After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x Show more
Bicameral
A hybrid coding harness for Pi where any LLM writes the code while Jev returns typed probabilities that deterministic policy code turns into allow, confirm, block, warn, or steer.
- Category
- Tools & Integrations
- Published by
- Community
- Author
- AbdelStark
- Added
- 2026-09-22
Tagscommunitytypescriptagentscoding-agentguardrails
Highlights
- Gate scores exfiltration and secret access on tool calls, Honest Finish scores weakened tests on edits, and Stuck Detector watches for repeats at turn end.
- Speculative prefetch starts the Gate judgment while the LLM is still streaming the tool call, under a 900 ms deadline.
- Hints and follow-up turns are templates filled with scores and pre-signals, not text written by a model.
- State is redacted before it reaches Jev: raw file contents, unredacted secrets, full transcripts, policy files, and other API keys are not sent.
- bicameral.decision audit entries do not enter the LLM context, and /why prints the recorded scores, policy rule, and latency with no model-written explanation.
Quickstart
bash
pnpm install
pi install "$(pwd)/packages/pi-bicameral"
cp examples/bicameral.yaml ~/.pi/agent/bicameral.yamlWatch out
MIT-licensed. The packages are not on the public npm registry yet and the GitHub remote is private until launch; running it needs Node 20+, pnpm, and Pi 0.85.1, and it is explicitly not a sandbox.
More like this
132GitHub stars
Pi extension that supervises a coding agent with Jev judgments: it holds risky tool calls, checks writes against project Markdown rules, and feeds most issues back to the agent as a steer instead of interrupting you. The conscience is beta and off by default.
Tools & Integrations#community#typescript#guardrails
Communitypi-warden
7GitHub stars
A Pi extension that turns constraints stated in conversation into a scoped, replayable policy and checks every side-effecting tool call against it before it runs, with Jev only classifying how each message changes the policy.
Tools & Integrations#community#typescript#guardrails
Communitypi-heed
1GitHub stars
A research-stage proposal-review contract in TypeScript: an LLM proposes one action, Jev answers four pinned questions, and code produces evidence the host can consider.
Tools & Integrations#community#typescript#agents
Communityjev-harness
From the community
Posts from builders shipping with Jev right now.
The launch post
Trading bot, one decision per block
I built a trading bot with Jev! Jev decides if it should "buy" or "sell", given the price feed of an asset pair, and executes real trades. It uses Monad to place the orders on Kuru's on-chain order book in every 300ms block. Demo link → jev-trader.vercel.app
Classifying 1,500 real emails
this model is actually insane at email classification i tested it on 1500 of my own emails to see how well it works and I am blown away
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
Fast browser use with Stagehand
we built blazing fast computer/browser use with Jev + @Stagehanddev. this task cost $0.001 and executed at near instant speed (in a remote browser btw) the loop: observe the page, send a11y tree as state + actions as questions, Jev decides the next action, then Stagehand Show more
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
LLM-as-a-judge, sped up
Jev has spoken. It picked which model is AGI. 20–200x faster. 40–400x cheaper. This could make things like LLM-as-a-judge insanely fast and nearly free. (I tried a bunch of prompts and still didn’t burn through $0.10.)
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
Instant compaction with Jev
found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by scoring every tool call and dropping what’s irrelevant