Playground · typed decisions
Ask a deal a question. Get a probability, not a paragraph.
Describe a situation, ask Jev typed questions, and get calibrated answers back in under a second. Then set the thresholds yourself. Jev reports the odds; your policy makes the call.
- 1Describe the situation
- 2Ask typed questions
- 3Run Jev
- 4Set your policy
Saved deals
Saves the editable situation, questions, policy, answers and readable CIM research context to your pass's private Cloudflare storage.
1Situation
0 charsStart from a deal-team template
Sample or redacted PDFs · up to 25 MB. Text is extracted in your browser; review it below before running Jev.
Screening evidence · verify against the PDF
Source excerpts, not confirmed metrics. Missing disclosures remain unknown. Only the editable situation below is sent to Jev.
What Jev reads (the state)
Jev reads text only and handles words better than numbers. Put thresholds in words ("more than 20% of revenue") and let code check the math.
2Questions
0 / 8CIM research questions
Each question starts an ephemeral Cloudflare Worker and keeps running in the background. Review the public search terms first.
Upload a CIM to import its readable source and research context automatically.
Uploaded CIM · all readable research context
Source excerpts are unconfirmed. Full readable text is retained here separately from the shorter Jev screening excerpt. Only the reviewed search terms below leave your browser for public research.
Run Jev answers the typed questions in one call. Research questions run separately in background Workers.
3Answers
4Your rules
applied instantly · no new AI callsJev says how likely each answer is. Your rules decide when it answers for you and when it checks with you first.
Agent · durable deal-flow triage
Hype, fact, or signal?
A durable agent works through your deal-flow feed. For each item, one Jev call answers five questions: relevance, duplicates, hype/fact/signal, thesis fit and red flags. No LLM writes anything, so a full sweep costs fractions of a cent. Close the tab and it keeps going. Open the link on your phone and you'll see the same run.
Pipeline
Start the agent to watch Jev triage the feed.
Your rules
board re-sorted · no new AI callsThe agent asks Jev about every item. These four rules turn Jev's answers into the board. Change them any time: the board re-sorts instantly, without asking Jev again.
Pi agent · Cloudflare Sandbox
How Pi worksAn agent harness in a Cloudflare Durable Object, one per session. Every step is saved there before you see it.Thinks withKimi K2.7 Code on Workers AI, through AI Gateway. The model is for the event; one setting swaps it.Reads the deal board, drops in new deals, and asks Jev for judgments.Does its maths in Python in a Cloudflare Sandbox: Linux, no internet.Jev scores every shell command before it runs.Read the Pi guide →Ask the desk. Pi does the work, Jev guards the shell.
A Pi Durable harness runs inside a Durable Object, so every turn, tool call and result is committed before you see it. It reads the board, drops in new items, and does its maths in Python. Its hands are a Linux sandbox with no internet, and Jev scores every shell command before it runs. It thinks with Kimi K2.7 Code, a model hosted on Workers AI for the event. How Pi works
1Ask Pi
Pi works on your latest Agent run. If there isn't one, it starts one first.
2Jev command gateJev command gateBefore any sandbox_bash command runs, Jev answers three questions about it in one call: how risky it is (safe, risky or destructive), whether it runs remote code, and whether it sends data out.Code blocks it at destructive ≥ 50%, remote code ≥ 70% or exfiltration ≥ 70%. A blocked command never reaches the sandbox, and every check is logged here.How the gate works →
0 checks- No commands yetEvery
sandbox_bashcall is scored here first: safe, risky or destructive; remote code; exfiltration.
3TranscriptTranscriptWhat Pi's Durable Object has saved, in order: your prompt, Pi's messages, each tool call with its result, and the gate's verdict on every shell command.It refreshes about once a second while Pi works, and comes back when you reload. The chips show your session, the model and the sandbox.Reading the transcript →
How it works
A System One model in front, your policy in charge.
Jev is a decision model. It doesn't generate text: you send a state and typed questions (Yes/No, Pick one, Rate), and it answers every question in parallel with calibrated probabilities, usually in well under a second. That makes it cheap enough to check every item, every minute.
The stack behind this page
#run=… link.policy.js is shared with the browser.jev-gateway.jev-latest → jev-1.13.0Calibrated probabilities. Billed on input tokens; output is free.Why it's called a Pi agent
It follows the Pi Durable pattern IndyDevDan laid out: a proactive agent is a durable agent plus a trigger, inside a sandbox. Clients are thin and detached.
- Durable agent: the DealAgent Durable Object. It keeps working when no one is watching.
- Trigger: a new item dropped into the feed. A cron sweep of real sources is the next step.
- Jev everywhere: gate, dedupe, classify and rank are all Jev questions in one call, so tokens stay thin.
- Sandbox: the Pi tab runs the Pi Durable harness in a Durable Object, and its hands are a Cloudflare Sandbox container with no internet. Jev scores every shell command before it runs.
The part most demos skip
- Probabilities aren't verdicts. Jev is calibrated, but the thresholds are your policy. That's why every slider re-sorts the board with zero model calls.
- Words over numbers. Jev reads text and handles words better than arithmetic. A real deal screen checks EBITDA math in code.
- Fictional data. Every company and item on this page is made up for the demo.
- Demo budget. Inputs are capped and requests are rate-limited per network. Nothing you type is stored by jev-gateway, which audits only counts and types.
The next step is wiring your real sources (inbox, broker teasers, CRM) into the agent and tuning the questions on your past deals. Talk to Jordan after the session.