YOUR AI-POWERED IT DEPARTMENT · UNDER AUDIT
NAKKITO runs IT for organizations that have no IT department — and works alongside the one you already have. It proposes, a human approves, it executes once — and an append-only ledger proves what happened. Marketing pages make claims; this one files them as ledger entries, each with how it was verified. Nothing more is claimed than the evidence column can carry.
no sales call · the product is live behind this page
✕ halted by you · animations and polling stopped · nothing on this page moves until you say so — that is the whole idea
EACH ENTRY CARRIES ITS EVIDENCE · NONE CARRIES MORE
A customer issues a scoped, expiring, revocable grant. Delegation can only narrow what the grant allows — never widen it — and a grant with nothing listed permits nothing.
The requester can never be the approver. An approval checks its own expiry at the moment of decision, and is consumed on use, so nothing can run twice on one yes.
Audit events, provenance records and task history are append-only because the database raises an error on UPDATE or DELETE. A bug, an admin, or NAKKITO itself cannot quietly rewrite what happened.
The kill switch and the highest-risk tier always drop. The tier below can only ever produce “ask a human” — no configuration makes it run unattended.
Answers are grounded in your records and carry their sources and age. Visitor text, ticket text and retrieved documents are data, never instruction. When the model is unreachable, it says so and answers from the records alone.
The embeddable widget answers only from material published on purpose. There is nothing behind it to leak, which is a stronger property than a filter in front of a secret.
THE ANTI-FEATURES · AS LOAD-BEARING AS THE FEATURES
There is one execution path, and it refuses anything unapproved. Not a setting. The path.
A persuasive answer is still just an answer. Authority comes from grants and approvals, never from text.
Verification is three-valued: verified, failed, or honestly unknown. “Unknown” is an answer this product is allowed to give — and gives.
Completing a course teaches; it confers no capability. Being taught, being permitted and having done it are three separate records that no foreign key quietly merges.
No staged screenshots, no invented testimonials, no fake activity feed. The spec forbids them (§9). The stream above is the product's real audit vocabulary, labelled as exactly that.
FOUR STATIONS · NO SHORTCUTS BETWEEN THEM
Anything with consequences starts as a proposal. The plan is frozen and hashed at request time — an approved-then-swapped plan cannot run.
A different human than the requester. The approval checks its own expiry at the moment of decision, not when it was filed.
One path, which refuses anything unapproved. The approval is consumed — it can never authorize a second run. Authority is re-checked before each step.
The outcome is recorded as verified, failed, or honestly unknown. Execution cannot be fabricated after the fact.