Granite Switch 4.1

# Fix list: Granite Switch 4.1

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/granite-switch, the October 2026 research run, assessed 17 October 2026. Grade C, 59.4 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 Granite Switch 4.1: 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, 46 out of 100, up to 10.8 more on the total

Why it scored 46: Scored on the local-software lines, since the owner serves the checkpoint with vLLM. The README installs with `pip install "granite-switch[vllm]"` and states Python 3.11 or later and PyTorch 2.11 or later, while the model card installs from a Git clone and states Python 3.9 or later and vLLM below 0.28, against 0.26 to 0.30 in pyproject.toml. PyPI answered with a bot check, so the package page was not read (12 of 20). Public CI runs pre-commit and CPU unit tests on Python 3.11 and 3.12 on every push and pull request, and GPU suites run on a self-hosted runner when a maintainer asks. There are 102 test files. GitHub's pages were not readable from our network, so whether main is passing is unchecked (15 of 25). Open crash and regression issues couldn't be read for the same reason. Neither Hugging Face repository has a community thread (10 of 25, unchecked). CHANGELOG.md follows Keep a Changelog and semver but has one entry, 0.1.0 MVP, the 15 October bump to 0.1.1 has none, and the repository has no tags (4 of 15). The checkpoints carry no preview label, but the composer is 0.1.x and its changelog says the API is not yet stable (5 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, 54 out of 100, up to 8.1 more on the total

Why it scored 54: Read as software the owner runs. No account, key or gate, so there is no secret to leak, and SECURITY.md tells operators to add authentication to the vLLM server (14 of 30). No write actions (15 of 20). It reads untrusted text and audio by design and carries a safety adapter with a jailbreaking criterion. The card warns that a mistranscription can silently change what an adapter answers and that the guardian adapter is no substitute for safety testing. Nothing documents how hostile text in the state can steer a decision (7 of 15). No call log of its own beyond vLLM's. Decisions return a probability for every option (5 of 15). SECURITY.md takes private reports through GitHub or the maintainers, IBM's security.txt is valid with PSIRT and HackerOne, commits need a DCO sign-off, and the decision adapters in Granite Libraries are signed with certificates. The composed 4.1 checkpoints have no signature file, where the 8B preview had one from 4 May 2026 (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. Schema & documentation, 55 out of 100, up to 7.3 more on the total

Why it scored 55: Read for a model the owner serves. The route is vLLM's OpenAI chat-completions API, and the adapter contract is the chat template, `adapter_index.json` and one `io.yaml` per adapter. There is no specification file of its own (10 of 25). We found no llms.txt. The model cards, BUILD.md, the adapter card and the repository docs are Markdown on Hugging Face and GitHub (5 of 10). The card says what each adapter does, that decision probabilities are a prior and not a certified classification outside trained domains, that the guardian adapter is no substitute for application testing, and the speech limits (16 of 20). A decision request is free-form prompt text with token ids the caller computes, and the guardian criterion is free text. The adapter card defines choice, noul and score types, but nothing on the endpoint enforces them (6 of 15). Python and curl examples for both adapters and audio, a chunking recipe for wide questions, and two documented HTTP 400 cases (11 of 15). Versioned model names, Hugging Face commit history and a changelog with one entry (7 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.

## 4. Agent ergonomics, 66 out of 100, up to 5.5 more on the total

Why it scored 66: Read as an endpoint an agent calls for a decision, as with Vela and Decision 3.0. A decision returns one token and its label logprobs, a guardian check up to 16 tokens, and the state turn can share vLLM's prefix cache with chat. Context is 131,072 tokens (21 of 25). The caller picks the adapter per request, the options and the label ids. More than 128 options need one request per 128 (13 of 20). Errors are vLLM's, with the 128-id limit and the audio prompt-length error documented. A guardian answer is JSON in text with no stated handling for a malformed reply (8 of 20). Requests are stateless and the examples set temperature 0, so a decision can be repeated safely (16 of 20). The OpenAI SDK works, and Mellea wraps the guardian adapter, but the decision adapter has no Mellea call yet, label ids need the tokeniser, and vLLM needs a GPU (8 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.

## 5. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 60: Self-hosted rule. Nothing to buy, and the Apache 2.0 licence allows commercial use at no charge (20). Free to run with no card (20). The weights are not gated, so an agent can fetch and run them with no sign-up (20). No payment protocol (0). IBM hosts no endpoint for Granite Switch that we found, and Hugging Face lists no inference provider for it.

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.

## 6. Transparency & trust, 74 out of 100, up to 2.3 more on the total

Made of editorial 61, provenance 87.

Why it scored 74: Weights, adapters and code are Apache 2.0 with no gate, on an Apache 2.0 Granite 4.1 base. The decision adapter card adds that base-model use is also governed by the Granite model terms (28 of 30). Inputs stay on the owner's hardware, so no privacy statement applies. Neither the Switch cards nor the decision adapter card describe training data (18 of 30). No deprecation policy. The preview repositories stay online, but the 8B preview was overwritten in place on 3 May and 15 October 2026 (5 of 20). No telemetry code found in the package source and no statement either way. Weights download through Hugging Face (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):

- Status page: not found (0 of 10)

## 7. Maintenance & community, 81 out of 100, up to 1.7 more on the total

Why it scored 81: Read for an open-weight model. The 8B and 30B checkpoints were published on 14 October 2026, three days before the check (30). In the last 90 days, the 8B and 30B releases on 14 October, the decision adapters in `granitelib-core-r1.0` and the 8B preview rebuilt with audio on 15 October, and repository changes including the default speech encoder on 30 September (20). The repository's 300 newest commits come from more than ten authors, with pull requests numbered to 148 and merged within days. Issue reply times couldn't be read (12 of 25, unchecked). An official Python package with a vLLM plug-in, and Mellea for the guardian adapter. The decision adapter has no Mellea call yet (11 of 15). CI, a lock file and dependency ranges that track vLLM 0.26 to 0.30 (8 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.

- firstReleased is 1 May 2026, when the 3B, 8B and 30B preview checkpoints were first uploaded. The 8B and 30B release with the decision adapter and speech is 14 October 2026, and the launch post is 15 October. The scout's dates for the launch were right.
- The scout said the post claims Granite Switch can stand in for `/v1/systemone` calls. The post attributes that to native llama.cpp support. The repository has no `/v1/systemone` route and the cards document only chat completions. llama.cpp support was not checked, and no deduction was taken.
- unchecked: the granite-switch package page on PyPI, which answered with a bot check. The README installs from PyPI and the model card from a Git clone.
- unchecked: GitHub stars, open issues, reply times and whether CI on main is passing. GitHub's web pages were not readable from our network, so these come from the clone only.
- unchecked: we did not run the model, so the install lines, GPU memory and the vLLM plug-in's behaviour are from IBM's pages and source only.
- The decision adapter's training data is not described in the Switch cards or the adapter card.
- IBM's Decision Index figures (42.9 on the 8B, 50.1 on the 30B) are IBM's, on the public fast subset. The adapter card gives 42.56 and 50.91 for the same aLoRAs, on the full v0.3 public set by its heading.

## Weaknesses

- The decision adapter has no typed request format. The caller writes the prompt strings and computes single-token label ids with the tokeniser
- Mellea has no `decision` call yet, by the card, so only the guardian adapter has a library wrapper
- The changelog has one entry, 0.1.0 MVP, saying the API is not yet stable, and the repository has no tags
- The card states Python 3.9 or later and vLLM below 0.28, while pyproject.toml states Python 3.11 to 3.13 and vLLM below 0.31
- The speech encoder is English only and not trained jointly with the model, per the card

## 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.

- Pass `chat_template_kwargs: {"adapter_name": "decision"}` or `"guardian-core"`. Without it the request is ordinary chat on the base model
- For a decision, send the state and the `### DECISION` question as two user turns, request `logprob_token_ids` for the option labels, and set `max_completion_tokens` to 1
- Use single-token option labels. Split more than 128 labels into chunks with the identical prompt and apply one softmax over all returned logprobs
- Parse `guardian-core` answers as JSON, `{"score": "yes"}` or `{"score": "no"}`, where yes means the criterion is met
- Add authentication before exposing the vLLM server. Neither vLLM nor the checkpoint requires a key by default

## What the review panel asked for

- tagged releases
- a new repository for each rebuild
- Publish GPU memory sizing
- Add a typed decision request
- Publish a changelog
- sign the composed checkpoints
- document injection effects

## 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.