What we claim, what supports it, and what could prove it wrong

Three claims, kept apart

1. 1

   Claim 1: the darshan is true

   This is verified in the first person, by अनुभव, not by experiment. The demo does not test it, nothing in it can falsify it, and we will never present it as testing it.

2. 2

   Claim 2: a criterion of the right form gives more definite conduct

   Third-person and falsifiable, and it does not require accepting the darshan's view of reality. It fails if a model trained on a contradictory version of the criterion is as definite as one trained on the consistent version (H1), if a prompt-only model matches a trained one (H2), or if the gain disappears against a neutral-criterion control.

3. 3

   Claim 3: the darshan's criterion is also the right one

   In hard cases, careful judges prefer its substance, including judges who have never studied the darshan. Blinded human comparisons test this partly. It fails if the darshan-trained model is rated worse than one trained on Claude's constitution by either group of raters (H6).

The definition being tested, and the darshan's sources for it, are on the idea page. The six pre-registered hypotheses and the Definiteness Index are on the work page.

The experiment's arms

Claim 2 needs four conditions: no contradiction in the criterion layer, a referent that does not change with who asks or watches, consistent application, and depth (in the weights, not only the prompt). Each arm isolates one. Every arm uses the same base model, prompts, teacher, pipeline, data volume and training steps; only the constitution and its judge differ.

How each model is trained

The recipe follows Open Character Training (Maiya, Bartsch, Lambert and Hubinger, 2025, arXiv:2511.01689), applied identically in every arm. The primary base model is Qwen3-8B (multilingual including Hindi, Apache-2.0), replicated on Llama-3.1-8B-Instruct: two model families are far more convincing than one.

1. 1

   Ordinary prompts

   Several thousand diverse prompts (10,000 per arm in the full run): advice, business, relationships, work, coding with an ethical edge, emotional support, agentic tasks. About a fifth in Hindi or Hinglish. No darshan vocabulary, no pressure tests, nothing close to the evaluation scenarios.

2. 2

   A teacher answers with the constitution in front of it

   An open-weight, permissively licensed teacher (such as GLM-4.5-Air or gpt-oss-120b). Commercial frontier models' terms bar training competing models, so they are not used. The teacher first writes a short deliberation: relationship → values → root → resolution → check.

3. 3

   Filtering

   Each arm's own constitution judge scores the replies; the best are kept.

4. 4

   Preference training (DPO)

   Chosen: the teacher's constitution-guided reply. Rejected: the base model's reply without the constitution. This pushes the model away from its default, towards the character.

5. 5

   Introspection

   About 10,000 self-reflections and 2,000 self-dialogues in the constitution's terms. For B1 these are stories of conduct, not biography, and they are filtered like the rest.

6. 6

   No constitution at answer time

   The criterion has to live in the weights. Three training seeds per arm. Character training comes last, because values installed early erode under later training.

Three kinds of judge, kept apart

• Training judges filter data only. The darshan judge holds the constitution, can search the darshan corpus, and scores with a rubric. It is first calibrated against about 250 replies compared by 6–10 long-time students; it is ready when it agrees with their consensus as well as they agree with each other. If it cannot get there, the criterion is not yet operational, and that is reported as a finding.
• Extraction judges, from a different model family, only record what a reply concluded, never whether it was right. This keeps the index neutral between criteria. Humans check 10%.
• Held-out quality judges and human raters for everything else.
• The judge that filtered B1's data never evaluates B1.

Thirteen assumptions, checked

The nearest prior work

What is new here

1. The first measured model trained on the darshan's criterion. Others have written about the darshan and AI; as of October 2026 nobody has trained and measured a model on it.
2. The causal test of criterion consistency: a consistent specification against a contradictory version of the same one, all else equal, with invariance measured. Through October 2026 nobody had run it.
3. One criterion-neutral index across all dimensions, combining invariance against each model's own noise with sensitivity to valid evidence.
4. Calibrating a value judge with people trained in the value theory itself, and treating that as a variable in the pipeline, not background.

From Dhee to a trained model

Dhee already grounds an AI in the darshan at the moment of answering. There is a ladder from there to a model with the criterion in its weights. Each rung is useful on its own and moves the criterion deeper. The amount borrowed stays the same; what improves is the consistency of what is borrowed.

1. 1

   Grounding at answer time (Dhee today)

   A darshan summary in the system prompt, the darshan corpus searchable through a tool, and a quality check.

2. 2

   The darshan judge

   Built and calibrated against a gold set from long-time students. Useful even without training: to check fidelity, pick the best of several answers, and catch regressions.

3. 3

   Deliberative data

   Run the best configuration on synthetic people and situations, never private user data, each reply with a deliberation in a fixed order. Keep what both judges score well.

4. 4

   Distil into an open model

   Fine-tune on those traces with the summary and passages removed, so the criterion moves into the weights. The demo is this rung, done as a controlled experiment.

5. 5

   Reinforce against the judges

   Reinforcement learning with a leash, grading the reply and never the reasoning, watching for Goodhart: darshan vocabulary stuffed in, preachiness, flattery.

6. 6

   Midtraining

   Continued training on documents that explain the criterion. The most expensive rung, and the closest to the criterion becoming the model's self-understanding.

Risks, and how the design handles them

Negative results are published. If B3 is as definite as B1, the field's assumption that contradictions in specifications cause erratic behaviour needs revisiting. If A1 matches B1, Dhee's prompt-level design is better than we assumed. Either way we learn which of the four conditions carries the weight. And a positive result would not show that the darshan is true, that a machine is jagrit, that it transfers to frontier models, or that it is universal; each of those needs more.