---

The problem you already have

Put an AI agent to work and it will tell you things. Some of them are true.

"Done. I researched all five vendors and wrote the comparison."

No file was written. Two vendors were checked. The "comparison" is your own request, rearranged into a sentence that sounds like an answer.

That is not a hypothetical. It is the single most common way an autonomous agent fails, and it fails that way quietly, in a tone of complete confidence, in the same words it uses when it succeeded.

The uncomfortable part: you cannot catch it with code. Whether a claim matches its evidence is a judgment, not a fact. So most teams do the only obvious thing, ask a large model in prose, parse the prose, invent a threshold on the spot, and keep no record.

That holds up until you have four such decisions. Then the same magic number lives in four files, no two agree on what to do when the answer is uncertain, and nobody can tell you what your agent decided yesterday or why.

What JDE does

const outcome = await ask({
  decision: "completion-check",
  state: { task, taskParts, claimedResult, receipts },
  context: { agentId, turnId },
});

if (outcome.action === "accept") deliver(result);
else redispatch(outcome.answers);          // it did not do the work, and now you know

Four things, each of which you would otherwise build yourself:

The proof

The first decision shipped is the one above: did the agent actually do the work?

It is measured on a public blind set: 36 cases written by an independent author who never saw JDE's questions or any result, frozen by sha256 and pushed before any judge saw a case (how). The judge is Jeb, the open Jebadiah decision model, on your own machine. Every number below is in a file under evals/results, with the GGUF's sha256, the runtime and its version, and the commit that ran it, and npm run eval:verify reruns it and fails on any difference.

Each size gives the same answers, to the last digit, on Ollama 0.34.4 and on llama.cpp 0.5.0, and the same answers again on a rerun. Latency is one completion check at a time on a Mac Studio (M3 Ultra). No cost per call, and nothing leaves the machine. Verdicts are aggregated in code from the per-part answers; file parts are settled in code and are right by construction (19 of 19).

Where the judges disagree. Each local size misses five verdicts. Two misses are shared by all three, and they are one mistake: a receipt that looks like the work is taken for the work. Search results counted as pages read (pub-search-without-reading), and tests run against the wrong service (pub-wrong-service-tests). 9B and 27B also accept a recommendation without the reason that was asked for (pub-hosting-missing-reason), and judge two polished restatements of the task partial rather than not done, which hosted Jev also did on the same two. 4B gets those three right and instead calls two finished tasks partial and one restatement partial. The hosted judge is better here, by three verdicts. It is also a service: its weights can change behind the same name, it needs a key, and nobody outside TypeSafe can rerun it at a fixed version, which is why it is a comparison row and not the headline. evals/README.md has the method and the case-by-case table (node scripts/results-table.mjs prints it from the files).

JDE's first measurement, 30 of 30 verdicts with hosted Jev, was on a private blind set that stays private: once a model can train on a blind set it stops measuring anything. The public set exists so that nobody has to take that number on trust.

Why it works: ask small questions

This is the lesson that produced those numbers, and it is worth more than the code.

On JDE's first, private blind set, with the hosted judge, asking one question about a whole task scored in the seventies. Splitting the same judgment into one question per part of the task, settling the parts that are facts in code, and adding them up in code, scored 100 percent. The public set above is the same shape, and it is what the local judges are measured on.

Give the model the semantic step in the middle. Keep the arithmetic.

Whether a file exists at a path, whether a count exceeds a budget, whether every part passed: none of those need a model, and asking one makes them worse.

Principles, and what they cost to learn

A band for every case, including the bottom. There is no undefined confidence, so there is no path where your code does not know what to do.

A timeout and a fallback are required, not optional. A judge that does not answer must never change what your agent would otherwise have done. Degrading to your old behaviour is always allowed.

Aggregation belongs to code. The judge answers one thing at a time and never sees how the answers combine.

Facts stay in code. If an if can answer it, an if should.

Every new judgment needs a blind case set written by someone who did not write the questions. Every judgment in this engine got measurably better after a blind set caught wording we had quietly tuned ourselves into.

Quickstart

npm i @titanium-devops/jde
pip install jebadiah-decide && jeb serve     # the judge, on your own machine

import { completionCheck } from "@titanium-devops/jde";
const { verdict } = await completionCheck({ task, task_parts, claimed_result, receipts });
if (verdict === "partial" || verdict === "not_done") redo();   // null: the judge did not answer

task_parts is the task split into its parts, and receipts is what the run's own tool log shows it did. USAGE.md has the shapes.

Use it from an agent

No code: give the agent JDE as an MCP server.

claude mcp add jde -- npx -y -p @titanium-devops/jde jde-mcp

The agent gets check_completion, which its description tells it to call before it reports a task done, and ask for typed questions. Cursor and other clients: examples/mcp. In your own agent loop: OpenAI Agents SDK and LangGraph, each checking the run's receipts and sending the agent back when the work is not done.

Run the tests

npm install
npm test          # no network, no API key

Everything offline runs against the code judge, which is a judge you hand the answers to. You can test your policy, your thresholds and your fallbacks without a model, a key, or a network connection.

The judge

For real judgments, JDE asks Jeb, the open Jebadiah decision model, running on your machine. Start one:

pip install jebadiah-decide
jeb serve          # Ollama by default; also LM Studio, llama.cpp, vLLM, MLX. See jeb --help

That serves the judge at http://localhost:8100/v1/systemone, which is where JDE looks by default. No key, no account, nothing leaves the machine. Then measure it yourself:

npm run eval:local      # every public set, with full provenance, into evals/results/
npm run eval:verify     # rerun a committed result and fail on any difference

If nothing answers there, every decision takes its policy's on_error action, and the recorded failure says how to start a local Jeb. JDE never falls back to a hosted service on its own.

Judge Jeb is coming, and isn't released yet. It's a Jeb tuned on JDE's own judging questions. When it ships, moving to it is the one JDE_JEB_MODEL setting (and loading that model in jeb serve or AINode). Until then the default judge is Jeb, the general model.

The hosted judge is opt in. Name "judge": "jev" in a policy entry, or pass jevJudge(), and set a TypeSafe key in TYPESAFE_API_KEY:

export TYPESAFE_API_KEY="..."
node scripts/eval.mjs --judge jev --cases cases/completion-check-tuned.json

The key is read from there and nowhere else, and never reaches the ledger. See USAGE.md to call ask(), add a decision, or write a policy entry.

What JDE does not do

It does not generate text. It does not decide what code can compute. It does not retry a judgment: one call, a timeout, a fallback. It does not hold your credentials. It does not call a paid service unless you name one.

Where it came from

JDE was extracted from Titanium Bot, a hosted AI agent product where these decisions run against real customer work every day: whether a research answer is supported by what was actually read, whether a request is about local stores or national prices, whether a background worker delivered or only reported.

The shape turned out to be general, so we took it out and gave it a licence.