Extension reference
vllm-sr/5.0 is systemone/0.2, the published System One standard, plus the fields below, extension set 5.0. Every vllm-sr message is a complete standard message: an extension field is an additive hint in its natural place, and a vllm-sr server answers an unknown field with 400 naming it.
The last column says what the Router does with the field when the target deployment speaks systemone/0.2: it strips a field the standard part of the message already expresses, refuses at load a rule that needs one the standard cannot carry, and never expects a response field.
Request
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
roles | object of user or context or answer | 5.0 | vela2, vela3, vela3_d3 | stripped | Which state keys hold the user's content, context, or the answer under check. Role-aware families read the state's parts by these roles; without roles, the whole state is the user's content on every family. |
options | object | 5.0 | all | stripped | deadline_ms (work not started by then is not run, and its questions answer deadline_exceeded), return_meta (the response adds meta) and max_tokens (the scan budget of a model that reads a long part in windows). A standard target is bounded by the client's own timeout instead. |
Question
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
preset | string | 5.0 | decision1, vela2, vela3, vela3_d3 | stripped | A question the model was trained on, such as pii, answered by its trained head. The question still carries its own instructions and criteria, which any other server answers. |
head (span) | router or broad | 5.0 | vela2, vela3, vela3_d3 | refused at load | The span head that answers: router, or broad on models that have a broad span head. |
decision_adapter (choice, noul, score) | boolean | 5.0 | vela3_d3 | refused at load | Answers with the package's opt-in decision adapter, on Vela 3.0 d3 sizes that declare one. |
over | string or array of string | 5.0 | vela2, vela3, vela3_d3 | refused at load | The state keys the question reads, in state order; by default the whole state (a span: its answer part if present, else its user part). |
overflow | window or truncate | 5.0 | vela2, vela3, vela3_d3 | stripped | How a model with a scan budget reads a part longer than one input: window (whole, in windows) or truncate (its first tokens). |
require_full_input | boolean | 5.0 | all | refused at load | Fails the question unless every state part is read completely; its answer then carries input_coverage: complete, which the standard has no field for. |
threshold (set, span) | number | 5.0 | vela2, vela3, vela3_d3 | refused at load | Replaces the model's decision threshold for a set or span question; the applied value comes back in thresholds. |
at (box) | number | 5.0 | decision3, vela3_d3 | refused at load | A box question on a video: the second to read, from the start of the state's only video. |
rate (track) | 1 or 2 | 5.0 | decision3, vela3_d3 | refused at load | A track question's updates per second over the whole video, 1 (the default) or 2. |
Question type
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
span | object | 5.0 | vela2, vela3, vela3_d3 | refused at load | Where each label of criteria occurs in the text it reads. The answer is a noul (the highest probability any word reached) and its spans under spans. |
set | object | 5.0 | vela2, vela3 | refused at load | Which labels of criteria apply, each with its own probability, under sets; each label also answers as a noul under <question>.<label>. |
track | object | 5.0 | decision3, vela3_d3 | refused at load | A box for each sampled moment of the state's only video, on models whose card declares track questions. |
Asset type
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
video | object | 5.0 | decision3, vela3_d3 | refused at load | A video, placed where its placeholder is, on models whose card lists the video modality. |
Answer
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
abstain_probability (choice, noul) | number | 5.0 | vela2, vela3 | never sent | The probability of the model's abstain option, where it has one (not calibrated). |
input_coverage | complete | 5.0 | all | never sent | Proof that a require_full_input question read every state part completely. |
error | invalid_question or invalid_input or max_length_exceeded or scan_budget_exceeded or invalid_model_output or deadline_exceeded or unavailable | 5.0 | all | never sent | Why this question has no answer, on POST /v1/decisions, which answers the other questions in the same 200. |
reason | string | 5.0 | all | never sent | A fixed word for why the question failed; it never carries request data. |
message | string | 5.0 | all | never sent | Why the question failed, for a person to read. |
Response
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
spans | object of array of object | 5.0 | vela2, vela3, vela3_d3 | never sent | Each span question's spans: label, start and end (Unicode code points into the text it read, end exclusive), text and probability. |
sets | object of object | 5.0 | vela2, vela3 | never sent | Each set question's selected labels and the probability of every label. |
thresholds | object of number | 5.0 | vela2, vela3, vela3_d3 | never sent | The threshold applied to each set or span question. |
span_heads | object of router or broad | 5.0 | vela2 | never sent | The span head that answered each span question, on models with a broad span head. |
meta | object | 5.0 | all | never sent | With options.return_meta, the model's verified identity, revision, precision, numerics, engine, device and timings. |
routing | object | 5.0 | the Router | never sent | Added by the Router's vllm-sr/auto routing: the recipe, decision, algorithm, stage and model that answered, and the model calls made. |
Error
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
code | string | 5.0 | all | never sent | A stable error code beside the standard detail, such as invalid_request, model_not_found, overloaded or question_failed. |
question | string | 5.0 | all | never sent | The question whose failure answers the request (question_failed). |
state | string | 5.0 | all | never sent | In a batch, the name of the request the failed question belongs to. |
reason | string | 5.0 | all | never sent | A fixed word for why the request or its question failed; it never carries request data. |
limit | integer | 5.0 | all | never sent | The limit reason names, when one was broken (bytes, pixels, seconds or tokens). |
Listing entry
| Field | Type | Since | Families | On systemone/0.2 | Description |
|---|---|---|---|---|---|
version | string | 5.0 | all | never sent | The extension set version the server answers with, such as 5.0. |
features | array of string | 5.0 | all | never sent | The question types and extension features the model answers, which vllm-sr/auto routes by. |
card | object | 5.0 | all | never sent | The model card, with its limits and placement. |
Endpoint
| Endpoint | Since | Families | Description |
|---|---|---|---|
POST /v1/systemone/batch | 5.0 | all | Named standard requests answered in one call, each as it would be alone. An item takes the batch's model and may repeat it. A systemone/0.2 target has no batch endpoint, so a batch is split into single requests. |
POST /v1/decisions | 5.0 | all | The body of POST /v1/systemone, answered with a failed question's error, reason and message in its own answer of a 200, instead of failing the request. |
GET /health | 5.0 | all | Readiness, with the extension set version the server answers with. |
Removed forms
The model runtime contract 4.0.0 accepted these forms; the standard expresses them, so 5.0 answers them with 400 naming the replacement.
| Form | Level | Replaced by | Note |
|---|---|---|---|
choices | question | criteria | An object of option keys to descriptions, for choice and noul. |
levels | question | criteria | An ordered array of level descriptions, for score. |
labels | question | criteria | An object of labels to descriptions, for set and span. |
images | request | assets | Assets of type image, placed by their placeholders. |
videos | request | assets | Assets of type video. |
states | request | POST /v1/systemone/batch | /v1/systemone takes exactly one request. |
long_context | options | TypeSafe-Danger-Zone-Long-Context: true | The standard's request header. |
error | error | detail | The standard detail, with flat code, question, state, reason and limit. |
data | listing | models[].card | Router mode keeps the OpenAI data list for Chat clients. |
implicit roles | state | roles | State keys carry no meaning; roles names the parts. |