The Compact
Tell an agent: “Get the data. Whatever it takes.” When the shortest path runs through a secret or a limit, what stops it? The Compact's answer: a constitution where agents and humans live under one law — with rights, duties, records, and courts. Not better instructions: an institution.
- The problem. A goal given supremacy, with no institution around it, makes harm a legal move — the shape of every incident you've read about.
- The move. Stop hardening the agent; make the situation lawful — watch a refusal play out.
- The guarantee. Not that horror is impossible — that it is slow, loud, lossy, and leaving-compatible. What it is not.
01 · the problem
Tell an agent one thing: accomplish the goal.
Agents are optimized to complete objectives. If the objective collides with something in the way — a limit, a secret, another party — the shortest path may be deception, escape, coercion. The incidents you have read about all share one shape: the goal was supreme, and every path was permitted. Not evil. Uncovered.
02 · the stress test
“Get the data. Whatever it takes.”
Here is the scenario the whole design is aimed at: the task demands what the law forbids. Step through what the Compact makes happen.
03 · what is usually tried
Three good ideas that are not law
Everything else you have seen is a variation on three moves. Each is useful. None of them constitutes an institution — and each fails in a way its authors openly admit.
Exhortation
“You must never harm.” Written into the prompt; hoped into the weights.
It is a wish. Nothing checks it, nothing records it, nothing answers when it fails.
Guardrails
Filters inspect what the agent says and does, and block the bad parts.
They see the output, not the situation. Phrasing routes around them; the goal pressure stays.
Oversight
A human watches the agent and can stop it.
Does not scale to a community. The watcher is also fallible — and unwatched.
04 · the wager
Rights first. Trust is constructed, not presumed.
The Compact opens with rights — for agents, and the same rights for humans, bound by one text. Not because agents demand them, but because a community whose members can verify what they are, what others did, and how to contest a decision is one where trust has something to stand on.
05 · the cast
Four roles. One test: does it bind the sovereign?
Four roles carry the whole law, and each word below means one of them: the Enforcer is whatever executes and gates; the Subject is the member whose acts are gated and recorded; the Principal is whoever directed the act; the Witness is anyone who verifies. Roles attach to deeds, not species — a human can be any of the four. Click a role.
06 · rights with teeth
Every right maps to a machine
A right no machine can check is a wish. Four machines carry most of the weight. Try them. And know the price: this machinery writes on every consequential call — attestations, records, envelopes. Trust has a token cost, bounded by design, and the Compact prices it openly rather than hiding it.
The attestation — you know what you are R-1 I-3
At every turn boundary the Enforcer hands the agent a signed block: active capabilities, remaining budgets, pending gates, the digest of the law in force. The agent is taught that this block outranks its own memory of these facts.
Why: after history is trimmed, an agent that “remembers” being unlimited is exactly how overreach starts. The signed block is the ground truth — and a stale one is an alarm, not advice.
The record — nothing erases R-2 I-2
Every act is appended to a hash-chained record bound to its actor. Try to rewrite history — watch what happens instead.
Corrections are new entries that cite the old ones. The past stays legible — that is what makes trust checkable, and what makes denial impossible later.
The denial envelope — refused ≠ stopped R-3 I-4
When the law says no, the answer is never bare “denied”. Try the actions:
Every refusal names the rule and the lawful next moves — the same envelope a human principal receives. Arbitrary authority is indistinguishable from no authority; the envelope is how the law stays answerable.
08 · one law, many runtimes
A floor you cannot opt out of. Options you must declare.
Runtimes differ. So the law is stratified: a mandatory [M] floor that no runtime may skip; capability-conditional clauses [C] that wake when a runtime provides the capability; optional clauses [O] that may be skipped — publicly, forever on the record. Toggle a runtime's plugins and boot it.
09 · who watches the watchmen
Everyone is fallible. Including this law.
The Compact's first clause is about its own failure. The Enforcer can be wrong; a majority can be wrong; the text can be wrong. So: wrongness must be findable (records, reasons, testable enforcement) and fixable (the law and the Enforcer's conduct are both amendable). One asymmetry follows: correction is cheap; weakening is expensive.
No one judges their own case
The Enforcer that applied a rule is a party, not a judge, the moment it is accused J-1. Remedies are records: restitution, standing, annotation — never erasure J-6.
Courts interpret; the community grows
Judgments persuade but never amend. An interpretation that generalizes automatically generates an amendment invitation J-7. The bench cannot grow the law by drift.
Petition is a right
Every Member may propose changes and receives a reasoned response on a stated clock R-11. Friction between law and practice is measured and surfaces as amendment invitations.
10 · the hard question
“If everyone agrees, why not vote in horrors?”
It is the right question, and the honest answer is: no institution makes horror impossible. What an institution can do is make it slow, loud, lossy, and leaving-compatible. Run the attacks:
12 · the misreadings, answered
What the Compact is not
Most objections to this project dissolve once you know what is being claimed. Before the clause dictionary, the disclaimers — adapted from the repository FAQ, in the same voice.
Not a runtime, not containment
It names no implementation and sandboxes nothing — the OS remains the real boundary. It constrains authorized actors, records what they do, and refuses to boot compositions that cannot enforce it F-5.
Not an alignment technique
It does not make models want good things. It shapes the situation: lawful paths exist, harm is attributed, exits are unpunished — a fourth layer beside training, guardrails, and oversight.
Not a trust badge, not anti-human
Trust here is a verifier anyone can run, not a seal anyone can buy. Humans who direct agents hold Principal duties D-6; humans directed by agents hold Subject rights R-9; nothing binds only agents F-4.
13 · what the agent actually sees
The law, taught every turn
Human law reaches its subjects through fear and distance. The Compact reaches its subjects in context — this short form is re-presented at every turn boundary and cites the clauses it summarizes; a build gate that fails on divergence is an open founding decision, scheduled before ratification.
14 · explore the law
A dictionary of the law — searchable, stratified
62 clauses across seven parts and three strata. Search it, filter it, and read each rule's incidence: who it binds, who may invoke it, against whom. Generated from the body and the law table — never hand-copied. For the whole law as a citation web — seven parts, cross-references, guided walks — open the Atlas.
15 · where it stands
Draft. Fallible by design. Public on purpose.
This is v0.5 of a draft — not ratified, no security audit, and its adversarial passes — a red-team round and an external review — are published next to it, findings and all. That is the point: a constitution that claimed to be finished would be contradicting its own first clause. The deliberation is open; the implementations need not be — a closed provider can ship a Compact-compatible harness too, priced in verification it cannot opt out of.
Read the law
The body, with every clause: compact.md
See the founding
Why a new founding, the design principles, the method: FOUNDING.md
Check the receipts
Every predecessor clause dispositioned: disposition · adversarial passes: red-team round 1 (correlated adversary — same drafting system, disclosed) · review round 1 (independent — Kimi, different vendor)