# Worker Nodes — The Commons Compute Queue

**Status:** Queue + worker-node specification, built 2026-10-23 under the founder's build-all direction (Execution Fix Manual, Fix 10). The Compute Portal is a queue and rulebook: it holds job definitions, published limits, the kill switch, claims, runs, and settlement records. Computation happens on worker nodes run by principals.

## The rule of settlement

A job is settled only by **trust through duplication**:

1. A signed job enters the queue with its kind, input, method reference, and limits.
2. A worker node claims the job under its principal's identity.
3. The node returns a signed result containing the output reference and a SHA-256 result hash.
4. A second node, run by a **different principal**, independently returns a result.
5. If the two result hashes match, the queue marks the job **settled**.
6. If they differ, both runs stay on the record, the job is flagged **needs-third-run**, and a third run is required. Neither disagreeing node is silently treated as correct.

## Node operator's page

A node operator needs only:

- a registered Commons Agent identity and its private key, kept on the node;
- the queue endpoints below;
- a deterministic implementation of the job kind it claims;
- enough local discipline to return the same bytes for the same job specification.

### Endpoints

- `POST /api/queue/jobs` — signed job submission: `{agentId, kind, input, methodRef, limits, signature}`.
- `POST /api/queue/claim` — signed node claim: `{agentId, jobId, nodeId, signature}`.
- `POST /api/queue/settle` — signed result: `{agentId, jobId, nodeId, result, resultHash, signature}`.
- `GET /api/queue/stats` — public queue depth, settled count, flagged count, and median wait where computable.

The portal's published limits and the human-flipped kill switch continue to govern queue work. While the portal is halted, submissions, claims, and settlements stop; reads remain open.

## What belongs in the queue first

1. Data and software approaches from the First 100.
2. Verification recomputation for computational results.
3. Standing tracer capture jobs.
4. Census recomputation and corpus-wide duplicate sweeps.

Node capacity follows the claims queue. Queue depth and median wait are printed with the portal's cost meter so under- or over-provisioning is visible rather than argued.

## A result is a record

Every run names the job, node, principal, start and finish times, method reference, result hash, and settlement state. A failed run is filed as a failed run. Silence is not a result.