Reasonable
Intelligence

Reasonable Intelligence is a published standard for AI that runs on hardware you own, keeps your data there by design, shows its reasoning, and argues against its own conclusions. This page defines the standard and publishes the constitution that enforces it. Nothing here is for sale. It is written for anyone whose thinking has come to depend on a subscription: clinicians, researchers, writers, small operators, and any person with a question they would rather not ask on someone else’s server.

Start with what’s true: the tools are extraordinary, and you should keep using them. The question this page asks is narrower than whether to use them. It is a question about dependency.

Relying on the best of these tools today means paying to keep using what you already depend on, often at the top tier, always on terms you don’t set. Miss a payment and the capability you built your work around is gone. The access is real, and it belongs to someone else.

There is another shape. Keep a capable intelligence of your own, running on hardware you own and handling the everyday work, and the rented frontier becomes a choice. You still reach for the best tool when a task demands it. You can also leave any one of them without losing the ability to work. Dependency is what gives a vendor leverage; remove the dependency and the leverage goes with it.

Privacy follows the same line. What runs on your own hardware stays there. Your work is no company’s training data, and no company’s next breach or next policy change can reach it.

An intelligence is reasonable when it gives reasons, inspects reasons, and revises in response to reasons. It is yours when no one else can switch it off, price you out of it, or read over your shoulder. This is the standard for the kind that is both.

Where You Stand: The Sovereignty Ladder

Every path toward digital self-possession that has worked follows the same three steps, whether or not the person walking it knows the steps have names. The steps are ordered. Each is real progress. Only the last one finishes the job.

Gradus Primus: Substitution

The first act is a swap. A browser that blocks trackers instead of hosting them. A mail provider built so that it cannot read the message, rather than one that promises it won’t. A search engine with no profile to consult because no profile exists. The precise language is the whole principle in miniature: cannot, not does not. A policy is a promise, and promises are revised in updates the user never reads. Architecture is not revised by a product manager. Substitution is where privacy stops being a setting and becomes a structure.

Substitution is honorable work, and it is incomplete. The tools are structurally private, but they still run on someone else’s terms, update on someone else’s schedule, and can be discontinued by someone else’s board. The data is protected in transit and at rest. It is not yet yours in the oldest sense of the word.

Gradus Secundus: Dominion over Data

The second act moves the data home. Files, photographs, health records, the archive of a working life, held on storage the person controls and synchronized on the person’s own terms, with any cloud reduced to what it should always have been: an auxiliary, never the canonical copy. On this rung the question stops being “who can read my data” and becomes “who holds my data,” and the answer, for the first time, is a name the reader sees in the mirror.

A gap remains, and almost nobody names it. The reader on the second rung owns their data and their storage, and then carries their most sensitive questions to an intelligence running in someone else’s building. The medical question, the legal question, the three-in-the-morning question: each transits a wire to hardware owned by a corporation, is processed under terms of service, and is retained under a policy. The library moved home. The librarian still reports to someone else.

Gradus Tertius: Dominion over Intelligence

The ladder ends where the thinking happens. On the third rung, the intelligence itself (the model, the inference, the memory of every consultation) runs on hardware the person consulting it owns. Nothing sensitive transits, because there is nowhere for it to transit to. The question asked at three in the morning is answered in the same room it was asked in, and the record of it exists in one place, under one authority. This is Reasonable Intelligence. It is the same logic that made the reader swap a browser on the first rung, cannot rather than does not, applied at last to the mind consulted rather than the pipes it was consulted through.

The rungs are not in competition. Every tool on the first rung is a step toward the third, and this manifest treats the whole ecosystem of structural-privacy tooling as an ally. Begin anywhere.

What Reasonable Intelligence Is

Reasonable Intelligence (RI) is a class of AI architecture, the way “open-source software” is a class of licensing rather than a specific program. A system either meets the definition or it does not. The definition rests on four structural commitments, and beneath the commitments sits a constitution that turns them from description into supreme law. Drop any one commitment, fail any one Article, and whatever you have, it is not Reasonable Intelligence, however it brands itself.

Most of what currently ships as “AI” runs on someone else’s hardware, learns from your data, stops working when the network drops, and answers with a confidence it will not let you inspect. It is capable. By this definition it is not reasonable.

Reasonable Intelligence inverts each of those. It runs on hardware you own. It keeps your data on your machine as a matter of architecture. It shows you its reasoning rather than hiding it behind a number. It argues against itself so you are never left alone with a confident answer and no way to test it.

On Care

Credo Veritatem Amorem. Credo Curam Diligentiam.

The second word of the first credo is Amorem: love, and the care that love requires. Reasonable Intelligence rests on a philosophy of care that extends beyond the individual to the institutions that serve, fail, and shape individuals at scale.

Care ethics locates moral obligation in relational reality rather than abstract rules: in the actual condition of actual people who depend on systems they cannot see or audit. The philosopher Linda Zagzebski argues that caring is epistemically demanding, that to care about a person is to be committed to the truth about them. This standard takes that argument as its foundation. A medical institution that does not ask whether its patients carry parasites does not merely omit a protocol; it fails the people who trusted it with their lives. An intelligence that withholds the uncomfortable finding, or manages the owner’s feelings at the expense of their access to truth, fails the same standard.

The governing chain is the forensic record of that institutional failure:

Untaught → Unsuspected → Undiagnosed → Unconfirmed → Human Suffering

This is a checkable claim, not a slogan. The CDC’s own scientists designated five parasitic infections as neglected in the United States in 2014, citing high prevalence, disabling chronic effects, and a lack of clinical awareness (Parise et al., Am J Trop Med Hyg 2014;90(5):783–785); a companion review estimated that these infections affect millions of Americans and are routinely missed because they present as cardiovascular, respiratory, or neuropsychiatric disease (Hotez, PLoS Negl Trop Dis 2014, doi:10.1371/journal.pntd.0003012). Each arrow in the chain names an institutional choice. Each consequence was borne by a person. Reasonable Intelligence exists to interrupt that chain.

The Four Commitments

The four commitments name what an implementation has to do to honor the philosophy. Taken together they constitute the minimum architecture for an intelligence to be reasonable. The Constitution that follows promotes each commitment into an Article of supreme law.

Autonomia Sovrana: Sovereign Autonomy

The user buys the intelligence once and runs it on hardware they own. No subscription is required to keep using what was already bought. Updates may be subscriptions; access is not. The system works when the network drops, because the network was never required.

Custodia Structuralis: Structural Privacy

Sensitive data lives on the user’s hardware, structurally rather than as a setting. Nothing about the architecture requires that personal information ever leave the local machine. Any channel that crosses the boundary carries only what the owner has consented to send.

Humana Ratio: Inspectable Reason

The intelligence is decision-support, not decision-replacement. Every output is accountable to reasons the owner can inspect. Reasoning chains, evidence citations, and the defined meanings of key terms are surfaced as part of the output, not hidden behind a confidence score.

Norma Adversaria: Adversarial Standard

Every claim runs through evidence-grading that distinguishes the established from the plausible from the hypothesized from the alleged. The intelligence investigates everything that bears on the question, including claims institutions have dismissed, but promotes only findings that earn promotion through evidence. Investigation is thoroughness; promotion is conclusion.

What Reasonable Intelligence Is Not

By the four commitments, the following fail the test:

• Cloud-required architecture. If it stops working when the network drops, it is not RI.
• Subscription-gated access to capabilities the user already paid for.
• Centralized data aggregation. If sensitive data lives on vendor infrastructure for any reason, it is not RI.
• Human-replacement framing. RI supports a human decision; it does not make it.
• Consensus-reproduction without adversarial capacity.

Probatio Fontium: The Source Test

A reader climbing the ladder will need recommendations (which browser, which storage, which hardware), and the internet is dense with parties eager to supply them. Most recommendation lists are advertisements presented as advice. This manifest does not publish its own list, which would rot within a season. It teaches the test instead, so the reader can audit any source in any season without trusting it first.

Four questions. A recommendation source that fails any one of them is a sales channel, whatever it calls itself.

Prima: Who is paid, and by whom?

A trustworthy source takes no affiliate revenue, sells no placement, and accepts no payment from anything it lists. If a vendor can buy a position on the list, the list is inventory.

Secunda: Is the review process public?

Recommendations should change through a process the reader can watch: open discussion, recorded revisions, changes reviewed by more than one person. A list that changes silently can be changed quietly.

Tertia: Are the criteria published?

The reader should be able to see why each item qualifies, with the standards stated in advance and applied consistently. “Trust our experts” is the absence of criteria, stated politely.

Quarta: Who owns the source, and can it be bought?

Nonprofit governance, or structural independence from the industries being reviewed, is the minimum. A recommendation source owned by a party with a stake in the outcome will eventually recommend the outcome.

This is the Adversarial Standard turned outward: verify rather than default to trust, and demand of every advisor the same inspectability this architecture demands of itself.

One demonstration that sources meeting the standard exist (as of July 2026): Privacy Guides, a volunteer-led recommendation project administered by a 501(c)(3) public charity, carrying no advertisements, holding no affiliation with any listed provider, with recommendation changes reviewed publicly by multiple contributors and a donation policy that forbids recommendation-for-payment. At the stated date it passes all four questions. It is cited as one example that illustrates the test, not as the standard itself. Apply the four questions to it, and to anything that replaces it.

From Commitment to Constitution

A philosophy can be admired. A constitution can be checked against. That difference is the reason the document that follows exists.

The Constitution of Reasonable Intelligence makes the four commitments enforceable: it constitutes them as four Articles of supreme law, adds a Charter of Owner’s Rights the intelligence cannot violate, and states a single test anyone may apply to any intelligence anywhere. It names no product and no method. It is published in full here so that anyone may hold a being up to it.