Decider

# Fix list: Decider

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/decider, the October 2026 research run, assessed 8 October 2026. Grade B, 69.5 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 Decider: 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. Security & auth, 38 out of 100, up to 10.9 more on the total

Why it scored 38: Read as software you run. No account. The server has no authentication option, binds to 127.0.0.1 since 1.7.1 and the script comment says so (8 of 30). A decision model has no write actions, so there's nothing to approve (15 of 20). Nothing documents how hostile text in the state can move an answer. The README says rules written into a question aren't followed at this size (5 of 15). `/stats` counts requests, errors and rejections, and each response carries token usage. No request log (5 of 15). No SECURITY.md, disclosure policy or advisories found. Weights ship as safetensors, PyPI uploads use trusted publishing with no token in the repository, and workflow actions are pinned by tag, not commit (5 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.

## 2. 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, by the self-hosted rule. No payment protocol (0). No hosted API was found, so there's no hosted option to grade.

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.

## 3. Transparency & trust, 46 out of 100, up to 4.7 more on the total

Made of editorial 64, provenance 27.

Why it scored 46: Apache-2.0 for the code and model repositories, with the training code, data builders, teacher data and per-stage measurements published. The RL stage and the mixture-v2 builders aren't in the package, and the README says reproduction isn't byte-identical (28 of 30). Self-hosted, so inputs stay on the operator's hardware by construction, but we found no statement saying so. Training draws on about 95 public datasets whose licences we didn't review (15 of 30). The 1.0.x server was kept for one release as `decider.serve_v1` with notice in the changelog, and earlier weights stay under Hub tags. No deprecation policy (11 of 20). No telemetry code found in the package and no statement either way. Weights download through the Hugging Face Hub client (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: github.com/Mapika, no registry record we could read (0 of 15)
- Status page: not found (0 of 10)
- security.txt: not found (0 of 10)

## 4. Schema & documentation, 78 out of 100, up to 3.6 more on the total

Why it scored 78: Read for a model you serve yourself. The server is FastAPI with Pydantic request models, but `questions` is an untyped dict and `state` any JSON value, with the question rules enforced in `decider/systemone.py`. It follows TypeSafe's wire format and publishes no spec file of its own (15 of 25). No llms.txt. The README, five model cards and six docs pages are Markdown in the repository (5 of 10). The README says which model to use for which hardware and has a limits section with measured figures for each failure (19 of 20). Three question types, 2 to 255 options a choice and 2 to 10 levels a score, checked with a 422. State is free-form by design (11 of 15). Python examples with sample output and three example programs. `docs/SERVING.md` gives the 413, 503 and 422 bodies. No curl example and no error table (13 of 15). Release tags, Hub tags for earlier weights and a dated changelog (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. Agent ergonomics, 80 out of 100, up to 3.3 more on the total

Why it scored 80: Read as an API an agent calls for a decision, as for the other decision models. Answers are a probability per option, and a long state is read once with each question scored from the shared prefix. Context is 32k tokens, and a state over 32,768 tokens is truncated without an error (19 of 25). The caller sets the questions, any number a request up to 1,024 rows, with `independent` and a plain `/decide` form. No batch-of-states route (17 of 20). 422, 413 and 503 return a message that names the limit and its variable, documented in `docs/SERVING.md` (17 of 20). Calls are stateless and safe to retry, and the 503 body says to retry later. No Retry-After header (16 of 20). One pip install and a Python class with a default device order. The README says TypeSafe's SDKs work unchanged. Python only, and `scripts/serve.sh` comes from a clone (11 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.

## 6. Reliability, 90 out of 100, up to 2 more on the total

Why it scored 90: Scored on the local-package checklist, since Decider is open weights the owner runs, as for Kev and Strands Decider. `pip install decider-ai` from PyPI, 1.9.0 of 7 October 2026, with Python 3.11 or later stated and extras for serving, GGUF and Apple silicon (20). Public CI on GitHub Actions runs 26 test files in two jobs, without torch and with CPU torch, and the five latest runs on main had passed on 8 October. No test loads model weights, and the changelog reports 294 passing (22 of 25). 3 of 16 issues are open, a feature request, a Windows on ARM report with six comments and fixes shipped in 1.7.1, and one unsolicited promotion. Reported crashes (#5, #8, #21) were fixed in a release within days (23 of 25). `docs/CHANGELOG.md` has a dated entry for every release and tags match PyPI. The 1.3.0 minor release changed what `confidence` means, called out with a migration note (12 of 15). 1.9.0, though 1.0.0 is from 22 September 2026 and 21 releases followed in 16 days (13 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.

## 7. Maintenance & community, 87 out of 100, up to 1.1 more on the total

Why it scored 87: Read for an open-weight model. `decider-ai` 1.9.0 and the decider-31b weights on 7 October 2026 (30). 21 package releases since 22 September 2026 (20). 13 of 16 issues closed, with fixes released within days of reports #5, #8, #18 and #21, and outside pull requests merged (#2, #11, #13). One maintainer wrote 95 of the 98 commits under two names (21 of 25). A Python package on PyPI from the author, and TypeSafe's SDKs for the HTTP route. Not an MCP server, so no registry entry applies (10 of 15). CI passes on main. Dependencies have no upper bounds or lockfile, and the vLLM path needs its own environment pinned to vLLM 0.29.0 (6 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: the GitHub API answered with a rate limit for our address, so the star count is the rounded 1.1k from the repository page (47 forks) and the pull request list was not read
- unchecked: the licences of the roughly 95 training datasets, and the licence terms of the Gemma-4 bases under decider-12b and decider-31b, whose cards we didn't read
- The lead said the model card states it is unrelated to Strands Decider 2B. No such statement is in the repository or the decider-4b card. The independence statement there concerns TypeSafe AI
- The lead named three models. The family has more, among them decider-0.8b, decider-12b and decider-31b
- The leaderboard positions and all accuracy, calibration and speed figures are the author's or third parties' as cited by the author. We haven't run them
- No legal entity. The author is an individual, and commits appear under the names Mapika and Mark Marosi
- The slug `decider` sits beside `strands-decider`, an unrelated product with the same name

## Weaknesses

- The local server has no authentication option. It binds to 127.0.0.1 since 1.7.1
- States over 32,768 tokens are truncated (`DECIDER_MAX_STATE_TOKENS`) without an error
- No SECURITY.md, disclosure policy or security contact was found in the repository
- One maintainer wrote 95 of 98 commits, and the project is three weeks old
- The README says calibration on hard items is weak, the models are English only, and one pass can't do multi-step arithmetic
- No hosted API, and the HTTP server doesn't serve the GGUF files

## 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 weights by Hub tag (`v10`, `v2`) when results must repeat. The `main` branch of each model repository changes with new versions
- Read `x_p_max` for the top probability. Since 1.3.0 `confidence` on choice and score answers follows TypeSafe's rescaled definition, not the top probability
- Keep the server on 127.0.0.1 or put an authenticating proxy in front. It has no key option
- Count state tokens before sending. Over 32,768 the state is cut silently, and `/decide` caps context at 1,536 tokens
- Split multi-step arithmetic or multi-hop judgements into several questions, and don't write long rules into a question. The README says both fail

## 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.