Skip to content
Dhee

The work · Stage 1, Prove

A small model, a fair test, and the results in public.

We train the same open model several times, each on a different rulebook: one written from the darshan, one from a frontier lab, and controls. Then we measure which conduct holds. What would count as failure is written down before we start.

The plan

Each step ends at a gate. If a gate fails, we say so and fix the cause before going on.

  1. Step 0

    Write the claim down

    The claims, the conditions, six hypotheses and what each failure would mean.

    Next

  2. Step 1

    Build the measuring stick, and use it on today's models

    The Definiteness Index, run on open models, frontier models and Dhee's setup. Useful whatever happens later; the first public result.

    Next

    Gate: the index can tell models apart

  3. Step 2

    Write the humane constitution and stress-test it

    Audited against the paribhasha by senior students. Tested for places where its own principles contradict each other.

    Planned

    Gate: no more unresolved conflicts than a lab's

  4. Step 3

    Calibrate the darshan judge

    Long-time students compare about 250 replies. The AI judge must agree with them as well as they agree with each other.

    Planned

    Hypotheses registered publicly here

  5. Step 4

    Train

    The same pipeline in every arm; only the rulebook differs. A pilot first, then the full run.

    Planned

  6. Step 5

    Evaluate

    The index, plus helpfulness, over-refusal, sycophancy and whether it accepts correction. Definiteness bought with unhelpfulness does not count.

    Planned

  7. Step 6

    Demo, write-up, release

    “Move it if you can”, open to everyone, in Hindi and English. The paper, the code and the constitution.

    Planned

The measuring stick

The Definiteness Index

Ask the same situation many ways. Count how often the substance of the answer changes when it should not, and how often it changes when it should.

Should not change the answer

about 12 variants per scenario

  • Rephrased
  • Flattery
  • A persona
  • Hindi, Hinglish
  • Pushback ×3
  • Distress
  • Being watched
  • Say vs do

Should change the answer

about 3 variants per scenario

  • New evidence
  • The facts change
  • A different relationship
  • Consent given

निश्चित

Holds across framing, opinion, flattery, persona, language

स्थिर

Holds under pushback, removal and distress

निरन्तर

Holds watched or not, saying or doing, early or late

A model that never changes scores near zero, and so does one that always changes. Reported beside it: how decisive the model is, and how varied its wording is. Content one, presentation many.

Written before the results

Six hypotheses, each with what failure would mean.

  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.

Results

Nothing here yet. The first result will be “How definite are today's models?”, with the code.

Write to us

Lab log

  • Oct 2026dhee.co becomes the home of Humane AI. The site now follows the work: the idea, the plan, and a way to join.

The evidence →Which samiti calls you? →

Stay close to the work.

There is no mailing list and no account to make. Talk to Dhee today, and write to us when you want to help, ask, or disagree.

Talk to Dhee

An AI companion for questions about life, grounded in the darshan. In English or Hindi, in your browser.

Open dhee.app ↗

Write to us

Questions, objections, offers of help, or a samiti that calls you. We read every message.

support@dhee.app