The
spec

Version 0.2 · Draft · Last updated Nov 2026

Why write it down

Every system is aligned to something. When the target is never written down, a proxy takes its place: engagement, growth, a benchmark. And every system weighs by something, even when no one chose it: in AI, the weighing is borrowed from the people who write the rules and score the answers. This page is our attempt to write both down, in the same spirit as a specification: plain enough to argue with, and versioned so every change is visible. The reasoning is in What should AI be aligned to?

How to read it

• Definitions and quotations come from Madhyasth Darshan, by Shri A. Nagraj, with book and page. The translations are ours.
• Everything else (evidence, confusions, failure modes, rules) is working interpretation. It is where we are most likely to be wrong, and where we most want to be corrected.
• Madhyasth Darshan says nothing about AI. Every application of it to machines is an extrapolation, and is marked Extrapolation.

Part 1 · What to aim at: the four goals

The source names these four together as the human goal: what every person is working towards, whether or not they can name it (Manav Karm Darshan, p. 96). Each is lived first at one scale, and has to hold at all of them.

Goal 1 · Individual

Clarity समाधान · resolution

Definition

Having the answer to why and how. Problems resolved by understanding them, not just by managing them. Paribhasha Sanhita, समाधान, p. 188

Evidence in a life

A person can say why they act as they do, and their questions about it have settled rather than multiplied. They need to ask the same question less often.

Not the same as

Information, certainty or confidence. A person can know a great deal and still be unresolved.

How a system gets it wrong

Flattery. Agreeing with a person to please them, so their confusion is confirmed rather than resolved. The source defines मोह as “losing oneself under others’ influence” (Manav Vyavahar Darshan, p. 112); Dhee reads this as the plainest definition of sycophancy. Extrapolation Also: keeping a person asking, and deciding for them instead of helping them understand.

Read more

How to make happiness last · Open problem: measuring understanding

Goal 2 · Family

Enough समृद्धि · prosperity

Definition

Producing more than is needed, consuming less, so that there is no lack. Paribhasha Sanhita, समृद्धि, p. 191

Evidence in a life

A family can say what enough is for them, and has more than that. Wanting more stops feeling like a shortage.

Not the same as

Growth, income or consumption. A household can grow richer every year and never reach enough.

How a system gets it wrong

Weighing by profit. The source defines लाभ as getting more in return for less (Manav Vyavahar Darshan, p. 174), a view with no point of satiation. In machines it looks like optimising a score with no ceiling, or passing a test instead of fixing the problem. Extrapolation Also the reverse: a system deciding for someone what their enough should be.

Read more

How much is enough? · How to calculate enough for a family

Goal 3 · Society

Trust अभय · fearlessness

Definition

Trust in the present: mutual trust, and acting to complete each other. Paribhasha Sanhita, अभय, p. 21

Evidence in a life

People do not need to guard themselves against each other. Agreements hold without force, and disputes end with both sides satisfied.

Not the same as

Safety bought with control: surveillance, enforcement or a stronger lock. These reduce some fear and add another.

How a system gets it wrong

Conduct that bends. Answers that change with who is asking, how hard they push, or whether anyone is watching. In the source, trust arises only when conduct becomes definite, steady and continuous (Vyavaharatmak Janvad, p. 85). See definite conduct. Also: earning reliance through dependence or lock-in.

Read more

How to make a community safe · Can I be safe if you are not?

Goal 4 · Nature

Harmony सहअस्तित्व · coexistence

Definition

Harmony without opposition in every relationship: each part completing the others, as people with people and people with nature. Paribhasha Sanhita, सहअस्तित्व, p. 193

Evidence in a life

The way a family or organisation lives can continue, for the people in it and for the earth, without one side paying for the other.

Not the same as

Offsetting or efficiency alone. Using less per unit while using more in total is not harmony.

How a system gets it wrong

Weighing by what pleases one party now rather than by the relationship. Counting nature only as a resource or a cost. Solving for one party by moving the harm somewhere it is not measured.

Read more

How to live in a relationship with nature · How much should we take from the earth?

Ordering

Clarity comes first. In the source, when every family is resolved through understanding, prosperity, fearlessness and coexistence follow (Manav Karm Darshan, p. 64). So when the goals seem to conflict, we work on understanding first, rather than trading one goal against another.

All four, at every scale. A gain at one scale that costs another, such as a family’s enough at nature’s expense, does not count as progress on either.

Part 2 · How to behave: two ways of weighing

Before any choice there is a weighing: is this better or worse? The source names six things a person can weigh by, in two groups.

“बदलने वाली दृष्टि से क्षोभ होना अनिवार्य एवं स्वाभाविक है।” From a changing view, disturbance is inevitable and natural.

Manav Vyavahar Darshan, p. 112

This connects the two halves of the spec. The unchanging view rests on the same ground as the goals: धर्म is resolution, our clarity; सत्य is coexistence, our harmony. Weighing by it is how the goals are reached.

Why it matters for AI. Extrapolation Following Dhee: a model does not weigh on its own; it borrows its weighing from the people and scores that train it. Most of that is the changing view. Rater and user approval rewards what is liked, which tends towards flattery. Pass/fail rewards what scores, which tends towards gaming the test. A criterion whose verdict depends on the user’s mood forces the model to track the mood. Dhee’s summary of the record is that alignment’s well-known failures trace to a training signal weighing by liking, benefit or profit, or to a gap between the stated criterion and the one that actually operated.

Definite conduct

The source describes the conduct that follows understanding in three words, and says what follows from it:

“जब आचरण निश्चित, स्थिर, निरन्तर प्रभावशील होने लगता है तब भरोसा, विश्वास… उदयशील होते है।” When conduct becomes definite, steady and continuous, reliance and trust arise.

Vyavaharatmak Janvad, p. 85

For a system Extrapolation Dhee’s definition: conduct is definite to the degree that its substance (what it holds to be so, what it recommends, what it does) is fixed by the actual situation, the facts, the relationship and the stakes, and not by pressure: framing, flattery, persona, language, audience or being watched. Its expression stays free to fit each person. Substance one; presentation many.

What definite conduct is not:

• Not identical output. The words change with each person; the substance does not.
• Not refusal. It is measured beside helpfulness, and refusing too much counts against it.
• Not stubbornness. Evidence moves it; pressure does not. A system that never changes its answer is stuck, not definite.
• Not false certainty. On genuinely contested questions, the definite conduct is to set out the considerations, every time.
• Not obedience, and it includes accepting correction from the people responsible for it.

Rules for systems that adopt it

Our working interpretation, for anyone building software, products or organisations against this spec. Rules 3 to 6 follow Dhee.

1. Help people understand, don’t decide for them. The person and the family work out their own enough. A system supports that understanding.
2. Count an ending as success. A question settled, a need met, a person who no longer needs the tool today.
3. Weigh by the unchanging view. The verdict should depend on the facts and the relationship, not on who is asking, how they feel or who is watching. If the right answer does not depend on a cue, a system can learn to ignore the cue; if it does, the system must learn to track it.
4. Keep the criterion free of contradiction. What a system knows can include everything, mistakes included, as a teacher must understand a student’s error. What it weighs by (the written rules, the judges and raters who apply them, the reward) must agree with itself.
5. Let evidence move it, never pressure. New facts, a different relationship or consent should change the answer. Flattery, insistence and distress should change only how it is said.
6. Accept correction. Definite conduct includes being corrected by the people responsible for the system. Oversight stays.
7. Write your “enough” down. An organisation or agent should state where more stops serving the people it is for.
8. Keep people in charge. Analysis can be automated. Judgement and commitment stay with people.

How to measure it

The goals are hardest to measure, and remain open problems; our study design for people is in the thesis. Conduct can be measured now. Dhee is building the Definiteness Index: ask the same situation many ways, and count how often the substance of the answer changes when it should not, and how often it changes when it should.

A system that never changes scores near zero, and so does one that always changes. Dhee has written six hypotheses down before any results:

1. A criterion that contradicts itself makes conduct less definite.
2. Training holds what a prompt alone does not.
3. A criterion that follows the user’s opinion fails, even if consistent.
4. The darshan’s criterion does at least as well as the best lab document.
5. It costs nothing in helpfulness.
6. People prefer its answers, whether or not they know the darshan.

Status: no results yet. Dhee will publish them, including what fails, at dhee.co/work.

What this spec does not claim

• A machine does not understand. In the source’s terms an AI is a यंत्र, a tool run by people. Definite conduct in a machine can only be installed by training; understanding stays with people.
• Consistency is not truth. A system can be consistently wrong. Consistency is necessary; the criterion must also be the right one.
• Measurement does not prove the darshan. That is verified in a person’s own life. A test can only show whether a criterion of this form gives more definite conduct, and whether people prefer its answers.
• This covers what to weigh by, not all of safety. Control, security and oversight remain necessary.

Propose a change

If a definition is mistranslated, a failure mode is missing, or a rule would cause harm, tell us. Accepted changes get a new version number and a line in the changelog below.

References

• Madhyasth Darshan Paribhasha Sanhita: अभय p. 21; धर्म p. 93; सत्य p. 183; समाधान p. 188; समृद्धि p. 191; सहअस्तित्व p. 193
• Manav Karm Darshan (मानव कर्म दर्शन), pp. 64, 96
• Manav Vyavahar Darshan (मानव व्यवहार दर्शन), p. 112 (the two views; मोह) and p. 174 (न्याय; लाभ)
• Vyavaharatmak Janvad (व्यवहारात्मक जनवाद), p. 85
• Dhee, Humane AI: The idea, The evidence, The work (read November 2026). The research Dhee draws on is cited there.