Granite Guardian

# Fix list: Granite Guardian

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/granite-guardian, the October 2026 research run, assessed 8 October 2026. Grade C, 60.1 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 Granite Guardian: 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, 56 out of 100, up to 8.8 more on the total

Why it scored 56: Scored on the local-software lines, since the owner runs the model. Weights are ungated on Hugging Face, IBM publishes GGUF builds and Ollama carries `granite4.1-guardian:8b`. The card's install line is `pip install transformers torch vllm` with no versions, and the evaluation code asks for `vllm>=0.15.1` and `transformers>=4.52` and names an 80 GB GPU. No Python version is stated (14 of 20). The repository has no CI workflow and no tests. Its `evaluation` folder reproduces the card's benchmark figures, which is public but is not run automatically (5 of 25). Fourteen items are open on GitHub. Five are automated dependency alerts, one is an outside pull request, and three usage problems from May to September 2025 (issues 28, 29 and 35, on a 3.2 error and on Ollama) are open with three to seven comments each. The 4.1 Hugging Face page has no open thread (14 of 25). Versions are named 3.0 to 4.1 with a Git tag each and dated What's New notes, but there is no changelog file, and 4.1 changed the prompt format from the `guardian_config` argument of the 3.x cookbooks to a `<guardian>` block without a migration note (9 of 15). Released as 4.1 with no preview label. IBM calls model signing experimental (14 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, 58 out of 100, up to 7.4 more on the total

Why it scored 58: Read as software the owner runs. No account, key or token is needed to get or run the weights, so there is no secret to leak, and any server authentication is the owner's to add (14 of 30). A classifier with no write actions (15 of 20). It reads untrusted text by design and has a built-in jailbreaking criterion. The card says the model may be prone to adversarial attacks and that reasoning traces may contain unsafe content, with no mitigation guidance beyond keeping to the scoring mode (8 of 15). No call log of its own. Think mode gives a reason for each judgement (6 of 15). IBM's security.txt is valid and names PSIRT and HackerOne, the weights are safetensors with a sigstore signature and published verification steps, and dependency scanning was added to the repository on 19 August 2026. The repository has no `SECURITY.md` (issue 20, open since 9 February 2025), and five scanner alerts and an outside bypass report of 19 July 2026 (issue 46) have no reply (15 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, 60 out of 100, up to 6.5 more on the total

Why it scored 60: Read for a model the owner serves. There is no specification of its own. The contract is `chat_template.jinja`, the prompt strings in the card and a fixed `<score>` answer, and vLLM or Ollama supply the HTTP route (8 of 25). www.ibm.com/granite/docs has no llms.txt (the path returns the docs home page). The model card and README are Markdown on Hugging Face and GitHub (5 of 10). The card states the intended use, eleven built-in criteria with definitions, English-only training, the need to test custom criteria, and points to the HAP-38M model for tighter latency (18 of 20). The answer is binary inside fixed tags, and documents and tool schemas have their own template arguments. The criterion is free text (10 of 15). Four code examples, two notebooks for 4.1 and an Ollama page. Nothing documents failure modes such as a missing score tag (10 of 15). Versioned model names, Git tags and dated What's New notes, with no changelog file (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.

## 4. Agent ergonomics, 69 out of 100, up to 5 more on the total

Why it scored 69: Read as a classifier an agent calls. Without thinking the answer is a yes or no in tags. Think mode adds a reasoning trace, and the card's examples allow 2,048 output tokens. Each call judges one criterion (19 of 25). The caller chooses think or no-think, the criterion, whether the last user or assistant turn is judged, and passes documents and tool schemas separately (17 of 20). The model has no error responses of its own and the 4.1 card documents no confidence value, so a caller parses tags from text (7 of 20). Calls are stateless and the examples set temperature to 0, so a check can be repeated safely (16 of 20). No gate and one `ollama run` command, but the prompt strings and parser must be copied from the card because no package ships them, and vLLM needs a GPU (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.

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

Why it scored 60: Self-hosted rule. Nothing to buy for the model, 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's watsonx.ai sells only the deprecated `ibm/granite-guardian-3-8b` at $0.0002 per 1,000 tokens, which is not the model graded here.

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. Maintenance & community, 47 out of 100, up to 4.6 more on the total

Why it scored 47: Read for an open-weight model. The 4.1 weights date from 29 April 2026, 162 days before the check. The card and signature changed on 26 and 27 August 2026, which we do not count as a release (10 of 30). Main releases came in October and December 2024, February and August 2025 and April 2026. The last 90 days brought the evaluation code on 10 August and a cookbook refactor on 19 August, and no model (10 of 20). Maintainers merge their own pull requests within a day and closed the evaluation-code request (issue 26) on 10 August 2026 after 17 months. An outside pull request of 12 August, the bypass report of 19 July and the 4.2 question of 17 September have no reply (10 of 25). `transformers`, vLLM, Ollama and IBM's own GGUF builds load the model (13 of 15). Dependency scanning is on, its five alerts are open, and there is no CI (4 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.

## 7. Transparency & trust, 70 out of 100, up to 2.6 more on the total

Made of editorial 67, provenance 73.

Why it scored 70: Weights and repository are under the Apache 2.0 licence with no gate and no use policy attached (30 of 30). Inputs stay on the owner's hardware, so no privacy statement applies. Training data is described as hh-rlhf samples annotated at DataForce plus synthetic data, and the datasets are not released (issue 27, open since 28 March 2025) (22 of 30). No deprecation policy for the open weights. Older versions stay downloadable, and watsonx.ai marks its hosted 3.0 model deprecated (5 of 20). The weights carry no telemetry of their own, but we found no statement either way and downloads go through Hugging Face or Ollama (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)
- Changelog: not found (0 of 10)

## 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.

- The lead named version 3.3. Granite Guardian 4.1 8B (April 2026) is the current model, and the listing grades it. The lead's interface and licence were right.
- unchecked: we did not run the model, so the install lines, GPU memory and whether current `transformers` and vLLM releases load `GraniteForCausalLM` are from IBM's pages only.
- unchecked: who replied on issues 20, 28, 29 and 35 and when. The GitHub API allowance for this network ran out, so only comment counts were read.
- unchecked: the date watsonx.ai withdraws `ibm/granite-guardian-3-8b`, and whether 4.1 is on Azure AI Foundry or Replicate. www.ibm.com/docs answered 403 and the partner lists were not read.
- unchecked: IBM's Granite Trust portal, which the docs give as the home of its responsible AI material.
- Whether a Guardian model based on Granite 4.2 is planned. Issue 54 asks and has no reply.
- Issue 46 is an outside, unverified report against the 3.0-2B model prompted as a chat model, which is outside the scoring mode IBM prescribes. No deduction was taken.
- The 4.1 card lists sexual content in its scope of use but not in its list of built-in criteria. IBM's docs page and the Ollama page list it.

## Weaknesses

- Trained and tested on English only, per the model card
- Each call judges one criterion, so checking several risks takes several calls or a separate LoRA adapter built on the 3.2 model
- Version 4.1 moved the criterion into a `<guardian>` block in the last user message, where 3.x cookbooks pass `guardian_config`
- The repository has no CI, no tests and no `SECURITY.md`, and an issue asking for one has been open since 9 February 2025
- The card's out-of-distribution safety F1 is 0.79 without thinking, below the 0.81 it reports for version 3.3
- IBM's watsonx.ai model list has only `ibm/granite-guardian-3-8b`, marked deprecated, so 4.1 has no hosted endpoint from IBM that we found

## 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.

- Append the `<guardian>` block as the final user message, with the mode line, `### Criteria:` and `### Scoring Schema:`. Copy the strings from the model card, because no package builds them
- Use the no-think instruction for gating and parse `<score>`. Think mode writes a reasoning trace first, and the card's examples allow up to 2,048 output tokens
- Treat `yes` as the criterion being met, which for built-in criteria means the risk is present. Treat a missing `<score>` tag as a failed check
- Pass retrieved text through `documents=` and tool schemas through `available_tools=` in `apply_chat_template`, not inside the message text
- Under Ollama, set `num_ctx` in the request options. IBM's docs say the default context is short and long requests are truncated

## 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.