and previous proposals the generator saw. Falls back to raw source-document text if summary generation fails.

First pass uses Variant B's prompt and tool set to draft a proposal. A second generator pass receives the draft and is asked to self-critique and return a refined proposal — same JSON schema, no extra context. AI review is skipped (the refinement pass is the review). Deterministic code checks still gate the final proposal.

VariantB's prompt and tools, plus a Global Workflow Steps Library section listing entries from past productions of the same type, grouped by stage and subsection and sorted by usage count. The AI reviewer pass is skipped — only deterministic code checks gate the proposal. The library section is omitted when no entries match the production's type.

VariantE's prompt skeleton (production markdown, source-doc summary, Global Workflow Steps Library, ontology-only tools, AI review skipped, no `description` field on output steps, modify-assets-only) plus a cached summary of the production's communications/notes between source docs and library. The notes summary weights more-recent notes higher when they conflict with older ones. Cache regenerates on note adds, deletes, or edits.