Chakravyuha: a browser ring-maze where every move is a Laya decision

# Chakravyuha — Laya guides Abhimanyu

A browser labyrinth where **every move is a decision by Laya**. Abhimanyu starts on the
outermost ring of a *chakravyuha* — the concentric battle formation of the Mahabharata —
and has to reach the **target at the centre** past the warriors standing in the rings.

The maze is serialised to text, sent to the Laya model (running locally) as a set of typed
questions, and the move Laya picks is the move that gets played. Where Laya is unsure — or
answers with something that cannot be played — the walk **marks the step red, plays the
correct move instead, and carries on**. So the path is either all green or a mix of green
and red, and that colour, not the outcome, is what reports the model.



*The screenshots here are taken against a mock upstream so that rendering them spends no
model time — the app itself has no stub mode.*

## What this is

**A fun experiment.** Viswa and Sarathy (his AI) built this to explore the Laya language
model by giving it a toy it could not bluff its way through. Two questions drove it:

1. **Can a model play a game purely through typed answers?** Not free text, not tool
   calls: a state goes in, a finite set of typed choices comes out, and the game obeys.
2. **Can its confidence score be trusted?** If a high-confidence move is always the right
   move, then the score is worth gating on. If it is not, the score is decoration — and
   that is worth knowing before anyone builds on it.

The maze is the excuse. The real output is the measurement: **green where Laya's own move
was played, red where it was overruled**, and a history page that asks whether its
confidence predicts its accuracy.

It is a toy, and it is deliberately unforgiving: a wrong move is visible on the board, and
every claim on the history pag