Decisions
An admission, a permit, a credit limit, a shortlist: what the agent decided, on what, with which model and version, and whether a person confirmed, corrected or rejected it.
When an agent decides, pays or carries out a process for you, someone will later ask what it did, on what, and who approved it. HardState keeps that answer in your institution's own chain, sealed in Bitcoin. Nobody can change or delete it afterwards without it showing, including you and us.
An admission, a permit, a credit limit, a shortlist: what the agent decided, on what, with which model and version, and whether a person confirmed, corrected or rejected it.
Before it pays, the agent records what it will pay, to whom, how much and why, and the person who authorizes it signs. Months later, an auditor can see who approved each payment.
Internal work an agent carries out step by step: what it looked at, which tools it called, what each step returned and where it failed, in words a manager can read.
An agent signs with a key limited to its own records, so it can't sign anything else. A person signs every review.
The system, model, version and outcome of each decision. Its input and output enter only as salted fingerprints.
Every step of a run (model calls, tools, lookups) folded into one record. Each step can still be proven on its own.
A reviewer confirms, overrides or rejects the decision with their own passkey, and the time is recorded.
At every interval the system states how many decisions it recorded. A missing beat or a count that doesn't match shows a gap.
Plain language first. The prompts, tokens and raw answers are there for whoever needs them, one click below.
To rebuild a decision later, for your own audit, a supervisor or a court, keep its evidence package: every step's content with the salts that open it, each step's path to the run, and the signed records. You store it; HardState never sees it. Because the provider's whole answer is committed, its reasoning, tokens and cost can't be edited afterwards either.
Keep your tracing tools. They serve your engineers, and their data lives in a store its operator can change. HardState serves someone who trusts neither you nor us.
For high-risk systems: education, employment, credit and insurance, essential public services, critical infrastructure, justice and biometrics.
Not legal advice. Whether a system is high-risk, and what it must log, is a legal assessment of your use case. HardState gives you a record anyone can verify; it doesn't decide what the record must contain.
Any language: one HTTP call and a signature. The SDK is @hardstate/ai, under Apache-2.0.
import { HardStateAI, ed25519Signer } from "@hardstate/ai";
const ai = new HardStateAI({
issuer: process.env.HARDSTATE_ISSUER,
signer: await ed25519Signer(process.env.HARDSTATE_AGENT_KEY),
system: {
name: "Permit pre-review",
model: "your-model",
version: "2026-11",
},
});
const d = await ai.decision({
// your reference, never a name
case: "PRM-2026-0412",
outcome: "approve",
// committed with a salt, never published
input: application,
output: modelAnswer,
});It proves integrity, time and order: what was recorded, by which key, when, and that nothing in the sequence was changed or removed. It doesn't prove that what was recorded is true, that the model reasoned well or that a review was careful. That part is still your institution's job, and now anyone can hold you to it.