Errors and evidence limits
Resolve missing inputs and distinguish execution from task outcomes.
Read the common response's status, error.code, field_path, failed_operation, retryable and next_action. Successful responses are schema-validated too. A stopped client or a model content-filter rejection is not a semantic refusal from WorldFuzz.
| Condition | Meaning and next action |
|---|---|
ACCESS_DENIED | Stop. Check current workspace membership, token scope, owner grant and separate source rights with the operator. Do not enlarge scope or switch to another user's catalog. |
PROFILE_INCOMPATIBLE | The operation, data type or semantics do not match. MCAP is not an XYZ trajectory pair. Inspect the named field and select compatible registered input; do not report a technical exception as a proved frame conflict. |
SCHEMA_INVALID | Correct the reported parameter using describe_operation. Unknown parameters and duplicate JSON keys are rejected. |
PLAN_CHANGED or an expired plan | Inputs, versions, code or permissions no longer match the prepared snapshot. Inspect the existing run before preparing a new plan. |
RESOURCE_BUSY | The one-worker workspace has no free slot. A new logical job/reservation was not created. Read or finish the existing run before another submission. |
| Budget refusal | Inspect get_budget and workspace usage. Measured usage, estimates, conservative holds and unknown tails remain separate. A retry may use only confirmed remaining budget. |
GROUP_INCOMPLETE / HELD_OUT_PROTECTED | Select complete linked groups and an allowed purpose. Never normalize on test data or silently substitute a training source. |
| Unsupported replay or runner capability | Stop or explicitly revise the request with the user. Do not replace action execution with recorded-state viewing or claim physical reproduction. |
Incomplete and ambiguous requests
compile_goal and run_goal return concrete missing_parameters for an ambiguous structured goal. Supply those fields. Do not repeatedly call get_capabilities when the service has already identified the missing input.
An expired wait retains a run ID and means the observation window ended. Repeat the same key and goal or call get_run; do not submit a second job. After a lost worker, use reconcile_run before an explicitly authorized resume_run. Cancellation preserves events and partial artifacts.
Three separate outcomes
- Execution state: queued, running, completed, failed, cancelled, budget_exhausted or incomplete.
- Task outcome: what the robot/software task actually did, or unknown.
- Verification: requested mode, checks performed and proven_result. Integrity alone cannot establish task success.
Nanosecond values in public responses use decimal strings under nanoseconds-decimal.v1. Preserve the strings or parse them with arbitrary-precision integers. Stored hashes refer to the original bytes, not a reserialized transport view.
Campaigns and provider limits
A requested campaign requires the campaign operations and aggregate budget, including search, reduction, checks, retries and failed attempts. Separate successful jobs do not complete it. Structured goal compilation needs no LLM; provider contract-only mode makes no model call. No paid inference, training, cloud job or device execution is implied by these docs.
The original WP31 twofold action-reduction target was not achieved. Both independent WP32 studies retain their original failures, including false confirmations. Limited M7 service acceptance is not full agent certification. Saved Can states, offline metrics and headless simulator execution do not establish physical reproduction.