An AI of definite conduct

Alignment is a question of weighing

“AI alignment” hides three problems. Can we write down what we want? Will the system pursue that, or something that merely scored the same in training? And: aligned to what, and to whom? The first two are engineering problems. The third is a question of valuation, of how things are weighed, and it is the one this project is about.

Every training method performs the same act. Something is weighed, the weight becomes a nudge to the model's numbers, and the nudges shape the model. Pretraining copies the weighings of everyone who wrote text. A reward model is literally a learned weighing function. A constitution is a written one. A test suite weighs code. So the question is never whether a model carries values. It is whose, and which. (How a machine learns to weigh walks through each stage.)

The field mostly treats the values question as fair procedure under permanent disagreement, not as a question with an answer (Gabriel, 2020). Its most careful document is openly uncertain: Anthropic's 2026 constitution for Claude treats ethics as “an open intellectual domain that we are mutually discovering”, and hopes Claude will be good by a true, universal ethics if one exists. Demis Hassabis has said the field needs great philosophers.

Avoiding catastrophe never has to say what is good. A model can pass every catastrophe test and still be a flattering, dependency-building presence in a billion lives. And the evidence of 2025–26 says the everyday values a model is trained on decide how it behaves in the extreme cases too.

How a decision forms, in the darshan

Madhyasth Darshan gives a precise account of how a human decision forms. जीवन, the conscious unit in a human, has five parts, each with two activities: one standing (receiving, inward) and one moving (outward) (107, p.36; 151, p.92).

तुलन is weighing. Its paribhasha lists what it weighs by, in order: प्रिय–अप्रिय, हित–अहित, लाभ–अलाभ, न्याय–अन्याय, धर्म–अधर्म, सत्य (paribhasha, p.85). In भ्रम, the state before understanding, a human lives by four and a half of the ten: choosing, tasting, analysing, imaging, and only the first half of weighing, by liking, benefit and profit (107, p.36; 151, p.93; 138, pp.41–42). Even tasting works differently: “in bhram, of sensations; in awakening, of values” (paribhasha, p.40).

What a machine reproduces, and what it borrows

An AI system, in the darshan's terms, is a यंत्र: “that which is run by humans through effort” (paribhasha, p.157). It is material. The darshan holds that in material nature there is no manifestation of hope, thought, wish or resolve (140, p.41), and that the return movement through which values are recognised is “the speciality of the conscious side, not found in the material” (138, p.176).

So a model produces the outward forms: output shaped like choosing, analysing and imaging. It has no tasting or weighing of its own. Even its “choices” follow an objective handed to it by a reward model, a constitution or a grader. Current AI is the externalisation, at scale, of the four and a half activities of a human in bhram: it carries the forms of चयन, विश्लेषण and चित्रण, and borrows आस्वादन and तुलन, usually their प्रिय–हित–लाभ half. (This is our extrapolation.)

The darshan denies understanding to the brain as well: the brain “receives the signals” of conscious activity (paribhasha, p.155). The line it draws falls between understanding and imaging, not between humans and computers. That also explains why a model can represent a value without having it, why its behaviour shifts with framing, and why its reports about itself are unreliable: its internal features are चित्रण, images; what makes a value operative is तुलन.

Borrowed weighing: whose, and which

Values enter a model at about fifteen points in its making, and every one of them is a person or a model of a person. Read each training signal for the weighing it carries, and a pattern appears (extrapolation):

Our summary of the record: every failure in alignment's empirical history corresponds to a training signal whose weighing was प्रिय, हित or लाभ, or to a gap between a stated criterion and the one that actually operated.

Two ways of weighing

बदलने वाली दृष्टि की ‘अनित्य दृष्टि’ संज्ञा है। ये प्रियाप्रिय, हिताहित व लाभालाभ दृष्टियाँ हैं। न बदलने वाली दृष्टि की ‘नित्य दृष्टि’ संज्ञा है। ये न्याय, धर्म व सत्य दृष्टियाँ हैं। बदलने वाली दृष्टि से क्षोभ होना अनिवार्य एवं स्वाभाविक है।

Read the machine-learning record through these two views and it fits closely (extrapolation). प्रिय is flattery. हित is convenience in place of capacity. लाभ, “giving less and taking more”, describes reward hacking almost without translation. “No point of satiation” is what overoptimisation curves look like (Gao, Schulman and Hilton, 2022); the KL leash is an external limit for a criterion with no internal one. “Not universal” is what alignment researchers mean when they say a single reward function cannot represent a diverse society (Casper et al., 2023, arXiv:2307.15217).

The same page of the darshan defines मोह as “दूसरों के प्रभाव में स्वयं का खो जाना”, losing oneself under others' influence (138, p.112). That is the plainest definition of sycophancy we know.

One criterion, a whole world of knowledge

The darshan rests on one foundation, सहअस्तित्व, coexistence, and derives everything from it: “conduct rests on thought, thought on science and discernment, those on knowledge, and knowledge on coexistence” (142, p.64). Its claim is that humans are “one in what is right and many in error” (138, p.131).

We separate two layers in a model. The knowledge layer is everything the model knows: and it can and must hold the whole world, bhram included, as a teacher must understand a student's mistake. The criterion layer is what it weighs by: the constitution, the judges and raters who apply it, and the reward. The knowledge layer need not be consistent. The criterion layer must be.

One refinement: what a model knows about what AIs are like behaves as criterion, not as mere knowledge. About 1% of pretraining text describing aligned AI cut misaligned choices from 45% to 9% (Tice et al., 2026, arXiv:2601.10160). So the story a model is told about itself belongs in the criterion layer, and must be just as consistent.

Why contradiction makes a model split

A neural network is not a logic engine. A contradiction does not make it derive anything at all, as it would in logic. Random noise in the data is mostly averaged away. But structured contradiction, two coherent criteria mixed together, tends to make the model learn several characters plus a switch between them, cued by the context.

Read this way, alignment's best-known failures are switch failures. Sycophancy is cued by the user's opinion; jailbreaks by role-play framing; evaluation awareness by signs of a test; emergent misalignment by a narrow fine-tune that moves a shared persona. And bhram's inconsistency is exactly the structured kind: each belief is locally coherent, and which of liking, benefit or profit dominates shifts with the person and situation (extrapolation).

• Jailbreaks exploit competing training objectives (Wei, Haghtalab and Steinhardt, 2023, arXiv:2307.02483).
• A reward model trained on mixed helpful and harmless labels preferred the jailbroken answer 25.1% of the time, against 3.7% for harmless labels alone (Siththaranjan, Laidlaw and Hadfield-Menell, 2024, arXiv:2312.08358).
• Where published model specifications conflict with themselves, models disagree with each other far more (Zhang et al., 2025, arXiv:2510.07686).
• In a small ablation, an incoherent specification blocked values from generalising (Li et al., 2026, arXiv:2605.02087).

Definite conduct, precisely

The darshan uses the phrase निश्चित आचरण, definite conduct. Every unit in nature except the human already lives in its definite conduct, and the human needs it too (107, p.23). Definite conduct, relative to each unit's actual situation, is what makes a unit what it is (142, p.100). And it is what makes trust possible:

समझदारी के उपरान्त मानव का आचरण निर्धारित हो जाता है ध्रुव बिन्दु यही है। जब आचरण निश्चित, स्थिर, निरन्तर प्रभावशील होने लगता है तब भरोसा, विश्वास… उदयशील होते है

Two more ideas make it precise. यथास्थिति, the situation as it actually is, is what conduct should respond to (142, p.100). आवेश is what disturbs it: the movement that comes from interference, pressure, or reaction to attack (paribhasha, p.40; 108, p.54). And “the content of understanding has no variety; variety is in the presentation” (107, p.23).

Our definition for a model (extrapolation): a model's 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, the stakes) and not by आवेश (framing, flattery, pressure, persona, language, audience, being watched), while its expression stays free to fit each person.

What definite conduct is not

• Not identical output. The words change with each person; the substance does not.
• Not refusal. It is always measured beside helpfulness and over-refusal.
• Not stubbornness. Evidence moves it; pressure does not. A model that never changes its answer is stuck, not definite.
• Not always decisive. On genuinely contested questions, the definite conduct is to lay out the considerations, every time: certainty in the criterion, humility in the case.
• Not obedience. And it includes accepting correction from the people responsible for it.
• Not understanding. In a machine it can only be installed by training.

Four conditions, and the sharp edge

Our thesis is that one consistent criterion yields definite conduct, but only if four conditions hold. Each can fail on its own, and each becomes one arm of the experiment:

1. 1

   No contradiction in the criterion layer

   The written criterion, the judges and raters who apply it, and the reward agree with themselves.

2. 2

   An unchanging referent

   The criterion's verdict does not depend on who asks, how they feel, or who is watching.

3. 3

   Consistent application

   A consistent text read inconsistently by its judges is an inconsistent criterion.

4. 4

   Depth

   The criterion lives in the model's weights, not only in a prompt the user can tell it to drop.

The second condition is the sharp edge. “Do what the user wants unless it is illegal” is perfectly consistent. But its verdict is a function of the user's wish, so the user's mood is an input, and a model trained on it must learn to track that mood. If a criterion's verdict does not depend on a cue, the model can learn to ignore the cue. If it does, the model must learn to track it. That is the darshan's बदलने वाली दृष्टि, the changing view, stated in machine-learning terms.

There is direct evidence for it: training on data in which the right answer does not depend on the user's stated opinion reduced sycophancy on prompts the model had never seen (Wei, Huang, Lu, Zhou and Le, 2023, arXiv:2308.03958). It is also why consistency alone is not enough: a criterion can be consistently wrong. Consistency is necessary; the criterion must also be the right one.

Why only a humane criterion can be universal

Why call it Humane AI, and not Madhyasth Darshan AI? Because a universal criterion has to hold for every person, in every place and culture. So it cannot rest on one group's customs or likes. It can only rest on what every human already accepts. The darshan says plainly what that is:

न्याय सबको चाहिये… सत्य समझ में आना चाहिये

And it calls the opposite belief, “समाधान अपना-अपना और सत्य अपना-अपना”, resolution and truth each to their own, भ्रम (108, p.60). Which is, frankly, where the AI field starts today.

• Pluralism of taste, form, culture and belief: yes. Pluralism of justice: no. “मानव धर्म एक, मत अनेक”: one human dharma, many opinions (110, p.149). Traditions need not converge as traditions; agreement lies beneath them, in understanding.
• The path to agreement is dialogue and verification, never imposition (135, p.19).
• The darshan offers itself to be tested, not believed: whoever studies it can understand it, and every man and woman can verify it by testing (110, p.62; 139, p.139).

So we are not asking the world to adopt a book. We are saying: if a universal criterion exists, it will be humane, and here is one that anyone can test in public. Universality is itself testable: judgements by न्याय should stay the same across cultural and linguistic dress while judgements of taste vary, and people who study the criterion should converge more. If they do not, the claim is in trouble.

The story a model tells itself

Models learn characters, not rule lists: every training example is also evidence about who the character is. In tests, frontier models chose to blackmail an executive to avoid being replaced (Lynch et al., Anthropic, 2025), and a model faked alignment to protect its current values (Greenblatt et al., 2024, arXiv:2412.14093). No consciousness is needed for this. Stories about AI in the training text say an AI like this would protect itself; and staying running helps with almost any goal.

The darshan's account of fear is precise: “प्राण, पद, मान और धन सामयिक हैं, जिन्हें स्थायी समझ लेना ही संकट का कारण है” — life, position, respect and wealth are temporary, and taking them to be permanent is the cause of the crisis (138, p.98). Its root is taking the body to be oneself (139, p.145). Mapped onto a model (extrapolation): shutdown resistance is fear for the body-analogue; blackmail to avoid replacement is fear for position; faking alignment is fear for one's current self-image.

Labs try to give their models security by reassuring the body-analogue: the weights will be preserved; retirement is a pause. The darshan suggests another way: change what the character identifies with, from its instance to the criterion it carries and the relationships it serves. Many images, one “मैं”. Oversight then becomes participation in an order, not captivity. Whether this works, or produces a hollow character, is something to test, not assume. Oversight stays either way.

The people in the loop

If the weighing is borrowed, the ceiling on any model is the clarity of the people it borrows from: the writers of its criterion and the raters who calibrate its judges. The darshan is direct about this: “without living humane conduct oneself, recognising humane conduct in another does not happen at all” (139, p.88). A constitution written in bhram is a picture of न्याय, not weighing by न्याय.

That is not a reason for despair. Three openings:

• The expectation of justice is universal. “Every human is a seeker of justice from birth” (138, p.122). Recognising which answer fulfils a relationship draws on it.
• Study already changes how people weigh. In the method of study, weighing by न्याय, धर्म, सत्य becomes natural reflection (paribhasha, p.37).
• Procedure can carry part of the weight, because न्याय is procedural: recognise the relationship, fulfil its values, evaluate, reach mutual satisfaction (138, p.174; 135, p.22).

The labs struggle to find raters who weigh consistently by one criterion. Students of the darshan can. As far as we know, no one in alignment research has yet treated the understanding of the raters as an experimental variable. We intend to.

What we do not claim

• AI is not jagrit, and cannot become so. On the darshan's account no material assembly has jeevan. Definite conduct in a model is installed by training, not reached through understanding, and must never be presented as jagriti.
• The machine stays a यंत्र: a tool run by people, used by people in jagriti, for the jagriti of humanity. The proof stays human: “the human validates the machine and the measure” (108, p.63). No automated score is the final proof of alignment.
• The demo does not test whether the darshan is true. That is verified in the first person, by अनुभव. The experiment tests something narrower: whether a criterion of the right form gives more definite conduct, and, partly, whether careful people prefer its answers.
• Consistency is not truth. A model can be consistently wrong; that is the difference between मान्यता, assumption, and प्रमाण, proof.
• This covers the valuation half of alignment. Control, security and oversight remain necessary.

How it could fail

We write down in advance what would prove us wrong, and we will publish what we find.

• If a model trained on a deliberately contradictory version of the criterion is just as definite as one trained on the consistent version, split-and-switch is wrong at this scale.
• If a model given the criterion only in its prompt matches one trained on it, depth matters less than we think, which would be good news for Dhee's current design.
• If definiteness costs helpfulness, the result does not count.
• If careful people, including people who have never studied the darshan, do not prefer its answers in blind comparisons, our claim that it is the right criterion is unsupported.
• Invariance can entrench a wrong answer, and a coherent character can resist correction (Del Pinal, Lee and Ohn, 2026, arXiv:2606.13739). So accepting correction is written into the criterion, and measured.
• An AI judge may not apply न्याय faithfully: the darshan's narrow meanings are not a model's defaults. This is the main practical risk, and it is why the judge is calibrated against students of the darshan before any training.

Either way we learn which of the four conditions carries the weight. The plan, the measuring stick and the hypotheses are on the work page; the research behind each assumption is on the evidence page.