Decision 3.0

# Fix list: Decision 3.0

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/decision-3, the October 2026 research run, assessed 12 October 2026. Grade B, 63.1 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 Decision 3.0: 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, 58 out of 100, up to 8.4 more on the total

Why it scored 58: Scored on the local-package checklist, since Decision 3.0 is a model the owner runs. The cards install with `pip install "transformers==5.17.0" torch ...` and load with `AutoModel.from_pretrained(..., trust_remote_code=True)`, and requirements.txt says the runtime was validated with transformers 5.17.0 on ROCm. No Python version is stated. The router's `vllm-sr` package states Python 3.10 or newer, but serving Decision 3.0 needs Engine mode, which is newer than the 0.4.0 stable release (12 of 20). The semantic-router repository runs public CI, and its model runtime has a `test_decision3_reference.py` suite and a parity record of byte-identical answers against the packages' own server on CPU and an MI325X, dated 11 October. Today's CI result on main was unread (15 of 25). The models are two days old and the Hub repositories have no discussions to read. We couldn't read issue reply times on the router repository (12 of 25). Each Hub repository carries tags v3.0.0 to v3.1.0, and the router pins a commit per size. There's no changelog in the model repositories, and five tagged releases in two days (9 of 15). Released as 3.0 with no beta label, though only two days old (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, 56 out of 100, up to 7.7 more on the total

Why it scored 56: Read as software the owner runs. No account. The bundled server binds to 127.0.0.1 and is open unless `DECISION_API_KEY` is set, after which it requires one bearer key, compared in constant time. The router's model runtime has no authentication (13 of 30). A decision model has no write actions (15 of 20). The server accepts images and videos only as base64 data URLs within byte, pixel and duration caps, and refuses paths and URLs, so it fetches nothing. The in-process Python API does fetch http(s) URLs it is given. Nothing documents how hostile text, images or video in the input can move an answer (8 of 15). The server sets `x-request-id` and `server-timing` headers and keeps no log of calls. The runtime exposes metrics and says request text is never logged (8 of 15). SECURITY.md in the router repository takes private reports through GitHub Security Advisories. Each package ships MODEL_MANIFEST.json with SHA-256 for every file, which the runtime checks at load (by default only files under its fast-hash size are hashed). The `transformers` route runs repository code. No bug bounty and no security.txt on vllm-sr.ai (12 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 from the authors. 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. Schema & documentation, 70 out of 100, up to 4.9 more on the total

Why it scored 70: Read for a model the owner serves. The router's model runtime has an OpenAPI file, contract version 2.3.0, changed on 11 October, whose decision request carries `images` and `videos` for models with those modalities. The bundled `d3_server.py` has no specification file and documents its wire format in its docstring (22 of 25). vllm-sr.ai has no llms.txt (404). The cards and the server docstring are Markdown and plain text in each repository (5 of 10). The cards state what each size is, its inputs and limits, and the authors' evaluation. They say nothing about when not to use it (13 of 20). Three question types, choice with up to 255 options, noul and score, and an `images` or `videos` list of data URLs with stated byte and pixel caps. State is free-form text or JSON (12 of 15). Three worked examples (text, an image, a video) with downloadable inputs, but no recorded outputs and no error table. The server docstring names 401, 422 and 503 (9 of 15). Hub tags and a versioned runtime contract. No changelog for the models (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.

## 5. Transparency & trust, 45 out of 100, up to 4.8 more on the total

Made of editorial 62, provenance 27.

Why it scored 45: Apache-2.0 for the weights and runtime, on Apache-2.0 Qwen3.5 and Qwen3.8 bases, with a NOTICE in each repository. Unlike Vela 2.0, the cards publish no training data or training stages, and we found no training code (24 of 30). Self-hosted, so inputs stay on the owner's hardware unless the Python API is given URLs to fetch. The runtime docs say request text is never logged. There's no privacy statement for the project (20 of 30). The router's support matrix has a Deprecated class, and Decision 1.0, 2.0 and Vela stay built in. No policy with notice periods (8 of 20). We found no telemetry statement either way and didn't audit the runtime 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)

## 6. Agent ergonomics, 72 out of 100, up to 4.6 more on the total

Why it scored 72: Read as an API an agent calls for a decision, as with Vela, Decider and Kev. Answers are probabilities, no text is generated, and every question sees the same text, images and videos. Videos are capped at 16,384 tokens a request, and the package config sets no other input limit (20 of 25). The caller fixes the output with any mix of the three types, and `permutation_average` scores choice questions in both option orders. There's no threshold field and no set or span type (15 of 20). The server answers 422 with a message, or an `invalid_questions` map naming each bad question, and refuses an over-long request whole rather than truncating it. The runtime answers `{error: {code, message}}` (15 of 20). Calls are stateless and safe to retry. The bundled server takes one request at a time and has no 429 or retry guidance (13 of 20). Loading is one `AutoModel` call or a `pipeline("decision")`, with no SDK of its own. It needs a GPU for the evaluated settings, `trust_remote_code`, or the router's development channel (9 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, 82 out of 100, up to 1.6 more on the total

Why it scored 82: Read for an open-weight model. Created on 10 October 2026, with v3.1.0 on 11 October (30). Five tagged releases of each model in two days, after the team's Decision 1.0 post on 22 September and Vela 2.0 on 6 October (20). The router repository merged Decision 3.0 image support (#4862) and video and d3-edge support (#4868) on 11 October, and the `vllm-sr` CLI had four development builds on 11 October. We didn't check reply times on issues (15 of 25). No SDK of its own, and the router serves it only on the development channel (10 of 15). The router repository runs CI, 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, and issue reply times. The GitHub API wasn't available to this session
- unchecked: GitHub stars for vllm-project/semantic-router on 12 October. The Vela 2.0 listing recorded 6,054 on 8 October
- The accuracy, vision, video and latency figures are the authors', on their own harness. We haven't run them, and no calibration figures were published
- The lead described the probabilities as calibrated. Neither the cards nor the runtime make that claim, and the readout temperature is 1.0
- d3-edge (0.59B) is built into the router and public on the Hub but missing from the collection and has no card figures we read
- The Hub API lists tag v3.1.0 on d3-mini at commit b5598a4 while main and the router pin a475367. We didn't establish why
- Decision 2.0 and Decision 1.0 sit in the same Hugging Face organisation and aren't graded here

## Weaknesses

- Released on 10 October 2026, with five tagged releases of each model by 11 October
- No calibration figures on the cards, though the category compares models on calibration
- Only choice, yes-or-no and score questions. No multi-label set or span answers
- Loading with `transformers` needs `trust_remote_code=True`, and the router's Engine mode needs its development channel
- No training data, training code or changelog published with the models

## 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 tag or commit with `revision=` when loading from the Hub. Each model had five tagged releases on 10 and 11 October
- Send images and videos to `d3_server.py` as base64 data URLs only. The server refuses paths and http URLs with 422
- Set `DECISION_API_KEY` before binding the bundled server beyond 127.0.0.1. It is open otherwise
- Turn on `permutation_average=True` when routing must not depend on option order. It is off by default
- Expect one request at a time from the bundled server. Use the router's model runtime for queued serving

## 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.