KHNUM FORGE
KHNUM FORGE
What our agents do, what they never do, and where each limit stops.
Our agents draft, review and answer, and a person decides. Each says it is an AI, works only inside the scope you gave it, and can be stopped by your administrators. How we protect your data is on the Security page.
When one of our agents changes a document, the change arrives as a marked proposal. Every change after the starting text stays a proposal until a person on your team decides it. Your people accept, reject or counter each one, and the record shows who decided, and when.
Our agents never accept a change for you, and never send a document to a counterparty.
Proposals. The first version of a thread is the starting text. The system refuses to build the next version while any change is undecided. Each accept, reject or counter is recorded with the verified email address of the person who made it and the time. Our audit log separately records, with the server's time, who applied the decisions.
Dates. Your calendar's dates come from your own documents and from fixed deadline rules. A date Tyr reads from an executed paper is saved with the passage that states it, and reaches the calendar only when a person confirms it. Athena, our project manager, books the approved project team's calls on a schedule.
In every chat, the agent's first reply says it is an AI and not a person. Every email an agent sends carries a line saying so, and so does the working copy of a paper Tyr drafts. Our code adds these lines, so a client's own settings cannot remove them.
Our agents never open a chat, or send an email, without saying they are an AI.
Each agent is set up for one client, with the practice areas, industries and channels you chose when it was deployed. It can carry out only a fixed list of actions. Changing its setup takes a deliberate act by an authorized person, and everything it produces stays in your own workspace.
Our agents never give themselves new practice areas, industries or channels.
Fixed at deployment. Only a Khnum Forge operator can deploy an agent or change its setup. What an agent can do comes from a list in code, and no output from a model can add a practice area, industry, channel or agent.
Jurisdictions. These come from your firm's own profile. When a deal your firm brought names a state whose law we already hold, a fixed rule adds that state and records it.
Ask Tyr. It answers from your firm's own jurisdictions and industries. When none are recorded, it holds to federal law and your firm's records.
Anyone signed in to your firm can press Stop while a reply is in flight, and nothing from that turn is stored, drafted, queued or mailed. Administrators can pause one agreement, matter or project, or the whole firm. A paused agent answers with one line naming who paused it and when. Nothing it was asked to do proceeds: no queued work starts and no mail goes out. Mail to a paused agent is held too, with one exception: our memoir product’s family inbox still honors a request to stop contact at once.
You do not have to ask us to stop an agent.
Stop this turn. After its model call and before it stores or acts, the agent discards its output if a stop was pressed. A stop lands only while the turn runs; a finished turn keeps what it saved. Ask Tyr acts on nothing, so it has no Stop button; a firm pause silences it.
Pause. A pause records who paused it (by verified email), the server's time and any note; each stop, pause and resume is an audit entry. On an agreement the portal asks you to confirm first. A paused chat gets one line and no model call. Queued jobs and mail are held, including our project manager's reminders and calls.
The data itself. We can lock your data by disabling your firm's own encryption key, which makes it unreadable within about a minute and can be reversed.
When you have not given Tyr a term of a deal, the draft leaves a blank line for it and lists every blank. When a review cannot confirm something, it says so. When nothing on file answers a question about your own agreements, Tyr says that and passes it to your legal team, and a general legal question is answered as general information.
You see every blank a draft leaves, and every value we filled in ourselves.
Blank until filled. The facts an agreement needs are declared up front. Before it drafts, Tyr stops and asks about any fact that decides whether a section of the agreement exists. For every other term you have not supplied, it leaves a blank line and asks you, rather than making up a number. Anything Tyr fills in itself, such as a standard default or something read from your own description, is shown to you.
Reviews name their gaps. A finding Tyr cannot tie to the paper's own wording is marked down to a note. A requirement it could not check is flagged as unchecked, and a draft with open holes is not labeled ready.
Questions. A general legal question is labeled as general information, not your firm's position. Where we can, answers cite the document they came from. We do not claim every chat answer is verified against your records.
Our writing and illustration agents screen every request before anything is produced, whether it comes from you or from another agent. The screen stops a request for a named real person's likeness without a release, for someone else's trademark or logo, or for a copyrighted character.
We never clone a voice, and our writing agents refuse a real person's likeness without a release.
A firm-wide policy. A responsible-AI policy holds the content work to bounds Khnum Forge sets for every client. It sits above client policy, only Khnum Forge can set it, and it is read live, so a change applies on the next request.
Both sides of the prompt. It governs the person who briefs the team and the agents that prompt one another. A person who is stopped is told why and asked to revise. An agent must try again with a compliant prompt, and after a set number of tries the request is refused.
Generic, cultural, national and public-domain references stay open (a Chinese dragon, the Statue of Liberty).
Images. You cannot upload your own images into your published content. Your logo is the only exception; every other image is generated by a licensed image model or fetched from a public source.
Voice. We never clone, imitate or train on anyone's voice, living or deceased, even if the person or their family asks. Our assistant speaks only in a licensed synthetic voice that belongs to no real person.
Said plainly: this screen covers our content work and reads requests, not the finished image. A model makes each decision, so it is a screen and not a guarantee.
Every change made through your portal, every approval or rejection a person makes, and every sign-in we refuse is written to a tamper-evident, write-once record. Each entry links to the one before it and carries the signed-in person's verified identity. For every agreement Tyr drafts or reviews, you can also open each step and read what the system was asked and what it answered.
We never keep a record that hides a change: a changed or missing entry breaks the chain, and a check shows where.
The proof. Entries go to an append-only log. Each carries a SHA-256 fingerprint of the entry before it, so changing or removing one breaks the chain. The identity comes from the signed-in person's verified token, never from what a caller says.
What it holds. What happened and who did it, including every stop, pause, resume and key change. It does not hold the full text of every conversation, and automated background work is logged under the system's own name.
The readable record. For each agreement Tyr drafts or reviews, each drafting and review step opens to the prompt and the reply, clause by clause. The chain is a separate record, kept by us.
Outside its scope an agent refuses where our code enforces it, and Tyr passes questions about your own case to your legal team. On your matters agents act only on a person's decision, and each agent answers in its own name, never as a person.
Our agents never claim an authority they do not have.
Outside its scope, an agent refuses where our code enforces it. You cannot use an agent for legal advice on your specific case: Tyr hands those questions to your legal team, and each agent opens only the work areas your firm set up. Our writing agents check each request before drafting and refuse out-of-scope requests and attempts to talk them out of their rules. A model makes that check, so it is a screen and not a guarantee.
On your matters, no action without a person's decision. Nothing goes to a counterparty, and no counter-draft is built, until a person on your side has decided every change. A contract is marked signed only on a person's instruction. Published content waits for approval unless you turn on automatic posting. Scheduled project work, such as reminders and calls, runs on your firm's schedule without a person approving each one.
In its own name, never as a person. Our agents never send mail as a person. Every message comes from the agent under its own name.
No claimed authority. Our agents identify themselves by their own names, not as people or attorneys. Their drafts carry blank signature lines for you to sign, and Tyr tells you when an answer is general information and not legal advice.