Beats by Itihasa · a storytelling research lab

Beats: Narratives & Models

Storytelling in Residence to Frontier Research & Startups
§1 Essays

The Beat Room is where I write it down, one essay at a time.

The Beat Room is my newsletter on storytelling, AI and venture. Newest first. It also goes out on LinkedIn.

§2 Work

This is the work.

Micro documentary

Promise of Pramaana poster

Promise of Pramaana

One Proof at a Time. about 5 minutes, 2026.

A micro documentary on the meeting of ancient mathematical wisdom and modern computational proof systems, following Pramaana Labs as they turn centuries of knowledge into verifiable frameworks.

Written and directed by Siddharth Ram
Co-director Mohammed
Edited by Daksh
Designed by Anto
Camera Balu
Produced by Pramaana Labs

Videos

Events and rooms

  • Verification Summit, inaugural edition

    San Francisco, June 2026

    The first Verification Summit, anchored by Vinod Khosla. I built the storytelling around it, including the film made for the room, AI Stops Saying Sorry.

    Watch AI Stops Saying Sorry

  • Follow The Frontier

    Bangalore, Fridays, November to December 2026

    BoldCap's invite-only series that brings researchers and builders into one room to follow what is actually moving at the frontier.

  • OpenAI Demo Party x BoldCap

    Bangalore, 2026

    A demo party for builders rather than a conference. The film from the night is in the videos above.

Websites and brands

  • Pramaana Labs

    Website design and brand storytelling for an AI verification lab. The brand had to make formal proof feel like trust, not like mathematics.

  • BoldCap

    Brand identity, website and narrative strategy for the fund I work at. The thesis on the site is written the way the partners talk.

Shows

  • Attention Is All You Need

    YouTube, in production

    A solo essay-film show about the gap between frontier AI and the culture it lands in. Episode one is the manifesto; episode two is on the future of fashion.

Research work

BeatLib, the retelling benchmark, Beatcode and the rest of the RLNR program are in the research section, with their status.

§3 The lens

Civilisations are built on innovations and narratives.

01Innovation changes what is possible. Narrative changes what people believe. The greatest problems in the world get solved where the two meet.

02Religions were the first media companies. A text, a programme of practice, and a community bound by one story, carried across centuries by believers who became evangelists. No distribution deal has ever matched it.

03A frontier lab or startup is a mission in search of believers. My work is to give a mission its text, its practice and its community, so the people who understand it become the people who carry it.

§4 Residence

As a Storyteller-in-Residence, I work with your firm in three ways.

01Writing Room. Your team and I form a writing room to engineer the firm's narratives and push them out, directly or through PR. I own it and I drive it.

02Programs. I structure programs that build a community of practice around your firm, so the people who should believe in the work have a place to gather around it.

03New Media. If you want to go beyond narratives and build a media house of your own, I put it together: newsletters, podcasts, documentaries, shows and events.

 I take a few residencies a year, usually by introduction. Write to me.

§5 Research

I am learning by research, and this is what I am working through right now.

This is personal research, not a lab programme. I am a storyteller learning how models are trained to write, and the question I keep returning to is whether story structure can be checked the way code is checked. I call that idea RLNR, reinforcement learning from narrative rewards, and these are the threads I am pulling on.

The model writesa story A verifier checksthe structure A judge readsfor taste The two becomeone reward beats, value shifts, commitments surprise, specificity, voice the reward trains the next draft

Fig. 1. The RLNR loop as I currently understand it. The verifier's share is the part a storyteller can make checkable, so that is where I am looking.

  • Retelling fidelity as a benchmark

    A story that works survives retelling. I am testing this by having models retell a story and checking which beats, value shifts and claims survive, against a beat graph I draw by hand first.

    trying now · a hand-run experiment, no GPUs needed

  • Beatcode

    Reinforcement learning on storytelling by making a model write a real story language, ink, so that compile success, branch reachability and declared value shifts become verifiable rewards alongside a judge. The question is whether you can code a story rather than generate one.

    sketching · a small open model, modest compute

  • Story as spec for agents

    The Storytelling Engineering question, made testable: do agents that receive a narrative brief keep the intent across handoffs better than agents that receive a specification? If they do, storytelling is an alignment tool, not a marketing one.

    thinking about · the question labs would care about

  • BeatLib benchmarks

    A skill has no benchmark of its own, so I measure the lift it gives a fixed base model on a fixed rubric, on briefs it has never seen. Constraint satisfaction today; a slop floor, a faithfulness check and a calibrated judge next.

    ongoing · in the open on GitHub

  • Autoformalising rule books

    What happens when a story rule book, the Ramayana or a constitution is formalised well enough for a machine to check a draft against it. This is where formal methods and story meet, and where the Natya Shastra's rasa becomes a typed vocabulary instead of a metaphor.

    exploring
§6 Reading

These are the papers I am reading, with the note I took on each.

First sweep, November 2026, added to monthly. The note is mine, not the abstract.

  • arXiv:2602.05125

    Rethinking Rubric Generation for LLM Judges and Reward Models

    Shen et al., Meta, Feb 2026

    The paper that gave me RLNR. It splits vague criteria until they bite and fires the ones that reward the weaker draft. Every writers' room does this; now it is a training signal.

  • arXiv:2601.17226

    Retell, Reward, Repeat

    Liu, Muston, Srirag, Joshi, 2026

    The only paper so far that put a narratology book (Todorov) into the reward. Closest thing to my retelling test, and the first I will try to beat.

  • ICLR 2026

    StoryAlign: Evaluating and Training Reward Models for Story Generation

    Xia, Peng et al., Tsinghua

    Finds that reward models prefer machine-written stories to human ones. That bias is the whole problem in one table.

  • EACL 2026

    LitBench

    Fein, Russo, Xiang et al., Stanford

    The reference dataset. The best judge agrees with humans 73 percent of the time; the gap is where structure has to do the work.

  • CHI 2024

    Art or Artifice? The Torrance Test of Creative Writing

    Chakrabarty et al.

    Fourteen tests written with expert writers. Model stories passed three to ten times fewer of them. Read it as a craftsperson, not as a benchmark.

  • arXiv:2606.04095

    POLARIS: Guiding Small Models to Write Long Stories

    2026

    A 9B model, four A100s, competitive long fiction. The budget that makes Beatcode possible on a residency's income.

  • ICML 2026

    Rubric Curriculum RL

    2026

    A moving rubric is harder to hack than a fixed one. Useful for anyone who has watched a writer learn to game the note.

  • AAAI 2026

    RLMR: Reinforcement Learning with Mixed Rewards

    2026

    Half the reward is a judge, half is a verifier, and the mix moves per sample. The closest existing design to the one I want to build.

  • ICML 2026

    NCP-Bench: Narrative Commitment Preservation

    2026

    Checks an interactive story turn by turn against a ledger of facts and commitments. A commitment ledger is a beat graph by another name.

  • arXiv:2605.27878

    Narrative Flattening

    Li, Zhu, Wu, Bao, Evans, Chicago

    Measures how post-training compresses stories toward one mundane regime. The scientific version of what I called slop in Understanding Taste.

§7 About

I am a failed founder and a failed filmmaker, and blending the two became my job.

I live in two worlds. One is frontier AI, which I see every week through the fund. The other is culture: cinema, the internet, my own family. Beats is about the gap between them.

Where AI is going is knowable if you pay attention. Where people are going is the mystery, and that is what I chase.

I write in the first person and I am not neutral. I work in venture and I am betting my own time on this.

I take a few residencies a year, usually by introduction. If you are building something at the frontier and want to talk, write to stories@itihasa.co, or find me on X and LinkedIn.