Contest pass rewritten
The second pass now runs against a held-out source set, which cut agreement bias on repeat questions by a third. Traces from before this date will diff noisily against new runs.
Azr0 is an agent runtime with no inherited priors. Every run begins at the origin, collects its own evidence, and reports what it actually found — not what it was told to expect.
Most systems arrive pre-loaded: cached beliefs, stale rankings, borrowed conclusions. They are fast because they skip the part that matters.
Azr0 resets that. An agent wakes with an empty state, a budget, and a question. It gathers, weighs, discards, and hands back a trace you can audit line by line. If it cannot show the path, it does not ship the answer.
No memory carried between runs unless you pin it. Same question, same conditions, reproducible trace.
Each retrieval, score, and rejection is written to the run log before a conclusion is allowed to form.
Competing paths are scored against evidence, then collapsed into a single output with its confidence attached.
The agent console runs in your browser. Ask it for status, scan the mesh, pull the ledger, or just type help and follow the prompts.
An agent is only as good as the evidence it refuses. This is the short version of how we build, and what we will not do to make a demo look faster.
Autonomy without accountability is just a faster way to be wrong. Azr0 treats the trace as the product and the answer as a by-product. You get the reasoning path, the sources it kept, and the ones it threw away.
Collect, contest, collapse. Three passes. The second one exists to attack the first — an agent that never argues with itself is an agent repeating its input back to you in a nicer font.
Empty state. The agent receives a question, a budget, and a deadline. Nothing else.
Sources are pulled in parallel and stamped with time, origin, and a trust weight.
A second pass tries to break the first. Claims that survive keep their weight; the rest are logged as rejected.
Surviving paths are ranked and reduced to one answer, carrying its own confidence score.
The trace is hashed. Re-run the same input and you can diff the two runs line by line.
No. The console is one surface. The runtime is the scheduler, the trace store, and the contest pass — the parts that decide whether an answer is allowed to exist.
Carried memory quietly becomes carried bias. Pinning context is opt-in and always visible in the trace, so you can tell which parts of an answer came from evidence and which came from history.
It returns the low confidence and the reason for it. A hedged answer with a visible gap is more useful than a clean sentence built on one weak source.
Yes — point it at your endpoints and the same collect / contest / collapse cycle applies. Request access on the Access page.
A live sandbox of the runtime. Type a command or tap one below. Everything here runs locally in your browser.
Each run declares its cost before it starts. Overruns stop the run rather than quietly degrading the answer.
No shared scratchpad between tasks. Cross-talk is the fastest way to launder a bad source into a good answer.
Every run exports as a signed JSON trace you can replay, diff, or hand to someone who does not trust you.
One contract, fixed supply, no mint function. The numbers below are the whole story.
0x34e0000000000000000000000000000000284ba3
Network: Robinhood Chain · Placeholder address for this build
It meters runs. Holding is not the product — spending is. Compute budget, priority in the queue, and access to private source connectors are all denominated in it.
No promises about price live on this page. If a project has to explain why its chart is a feature, it has stopped building.
Four phases. Dates move; the order does not.
Short notes, written when something actually changed.
The second pass now runs against a held-out source set, which cut agreement bias on repeat questions by a third. Traces from before this date will diff noisily against new runs.
Runs that exceed their declared budget now stop and return a partial trace instead of finishing on a cheaper path. Loud failure beats quiet degradation.
The public sandbox went live. Same command surface as the internal runtime, minus private connectors.
Sealed and versioned. Anything that reads a v1 trace will keep reading it.
Private connectors and higher run budgets are gated. Tell us what you want to point the agent at and we will reply within a few days.
That address is not in the index. The origin is still where you left it.