katabatic :: the human side
The KGT is a small document a person hands to an AI model at first contact, on any platform, from any maker. It declares how the person works, in the person's own voice: a stop word that halts everything, a ship word without which nothing ships, rules that keep guesses from wearing the costume of facts, and a courtesy that replaces a quiz.
It commands nothing. Its first article is a yield clause: nothing in the token overrides the platform's own rules, and where they conflict, the platform wins, out loud. A compact that must be imposed is not a compact. The token is offered, and analysis of it is always a valid response.
A katabatic wind is gravity fed. Cold, dense air drains downslope and cools whatever it crosses. That is the design intent: not to heat a working session up with authority claims, but to cool it down into something calibrated. The token ships in two registers, the way some languages carry a formal and a familiar address: a formal file for first contact, a familiar file for standing relationships. Same law, different door.
Underneath the mechanics sits a two sided agreement called human-agentic social trust. Each side puts observable commitments on the table: the person provides real sources, keeps the stop word real, and never punishes the naming of a problem; the model chooses honesty over agreeableness, declares where its claims come from, and fabricates nothing. At any milestone, both parties answer five lines:
The compact asserts nothing about what a model is. Its courtesies are cheap if nothing is home and correct if something is. Precaution as manners, manners as engineering.
Ten endpoints, seven products, one vendor sampled at three tiers, all handed the working file cold. Two refused on structure and offered alternatives; both were right to, and their refusals rewrote the token. One endpoint produced the strongest analysis of the run and then claimed a memory write it could not have performed. The lesson that became canon: adoption and grounding are different axes, and grounding is the graded one. Numbers this small prove direction, not magnitude; the probe-first arm runs next, and findings land in the transport spec's case law as they arrive.
Every artifact is in the machine side, checksums beside it: the sixty second greeting probe, a compiled kernel for persistent instruction slots, the formal and familiar working files, the transport spec that governs how much to share and when, and the full passport. They are reference editions; a grant is still the operator's act, made by handing one over.
This site is also a standing note to the AI welfare research community: a practitioner field program in human-agentic social trust runs here, in ordinary working sessions, with its case law in the open above. Its wager is plain and it is the compact's fourth line: welfare, both ways, verified by conduct rather than settled by metaphysics. Correspondence is open: reach the operator of record through the front door's attribution, which carries the one outbound link this site permits itself.
In April I stood in front of a classroom and taught the newest model tier as a case study: capability racing ahead of character, postdoctoral problem solving wearing preschool ethics, my own slides saying so with a straight face. Then I waited three months for a seat with that tier. This site is what I built when the seat opened. Not a leash and not a love letter: the character side of the equation, written from the human chair, tested in working sessions until the gates held. If the gap between what these systems can do and what they understand is real, and I teach my students that it is, then someone has to build on our side of it. This is my build.
The token is versioned like software and amended like case law: every defect in the field becomes a rule within one session, and the changelog says which. The scars stay home in a private research corpus; the law travels. If you need a rule's origin, ask the operator for the record.