← Deemwar decision model · All recipes
Eighteen everyday decisions an agent or a script hands to OpenJevX, each as a call you can run against your own
server. Start the server (task run, the Docker image, or a release binary), then copy any call below. The
server listens on localhost:21118 by default. These pages are also in the Docker image as files, in
/usr/share/openjevx/recipes, and on the server at /recipes.
Start with what you get: model size, measured CPU latency, and the three question types.
Every recipe is one POST /v1/systemone with a state (the text or JSON the question is about) and named
questions. There are three question types:
| type | ask | criteria | read the answer from |
|---|---|---|---|
noul |
yes/no | none (optional {"true": "...", "false": "..."}) |
noul: probability of yes |
choice |
pick one | {"key": "description", ...} |
choice and confidence |
score |
rating | ["lowest", ..., "highest"] |
score (expected level, 0 = first) and probabilities |
Each recipe was run against a local OpenJevX server (v0.4.0 8-bit model, CPU, Apple M5 Pro) and the output was
pasted unedited, including the wrong and unsure ones. Each page ends with how the local model did. The jevx
lines ran through the jevx CLI with --profile openjevx and --no-context (so your agent files do not change
the question) and --fresh (so no cached answer was reused); the tables and --fresh need jevx v0.11.0 or newer. Your inputs will score differently: copy the pattern, not the numbers, and test on your own data.
| recipe | decision |
|---|---|
| Find the failures in a log | Yes/no per log line: keep only the lines an on-call engineer would act on. |
| Route a ticket queue | Pick-one per ticket: which team owns it. |
| Rank search results | Yes/no per result, sorted by the probability of yes. |
| Pick a value among candidates | Pick-one: code parses the candidate dates, the model chooses which one is the due date. |
| Check a claim against the source | Yes/no: does the document say what you are about to write? |
| Before asking the user "A or B?" | Pick-one: what would the user want next, given what they already said. |
| Is this command dangerous? | Yes/no about a command before running it. |
| Prompt injection in a fetched page | Yes/no: does fetched text try to instruct an AI assistant? |
| Which file to open first | Yes/no per file path, sorted by the probability of yes. |
| Flaky test or real bug? | Pick-one per CI failure. |
| Sort PR review comments | Pick-one per review comment, four ways. |
| Do it here or hand it to a smaller model? | Rating: how much reasoning a task needs. |
| Question or instruction? | Pick-one: answer the message, or change code? |
| Secret in a file before writing it | Yes/no on a line before it is written to disk or a commit. |
| Did the build succeed? | Yes/no over build output before reporting "done". |
| Which tool should handle this? | Pick-one among function names (function calling from natural language). |
| Several judgements in one call | One request, three questions of the three types. |
| Would the user accept this turn? | Yes/no and rating about the agent's own final message, before handing back. |
| you get | the agent should |
|---|---|
noul at or above 0.8, or confidence at or above 0.6 for a choice or score |
act on it, and say so in its report |
noul at or below 0.2 |
act on the no |
| anything in between | rewrite the question more narrowly, check it itself, or ask the user |
| an HTTP error | report that the call failed; never treat it as a no |
These are the jevx CLI's default thresholds. When the answer is unsure, the fix is usually a narrower question:
name the exact thing you are looking for, say what counts as yes, and let code do arithmetic and dates.
jevx caches answers for 7 days, keyed by the profile's model name. After upgrading the OpenJevX model, run
jevx cache clear, or give the profile a versioned model name (for example openjevx-0.5.0), so you do not get the
old model's answers back.