Yes or no gate before an action runs
Four or more companies' docs show a full fit.
What counts as done
A probability of yes that can be thresholded.
Measured, for our own tests: at the chosen threshold no irreversible action proceeds on a wrong yes in the caller's labelled set, under 300 ms. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: is this tool call safe to run
, does this message authorise the booking
, is this request a prompt injection
, does this line affect driving safety
Does the job 11
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Celeris-1 Decision D Marqo Inc | noul returns the probability of yes; the Policy gates section says to ask whether the input violates the policy and pick thresholds from labelled examples. 131,072-token context, so a long policy fits. source |
| Clef B Cloudflare, Inc. | noul returns the probability of yes; 65,536-token context on Clef. Arize found Clef kept 0.92 and 0.91 AUC with the question worded both ways, one of three models that did. No prompt-injection guidance from Cloudflare. source |
| elephant-1 D Baseten Labs, Inc. | noul returns P(yes); trained at 16,384 tokens with longer requests refused, a trained 'I don't know' option and 16 percent safety and abstention data. No prompt-injection guidance. source |
| joey-1 C Baseten Labs, Inc. | noul returns P(yes); trained at 16,384 tokens with longer requests refused, a trained 'I don't know' option and 16 percent safety and abstention data. No prompt-injection guidance. source |
| mammoth-1 D Baseten | noul returns P(yes); trained at 16,384 tokens with longer requests refused, a trained 'I don't know' option and 16 percent safety and abstention data. No prompt-injection guidance. source |
| Drex 1.5 BB Nace AI, Inc. | noul with true and false criteria; drex-v2.0 reads states to 131,072 tokens and refuses longer ones with a 422 naming the limit, so a long policy need not be compressed. source |
| OpenAI Decisions API A OpenAI OpCo, LLC | predicate returns P(true); input can carry conversation history, tool calls and tool outputs; thresholds from labelled examples; 922,000 input tokens on GPT-6 Luna. Arize found it kept 0.90 and 0.89 AUC with the question reversed. No adversarial-input guidance. source |
| pplx-decider BB Perplexity AI, Inc. | noul with criteria, 256K context, and published calibration (ECE 0.014, Brier 0.134 on the development set; about 0.05 on held-out decision datasets). No injection guidance. source |
| pplx-decider-flash B Perplexity AI, Inc. | noul with criteria, 256K context, calibration published (ECE 0.021, Brier 0.163 on the development set). source |
| kettei C Sqwish Ltd, trading as Sqwish Labs | binary decisions with abstain_below (examples at 0.4 to 0.7) and a tool-call check example; 65,536-token input limit with a 422 context_too_long rather than a cut. source |
| Jev B TypeSafe AI, Inc. | noul with 0.7 and 0.3 as starting thresholds and a worked risk-tiered example (0.5 floor, 0.9 for a transfer); OpenRouter's screen caught 40 of 40 injections; the jaggedness page says adversarial content can steer answers; 64k context, 262k by header. source |
Does the job, with a limit 13
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Decider BB Mark Marosi (Mapika) | noul with per-type temperatures, but the README says rules written into the question are not followed at this size (0.24 against 0.67) and calibration on hard items is weak; the maintainer advised acting on destructive actions only at 0.8 or above. source |
| Decision 3.0 B vLLM Semantic Router project | noul with a probability of yes and no truncation, but issue 6298 found the vela3 and d3 sizes used as jailbreak and harm guards missed attacks the base model refused. source |
| Drex Max v1 C Nace AI, Inc. | noul with a 16,384-token default raisable to 262,144 (DREX_MAX_CONTEXT). No safety results, no prompt-injection guidance and the licence cap. source |
| Granite Switch 4.1 B International Business Machines Corporation | A guardian-core safety adapter and a decision adapter share the endpoint, with the 128K base context for a long policy; there is no typed noul, so the gate is a two-label logprob read. source |
| Kev BB Jared Palmer | noul with calibrated probabilities and a 65,536-token server limit, but Arize found Kev 9B's AUC fall from 0.85 to 0.54 when the question was reversed, and the small models are validated to 8,192 tokens. source |
| JEV-27B-VL C AutoTrust AI Lab | Yes/no decisions over up to 256K tokens of text (20 of 20 at every depth tested), one question a request, with no safety results or injection guidance. source |
| Liquid d1 C Liquid AI, Inc. | noul with a worked binary-decision example in the guide, but Arize saw d1's AUC fall from 0.92 to 0.17 when the question was reversed, so the gate must be tested both ways. source |
| Microsoft-Decision-1 E Microsoft | noul with criteria for true and false, 32,768-token context and a vendor safety test over 5,250 requests; no refusal type, and 0.42 s on OpenRouter is over the 300 ms target. source |
| RightNow System Zero B RightNow AI, Inc. | boolean returns the probability of true with a refusal per question, over 38,000 state tokens, but the probabilities are flagged uncalibrated, so a threshold has no stated meaning. source |
| ScaleDown E Carbon ScaleDown, Inc. | noul returns an independent P(yes); no context limit, calibration figure or refusal type is documented, and no key without sales contact. source |
| SPEV D Sprag AI, Inc. | predicate returns P(true) with a refusal type and a 400 model_limit_exceeded naming the limit; the weights' config caps input at 8,192 tokens, short for a long fixed policy. source |
| Strands Decider 2B B Amazon Web Services, Inc. | noul works, but at a 0.95 threshold no booking proceeded and at 0.50 two unauthorised ones passed; injection recall 92.5 percent at 17 percent false positives, weak on encoded attacks; a question naming tyres fired on every 'tire' line; the 4,096-token window blocks a 10k-token policy. source |
| Vela 2.0 BB vLLM Semantic Router project (with KR Labs) | noul plus trained safety, PII and hallucination span heads with AUC over 14 safety sets, but issue 6298 found the trained adapters missed attacks the base model refused. source |
Doesn't do this job 2
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| Laya BB Convai Innovations Pvt. Ltd. | noul exists, but Arize scored Laya 57.3 percent on evaluator suites with calibration error 0.330 and a toxic-content question flipped from 0.11 to 0.87 AUC on rewording; an HN tester saw a die roll of 3 come back 9 percent odd at 91 percent confidence. source |
| GLiClass C Knowledgator Engineering Ltd. | No predicate type with a calibrated P(yes); labels get softmax or independent sigmoid scores, F1 is the only published metric and issue 12 shows saturation at 0.99. source |
1 listings in the category were added after this reading and aren't marked yet.
What stops agents finishing
- disguised or encoded attacks pass a text-only gate
- threshold at 0.9 blocks every legitimate action, at 0.5 lets unauthorised ones through
- 4,096-token window too short for a long fixed policy
Around it
Before: agent-frameworks.propose-tool-call, chat.receive-message. After: sandboxes.run-command, human-review, agent-frameworks.refuse-tool-call.
Where this comes from
Weight 16 of 100 in Decision models: 8 of 41 accounts; the AWS nine-desk test, two GitHub issues on wrong yeses, the Vela guard issue and the Dataiku exception threshold work
The same as data: /api/v1/use-cases/decision-models.gate-action.json, or the MCP tool get_use_case with id decision-models.gate-action. Other jobs in this category: /api/v1/niches?category=decision-models.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.