# Fix list: Vela 2.0 From Anchor Terminal's listing at https://www.anchorterminal.com/tools/vela, the October 2026 research run, assessed 8 October 2026. Grade B, 66.5 out of 100. This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public. For a coding agent working on Vela 2.0: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published. ## 1. Reliability, 57 out of 100, up to 8.6 more on the total Why it scored 57: Scored on the local-package checklist, since Vela 2.0 is a model the owner runs. There are two routes. The model cards load it with `pip install torch "transformers>=5.17"` and `AutoModel.from_pretrained(..., trust_remote_code=True)`, with no Python version stated. The router's `vllm-sr` package on PyPI states Python 3.10 or newer, but serving a model with it is newer than the 0.4.0 stable release and needs the development channel (12 of 20). The semantic-router repository has public CI whose model-runtime suite includes five `test_vela2_*` files, and each model repository ships a parity record against the research scorer with zero decision differences on 230 rows. We couldn't read today's CI result on main, since the GitHub API was rate-limited and the Actions page we got was stale (15 of 25). The router repository showed 408 open issues on 8 October, 23 of them mentioning Vela, against 200 commits between 23 September and 8 October. The four Hub repositories have no discussions. Issue #4668 of 7 October tracks the 0.3B's CPU latency, which the release note puts at 79 ms against 16 ms for Vela 1.0 (14 of 25). The Hub repositories have no version tags, and the 4B and 9B weights were replaced in place on 3 October. The router pins the revisions it loads and publishes a dated release note (6 of 15). Released as 2.0 with no beta label on the cards, though the `vllm-sr` package is classified beta and the family is two days past launch (10 of 15). The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability): Hosted APIs, MCP servers, models and platforms. - 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own). - 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so. - 15, rate limits documented with numbers. - 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved. - 10, an SLA published for any paid tier. - 10, the surface agents use is generally available, not beta or preview. Local packages, SDKs, frameworks and stdio MCP servers. - 20, installs from an official package with supported runtimes stated. - 25, a public CI and test suite, passing on the default branch. - 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered). - 15, semver discipline and breaking changes called out in a changelog. - 15, version 1.0 or later, or declared stable. Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors. ## 2. Security & auth, 60 out of 100, up to 7 more on the total Why it scored 60: Read as software the owner runs. No account. The bundled `vela2_serve.py` binds to 127.0.0.1 and is open unless `VELA2_API_KEY` is set, after which it requires one bearer key, compared in constant time. The router's model runtime has no authentication, and its docs say to expose it only on a trusted network (13 of 30). A decision model has no write actions (15 of 20). The models are trained to flag prompt attacks, each question sees only the state and its own block, and the router's SECURITY.md names adversarial input to classification models in its threat model. Nothing documents how hostile text in the state can move an answer (10 of 15). The runtime exposes Prometheus metrics and, on request, the revision, digest and timings that answered a call, and its docs say request text is never logged. The bundled server keeps no record of calls (9 of 15). SECURITY.md takes private reports through GitHub Security Advisories, with severity classes and a disclosure policy. Weights are safetensors with a SHA256SUMS manifest in the decoder repositories, and the runtime checks every file's SHA-256 and never runs code from a model repository. The `transformers` route does run repository code through `trust_remote_code=True`, with no tag to pin. No bug bounty, and no security.txt on vllm-sr.ai (13 of 20). The checklist (https://www.anchorterminal.com/benchmark/#checklist-security): - 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option. - 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions. - 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10. - 0 to 15, audit logs or per-call visibility for the operator. - 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public. Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing. ## 3. Payments & pricing, 60 out of 100, up to 5 more on the total Why it scored 60: Free Apache-2.0 software with nothing to buy, so 20 + 20 + 20 for pricing, free use and no sign-up. No payment protocol (0). There is no hosted API. Hardware is the owner's cost. The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments): The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/). - 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which. - 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login. - 20, a free tier or trial that doesn't need a card. - 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API). Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied. Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol. ## 4. Transparency & trust, 48 out of 100, up to 4.6 more on the total Made of editorial 68, provenance 27. Why it scored 48: Apache-2.0 for the weights, code and documentation, on Apache-2.0 Decision 2.0 and Qwen3.5 bases, with MODIFICATIONS.md and ATTRIBUTIONS.md in each repository. The 0.3B's tokeniser keeps the Gemma Terms of Use, and training data isn't redistributed and keeps its own licences, some CC BY-SA. We found no training code (26 of 30). Self-hosted, so inputs stay on the owner's hardware. The runtime docs say request text is never logged and metrics carry no request content. There's no privacy statement for the project (22 of 30). The router's support matrix defines a Deprecated class and says nothing is in it, and the release note gives the lines that restore each Vela 1.0 model. No policy with notice periods (10 of 20). We found no telemetry statement either way and didn't audit the runtime's code for it. Weights download from the Hugging Face Hub (10 of 20). The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency): - 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms. - 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors). - 0 to 20, a deprecation policy or notices with dates. - 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted). The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two. Provenance checks not met in full (half of this category, computed from checked facts): - Legal entity named: not found (0 of 20) - Domain age: vllm-sr.ai, registered 2026-07-13 (under a year) (0 of 15) - Status page: not found (0 of 10) - security.txt: not found (0 of 10) ## 5. Schema & documentation, 78 out of 100, up to 3.6 more on the total Why it scored 78: Read for a model the owner serves. The router's model runtime carries an OpenAPI 3.0.3 file, contract version 2.2.0, served at `GET /openapi.yaml`, which covers `/v1/decisions` and its `/v1/systemone` alias. The server bundled in the model repositories has no specification file (22 of 25). vllm-sr.ai has no llms.txt (404). The model cards, USAGE.md, EVALUATION.md and TRAINING.md are Markdown in each repository (5 of 10). Each card states the size's intended place and its evaluation limits, and the router's release note names the signals where the 0.3B is behind Vela 1.0. We found no section on uses to avoid (15 of 20). Five question types with 2 to 255 options, 2 to 10 score levels and 1 to 255 labels, and enums for span head and long-input policy. State is free-form text or named parts (13 of 15). Recorded responses for every example, and error tables for both servers (400, 404, 413, 422, 429 and 503 with codes on the runtime) (14 of 15). The runtime contract is versioned and reports `api_version`, and the router has dated release notes. The models themselves have no tags or changelog (9 of 15). The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema): APIs and MCP servers. - 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool). - 10, llms.txt or Markdown docs served for agents. - 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference. - 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs. - 0 to 15, examples and documented error responses. - 15, versioning and a public changelog. Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference. ## 6. Agent ergonomics, 79 out of 100, up to 3.4 more on the total Why it scored 79: Read as an API an agent calls for a decision, as with Jev, Clef and Kev. Answers are probabilities and spans, `output_tokens` is always 0, and the decoders read the state once and fork it per question. Input is 16,384 tokens a sequence, or 8,192 on the 0.3B (20 of 25). The caller fixes the output shape with any mix of the five types, a `threshold` per question, a `head` override for spans and `/v1/bundle` for several tasks in one call (18 of 20). The runtime answers with `{error: {code, message}}`, and a failed question carries its own code without failing the others. The bundled server answers 422 with `loc`, `msg` and `type`, and 413 for oversize questions (16 of 20). Calls are stateless and safe to retry. The runtime answers 429 `overloaded` past 256 queued requests a model. We found no `Retry-After` header in the runtime code and no retry guidance (15 of 20). `typesafe-sdk` works against the bundled server per USAGE.md, and loading is one `AutoModel` call. Setup needs `trust_remote_code`, about 17 GB of GPU memory for the 4B, or the router's development channel (10 of 15). The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics): - 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries). - 20, pagination, filtering and output-size controls. - 20, actionable, documented error responses, codes and messages an agent can recover from. - 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations. - 15, sensible defaults, few required parameters, and official SDKs in at least two languages. Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs. ## 7. Maintenance & community, 84 out of 100, up to 1.4 more on the total Why it scored 84: Read for an open-weight model. Launched on 6 October 2026, with the four model repositories last changed on 7 October (30). Vela 1.0 on 18 September, the first Vela 2.0 upload on 29 September, re-exported weights on 3 October and the launch on 6 October (20). The router repository's 200 newest commits span 23 September to 8 October from more than eight authors, and the release note links issues the maintainers opened in the same week. 408 issues were open and we didn't check reply times (17 of 25). No SDK of its own. The cards point to `typesafe-sdk`, and the `vllm-sr` CLI on PyPI had a development build dated 8 October (10 of 15). The router repository runs CI, a security scan and a package check, with today's result on main unread (7 of 10). The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance): - 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older. - 20, at least three releases or dated changelog entries in the last 90 days. - 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15. - 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models). - 10, package health, current dependencies and CI. Models are read for deprecation notice periods and model churn rather than release counts. ## What we couldn't check What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it. - unchecked: today's CI result on the semantic-router main branch. api.github.com answered 403 for the rate limit and the Actions page we were served was stale - unchecked: reply times on issues and pull requests in vllm-project/semantic-router - unchecked: PyPI download counts for `vllm-sr`. pypistats.org answered 429 - unchecked: whether the model runtime or the `vllm-sr` CLI sends any telemetry. We found no statement and didn't audit the code - The accuracy, calibration and latency figures are the authors', on their own harness. We haven't run them - The listing covers Vela 2.0 only. Decision 2.0 and Vela 1.0 sit in the same Hugging Face organisation and aren't graded here - The 0.3B repository's LICENSING_STATUS.md is inherited from Decision-1.0-Kai and says no runtime code is bundled, while the repository does ship `vela2_inference.py` - No legal entity was found. The site footer names the vLLM Semantic Router Team, and KR Labs is credited as co-lead on the model cards ## Weaknesses - No version tags on the four Hub repositories, and the 4B and 9B weights were replaced in place on 3 October 2026 - Loading with `transformers` needs `trust_remote_code=True`, which runs Python from the model repository - The router's `vllm-sr serve MODEL` engine mode is newer than the 0.4.0 stable release and needs the development channel - By the authors' figures the 4B scores 31.63 on the Jev Decision Index 0.2.1 against 42.55 for its Decision 2.0 base - The model runtime has no authentication, and the bundled server is open unless `VELA2_API_KEY` is set ## What costs an agent a turn today The notes we give agents before they call it. Each one is a workaround an agent shouldn't need. - Pin a commit hash with `revision=` when loading from the Hub. The repositories have no tags and `main` has changed since launch - Send the served name in `model`, for example `vllm-sr/Vela-2.0-4B`. The bundled server answers 422 to any other name - Name span questions `pii`, `halu` or `toxic`, or set `"head": "router"`, to get the trained router head. Other labels go to the broad head - Keep input under 16,384 tokens a sequence (8,192 on the 0.3B). The bundled server answers 413 when the questions alone don't fit - Set `VELA2_API_KEY` before binding the bundled server beyond 127.0.0.1, and keep the model runtime on a trusted network ## When it's done Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.