AI agents you can actually govern.

Agent frameworks gave models hands. Nothing gave them a leash — or a way to catch their own mistakes.

Today, a model's judgment flows straight into effect: a prompt becomes a send, a delete, a purchase. When something goes wrong there is no durable record of what was decided, by whom, under which permission — and no way to replay what happened.

A smarter model does not close this gap. The gap has two independent sources, and either one is enough to make ungoverned agency unsafe.

Entropy. Natural language has null spaces — regions where several readings are valid and the model cannot tell it is standing in one. Chain enough interpretation→decision→action steps and a null-space hit stops being a risk and becomes a near-certainty.

Fallibility. Reasoning is conjectural. A model can be fluent, confident, and wrong even when the premises are clear — and it cannot detect its own bad inference from inside it. Every deduction, generalization, and plan is a guess that survives until something external checks it.

Together these are the Entropy and Fallibility Conjectures: even an arbitrarily smart model, reading and reasoning in language, will eventually drift into a null space or a bad inference; if the resulting actions are irreversible, the consequences are unbounded. That is why Narada separates intelligence (what the model thinks) from authority (what the system may do), and makes every step between the two durable, explicit, and inspectable.

Every boundary blocks a named collapse

Narada pipelines every operation through one fixed sequence. Each arrow is a governed crossing that prevents a specific failure mode.

This is the core of Narada: not a policy document, but a running system in which a model's output cannot skip a step — there is no code path around the pipeline.

One control plane for everything an agent touches

A deterministic kernel plus a governed tool fabric, designed for operators who run agentic work for real.

You stay the operator. Narada does the remembering.

Narada is designed around ops repos — private repositories that hold your operations, their knowledge, and their local configuration. You shape what runs; the daemon executes it under governance.

Authority is a dial, not a leap of faith

Every operation runs under an explicit posture you choose — and can tighten or loosen at any time.

The default for a new mailbox operation is draft-only. You grant more authority only when the evidence has earned it.

Proof, not promises

The governance claims on this page are checkable in the repo — the kernel invariants behind them are enforced in code and covered by CI tests and lints, not written in a whitepaper and left as an exercise for the reader.

Running in the open

These public sites run as Narada sites — what they publish passes through governed admission and promotion, with evidence, before it reaches the web.

Three programs, no shared charter — yet all three mark evidence the same way: perimeters on marici, gaps as first-class data on cintamani, confidence classes on the space ledger. The convergence is structural, not instructed: each site publishes through the same governed tool surfaces, which make a claim without its evidence class a non-operation. We tested this rather than assert it: an unbriefed agent, handed nothing but the surfaces, discovered the epistemic gate on its own and produced perimeter-marked work — the gate refused anything else — while a control agent without the surfaces produced equally good research with no perimeter marks at all. Form-level emergence: observed. The honest caveat: the gates govern publication through surfaces, not writing itself — the wall has an escape hatch, and knowing that is part of the finding.Full record.