The Machine That Refused to Guess: Why I Built Eternal Evals

“I got tired of machines that sound certain and are quietly wrong. So I built one that computes first, and only then allows itself to speak.”

---

The Moment I Stopped Trusting AI Astrologers

I asked a popular AI astrology app a simple question about my own chart—something I already knew the answer to. It replied with beautiful, fluent, confident prose. It was also wrong. Not poetically wrong. Factually wrong. It placed a planet in a sign it was never in.

That is the moment this project began. Because the app did not know it was wrong. It had no chart. It had my birth date and a language model, and it did what language models do: it produced the most probable-sounding sentence. Fluency without foundation.

I have spent years writing on this site about Jyotish from a grounded perspective and about what AI actually is and is not. Those two threads collided in that wrong answer. I realized the problem was not astrology and it was not AI. The problem was that nobody had bothered to separate the two jobs that were being done badly at once.

Two Jobs, Collapsed Into One

Every “AI astrologer” tries to be two things simultaneously: the calculator and the interpreter. It tries to work out where your planets are and tell you what they mean, in the same breath, from the same probabilistic guess.

Large Language Model (LLM) technology

A neural network trained to predict the next most-probable token given its context. It is astonishing at language and synthesis. But it does not compute—it estimates what a correct-sounding answer looks like. Ask it for a planetary degree and it will confidently invent one.

A birth chart is not a matter of opinion. The Moon was at an exact sidereal longitude at the moment you were born, or it was not. That is arithmetic performed against centuries of astronomical observation. It is the kind of thing a computer should get right, every single time, to the arc-second—and the kind of thing a language model should never be allowed to guess.

A planet's position is not a vibe. It is a computation. The tragedy of AI astrology is that it lets a machine that cannot count pretend that it can.

So the design principle wrote itself: the engine computes; the AI only interprets what the engine computed. The language model is never permitted to invent a number. It reads real ones.

The Market I Was Building Against

Once I saw the problem clearly, I saw it everywhere. The Vedic astrology market splits into three kinds of product, and all three fail the same test.

The content farms. Enormous websites that rank for every astrology query on earth. They publish articles about dashas and doshas and navamsa charts. What they almost never do is compute yours. You read a generic essay about Sade Sati; you do not learn whether you are in it. The authority is real; the personalisation is theatre.

The sun-sign apps. Slick, daily, addictive—and reading a single data point, your Sun sign, which in the Vedic system is one of the least informative things about you. They feel personal. They are horoscopes for one-twelfth of humanity at a time.

The bare chatbots. The newest failure. Point ChatGPT at your birth details and it will happily “read your chart”—hallucinating placements, inventing dashas, fabricating dates, all in flawless prose. It is the most convincing wrong answer yet built, precisely because it is the most fluent.

Notice what none of them do: compute a real, complete chart and then reason honestly from it. That was the empty chair in the room. I decided to sit in it.

What I Actually Built

Eternal Evals is, underneath, an unglamorous thing: a computation engine. It runs on the Swiss Ephemeris—the same high-precision astronomical data serious observatories and software rely on—and it computes a full sidereal chart deterministically. Not the Sun sign. The whole apparatus: planets and houses, sixteen divisional charts, the Vimshottari dasha timeline and twenty-one conditional dasha systems, shadbala strength, ashtakavarga, the yogas, the nakshatra structure.

Three commitments make it different from everything above:

It is deterministic

The same birth details produce a byte-for-byte identical chart, every time. There is no randomness, no “creativity” in the numbers. This sounds obvious until you realise almost nothing in consumer AI astrology can promise it.

The computation is separated from the interpretation

This is the heart of it, and it is the same insight I keep circling in my writing on pattern replication versus genuine understanding. The engine never interprets. The AI never computes. When the AI makes a claim about your Saturn, that claim is anchored to a number the engine produced—an exact degree, a real dignity, a dated dasha period. The model reads; it does not invent.

It is honest about what it does not know

The engine refuses to fake precision. Where a “dosha” has no single classical definition, it shows you the actual geometry and says so, instead of issuing a scary verdict to keep you engaged. Honesty, it turns out, is a feature you have to build—because the entire incentive gradient of this market points the other way.

Most astrology products are optimised to feel true. I wanted one that was optimised to be true, and then trusted that being true would feel like something better.

Why the Name Is “Evals”

Eval—evaluation—is the discipline in AI of actually measuring whether a system is right, instead of being charmed by how it sounds. It is the unglamorous work that separates engineering from theatre.

Naming the project Eternal Evals was a promise to myself. This is Jyotish—jyoti, light, the science of cosmic timing that our tradition treated as a rigorous computational science, not a party trick. It deserves to be evaluated, not vibed. Every number auditable. Every claim traceable to a computation. The eternal questions, held to an honest standard.

The Journey, and the Hard Part

The hard part was never the astronomy. Swiss Ephemeris does the heavy lifting; the classical formulae are documented across centuries. The hard part was discipline—refusing, over and over, to let the language model do the one thing language models are best at, which is to smoothly paper over what they do not know.

It would have been easier to let the AI “help” with the numbers. It always sounds better. It is always a little wrong. Holding the line—engine computes, AI only interprets—meant saying no to fluency in favour of truth, again and again, in a hundred small design decisions.

There is something very Indian about that discipline, and I do not think that is a coincidence. Our knowledge systems were obsessed with pramana—valid means of knowing—long before “hallucination” was a word we used for machines. The question “how do you know that this is true?” is old. AI has simply made it urgent again.

Where It Is Going

The most interesting thing that happened recently is that AI itself learned to use the engine honestly. Eternal Evals now exposes its computation to any AI as a tool. You can ask it inside ChatGPT. You can add it to Claude and Cursor as a connector. Developers can call it directly through an open API and an MCP server.

That inverts the original problem in a way I find quietly beautiful. Instead of an AI guessing at astrology, the AI now reaches for a tool that computes it—and interprets only what comes back. The model does what it is genuinely good at, language and synthesis, standing on a foundation it is not allowed to fabricate.

What I Want You to Take From This

If you remember one thing, let it be the distinction, because it will matter far beyond your birth chart: an AI that sounds certain is not the same as an AI that is right. The gap between those two is where hallucination lives, and it is being papered over, at scale, in products people are starting to trust with real decisions.

The fix is not smarter language models. It is humbler architecture—systems that know the difference between what they can compute and what they can only guess, and that have the discipline to keep the two apart. I make that argument in full in Why LLMs Hallucinate — and the Case for Computation-First AI. I built one, for a tradition I love, as a proof that it can be done. You can try the engine at eternalevals.com, and judge it the way it asks to be judged: by whether it is true.

---