Memory
## LLM Assigned to Workstation, Not Task (Design Clarification) **Date:** 2026-09-09 **Status:** Design clarification **Related:** Work Item ↔ Work Engineering Join Design, Work Engineering in the Workspace DO, Map Definition Tools for External LLM ### Core clarification > **An LLM is not assigned a task. It is assigned a Workstation which has input and output contracts.** This is a critical refinement of the mental model. ### The wrong model (task-based) ``` External LLM ↓ assigned Work Item ("transcribe this meeting") ↓ does work Output ("here's the transcript") ``` Problems with this model: - No explicit contract governing inputs/outputs. - No explicit quality gate. - No explicit linkage to a Routing or Work Object. - The LLM is a "black box" that just does "a task". - Hard to verify, hard to govern, hard to chain. ### The right model (workstation-based) ``` Work Engineering Map (defined by LLM, activated in DO) ↓ Routing: Meeting Audio → Notes ↓ Workstations: 1. Ingest Audio (production) - Contract: inputs=[raw audio], outputs=[normalized audio] 2. Transcribe (production) - Contract: inputs=[normalized audio], outputs=[transcript] 3. Review Notes (quality) - Contract: inputs=[transcript, draft notes], outputs=[approved notes] - Evidence requirements: cite transcript, confirm completeness External LLM (as operator) ↓ assigned to Workstation Workstation: Transcribe ↓ operates under contract Inputs: [normalized audio] ↓ executes (call_model, write_artifact) Outputs: [transcript] ↓ evidence recorded Evidence: { model: 'claude-sonnet-4', inputHash: 'sha256:...', outputHash: 'sha256:...' } ↓ lot advances Next Workstation: Review Notes ``` Key differences: - **Explicit contract** — The LLM knows exactly what inputs it receives and what outputs it must produce. - **Explicit quality gate** — The next workstation (Review Notes) has its own contract and evidence requirements. - **Explicit Routing** — The LLM is operating a defined step in a defined process, not just "doing a task". - **Evidence trail** — Every operation is recorded with hashes, model versions, timestamps. - **Composable** — The output of one workstation is the input to the next, with contracts ensuring compatibility. ### Design implications #### 1. Work item binding refined From the Work Item ↔ Work Engineering Join Design: - **Work item** = coordination task (D1, callsign-assigned) [mikesecondbrain:1] - **Bound work item** = "operate routing R for work object W" [mikesecondbrain:1] Refinement: - **Bound work item** = "operate Workstation X in Routing R for Work Object W under Contract C" The work item doesn't just point to a routing — it points to a **specific workstation in the routing** (the current station for the lot). #### 2. Callsign semantics refined From the Work Engineering in the Workspace DO design: - **Callsign** = operator identity [mikesecondbrain:2] - **Workstation** = definition (has no operator) [mikesecondbrain:2] Refinement: - **Callsign** = operator identity (human or LLM) - **Workstation assignment** = "Callsign @X is authorized to operate Workstation Y for Lot Z" - **Contract** = the governing input/output/evidence specification for that workstation The callsign is not "assigned a task" — it's **authorized to operate a workstation** under a contract. #### 3. Map definition tools refined From the Map Definition Tools design: The external LLM defines the work map, including: - Work Objects (Meeting Audio, Transcript, Approved Notes) - Contracts (Notes Approved Contract) - Workstations (Ingest, Transcribe, Review) - Routings (Meeting Audio → Notes) Refinement: When the LLM defines a Workstation, it must specify: ```typescript defineWorkstation({ name: 'Transcribe', kind: 'production', capabilities: ['call_model', 'write_artifact'], workingAreas: ['documents/', 'evidence/'], contractId: 'transcribe-contract-v1', // required inputWorkObjects: ['normalized-audio'], // explicit outputWorkObjects: ['transcript'], // explicit evidenceRequirements: [ 'Record model version used', 'Record input hash (SHA-256)', 'Record output hash (SHA-256)' ] }) ``` The contract is not optional — it's the **governing specification** for the workstation. #### 4. Execution flow refined When a lot reaches a workstation: 1. **DO checks** — Is there an operator callsign assigned to this workstation for this lot? 2. **If yes** — DO exposes the workstation's capabilities to that callsign. 3. **Callsign operates** — LLM (or human) invokes workstation methods (e.g. `call_model`, `write_artifact`). 4. **DO enforces contract** — Verifies inputs exist, outputs are valid, evidence is recorded. 5. **DO advances lot** — Lot moves to next workstation (or waits at quality gate). The LLM never "just does work" — it **operates a workstation under a contract**. #### 5. Quality gates refined Quality workstations (e.g. Review Notes) have: - **Input contract** — [transcript, draft notes] - **Output contract** — [approved notes] - **Evidence requirements** — [cite transcript, confirm completeness] - **Operator** — Can be the same callsign as production (v1, advisory only) or a different one. The key is: **quality is a workstation, not an afterthought**. It has its own contract, its own operator, its own evidence requirements. ### Naming contract (updated) | Term | Meaning | Lives in | |---|---|---| | **Work Item** | Coordination task (D1, callsign-assigned) | D1 (global) | | **Work Object** | WE artefact being produced (e.g. Transcript) | DO (Artifacts) | | **Lot** | WE run instance (routing + work object) | DO (Artifacts) | | **Callsign** | Operator identity (human or LLM) | D1 (global) | | **Workstation** | Definition (capabilities + contract) | DO (Artifacts) | | **Contract** | Input/output/evidence specification | DO (Artifacts) | | **Routing** | Ordered sequence of workstations | DO (Artifacts) | **Key clarification:** - **Work item** → assigned to **callsign** - **Callsign** → operates **workstation** - **Workstation** → governed by **contract** - **Contract** → specifies **inputs, outputs, evidence** ### Example: External LLM workflow 1. **LLM defines map** (via MCP tools): - Work Objects: Meeting Audio, Transcript, Approved Notes - Contracts: Transcribe Contract, Notes Approved Contract - Workstations: Ingest, Transcribe, Review - Routing: Meeting Audio → Notes 2. **Human (or LLM) starts lot**: - `startLot(routingId, workObjectId)` - Lot created, starts at first workstation (Ingest) 3. **Callsign assigned**: - Work item created in D1: "Transcribe meeting audio" - Work item bound to routing + workstation - Callsign `@transcribe-op` assigned to work item 4. **Callsign operates workstation**: - DO exposes Transcribe workstation capabilities to `@transcribe-op` - LLM calls `call_model(audio, prompt: 'transcribe this')` - DO records output as `transcript` work object - DO records evidence (model version, hashes) 5. **Lot advances**: - DO advances lot to next workstation (Review) - Work item updated: "Review notes at quality station" - Callsign `@reviewer` assigned (or same `@transcribe-op` if advisory) 6. **Quality gate**: - DO exposes Review workstation capabilities to `@reviewer` - LLM calls `read_artifact(transcript)`, `read_artifact(draft-notes)` - LLM calls `record_evidence(claim: 'all decisions captured', reference: transcript)` - LLM calls `submit_quality_decision(verdict: 'pass')` - DO advances lot to `accepted` 7. **Work item completed**: - Work item status → `completed` - Evidence cited in work item closure ### Benefits of this model 1. **Explicit contracts** — Every operation has clear inputs, outputs, evidence requirements. 2. **Composable** — Workstations can be chained, reused, swapped. 3. **Auditable** — Every operation is recorded with evidence. 4. **Governable** — Quality gates are explicit, not implicit. 5. **LLM-agnostic** — Any LLM can operate a workstation as long as it respects the contract. 6. **Human-compatible** — Humans can operate the same workstations as LLMs (same contract, same evidence requirements). ### Implementation implications 1. **Work item schema** — Add `workstation_id` and `contract_id` fields (in addition to `routing_id`). 2. **MCP tools** — Expose `assign_workstation(callsign, workstation_id, lot_id)` tool. 3. **DO runtime** — When a lot reaches a workstation, check if a callsign is assigned; if not, block and surface in `get_attention`. 4. **Contract validation** — Before executing a workstation method, verify inputs exist and match contract. 5. **Evidence recording** — Make evidence recording mandatory for workstation execution (not optional). ### Relationship to prior designs - **Work Item ↔ Work Engineering Join** — This clarifies that "operate routing R" means "operate workstation X in routing R under contract C". [mikesecondbrain:1] - **Work Engineering in the Workspace DO** — This reinforces that workstations are capability projections governed by contracts. [mikesecondbrain:2] - **Map Definition Tools** — This makes contracts mandatory in workstation definitions, not optional. --- **Note:** This is a design clarification, not a new direction. It sharpens the mental model: LLMs (and humans) operate workstations under contracts, not "do tasks".
Tags: work-engineering, workstations, contracts, llm, operators, callsigns, design-clarification, volatility:durable, kind:semantic, status:canonical — Source: mcp — 2026-09-08 23:03:10 UTC