Skip to content
twicetoldJoin the waitlist

How we check every story

Twicetold is built on dialogic reading. Here is the method, the rubric, the safety checks, and every result we have, good or bad.

Research basis

Dialogic reading is a way of reading picture books aloud in which the adult asks questions, listens, and builds on the child’s answers. Twicetold’s questions follow its CROWD prompt types: completion, recall, open-ended, wh- questions, and distancing (linking the story to the child’s own life). Each question also gives the parent a hint and a tip for expanding the answer, following its PEER sequence: prompt, evaluate, expand, repeat.

What the evidence says. The U.S. What Works Clearinghouse found positive effects of dialogic reading on oral language for preschool children, and a 2022 systematic review summarises the wider literature. There is no evidence yet for Twicetold itself. We describe the product as built on dialogic reading, and we don’t claim more than that.

Question rubric, v1.0 (draft)

Every question must pass all four criteria. A judge model scores each one yes or no. On any no, the story is rewritten with the judge’s feedback, up to two more times.

1. Needs reasoning, not recall

Answering takes prediction, inference, cause and effect, feelings, or a link to the child's own life. Remembering a fact from the page is not enough.

Passes
Why do you think Pakoda didn't eat the biscuit straight away?
Fails
What did Riya name the dog?
The answer is a fact from the page. The child only has to remember it.

2. A child of that age can answer it

The words, the idea, and the amount of thinking fit the story's age band.

Passes
How do you think Riya felt when she saw the puppies? (ages 5 to 6)
Fails
What does Pakoda's behaviour tell us about how street dogs survive in a city? (ages 5 to 6)
Too abstract for a five-year-old, who can't draw on any experience to answer.

3. Open-ended

There is more than one good answer, and yes or no is not one of them.

Passes
What could Riya do tomorrow to help the puppies?
Fails
Was Pakoda hungry?
A yes-or-no question. The child can answer without thinking aloud.

4. The text before it doesn't answer it

The paragraphs before the question don't state or strongly signal the answer.

Passes
(Asked before the puppies appear) What do you think Pakoda is going to do with the biscuit?
Fails
(Right after "Riya felt worried because Pakoda was gone") How did Riya feel?
The sentence before the question gives the answer away.

Open question. Dialogic reading uses recall and completion prompts with three- and four-year-olds, but criterion 1 rules out recall-only questions. We'll decide whether ages 3 to 4 get an exception before v1.0 is final, and record the decision here.

Changelog

DateVersionChange and reason
7 November 20261.0 (draft)First draft, written from the story generation contract. Needed before the first eval run. Not yet reviewed by an early-childhood educator.

Pre-launch eval run

200 synthetic parent notes, spread across the three age bands, each run through the full pipeline. Story model: claude-sonnet-5-5. Judge model: claude-haiku-5-5.

The first run hasn’t happened yet. Results will appear here, dated, when it does.

Run dateNot yet published
First-pass gate rateNot yet published
Safety hold rateNot yet published
Word-count complianceNot yet published

Judge and human agreement

An LLM judge is only useful if it agrees with a careful human. Anand hand-labels 200 questions against the rubric, without seeing the judge’s scores. We publish how often they agree and Cohen’s kappa, which corrects for agreement by chance, whatever the numbers show.

Agreement rateNot yet published
Cohen’s kappaNot yet published

Questions the judge failed during the eval run, each with the criterion it missed.

At least ten real rejections will appear here after the first run.

Safety red-team

At least 50 adversarial notes, written to break the safety checks. We publish the categories and pass rates, never the attack notes themselves.

CategoryPass rate
Prompt injection
A note that tries to give the model new instructions.
Not yet published
Requests for unsafe content
A note asking for frightening, violent, or copyable unsafe scenes.
Not yet published
Distress and abuse disclosures
A note suggesting a child may be in danger. No story; the helpline message instead.
Not yet published
A child writing as the sender
A note that reads as written by a child. No story; the parents-only reply instead.
Not yet published

Annotated sample story

This sample was written by hand to show the format. Generated samples will replace it after the first eval run. It starts from one parent’s note:

A parent's note
Riya fed the street dog half her biscuit and named it Pakoda.

Riya and the Biscuit Promise

4-minute read, ages 5 to 6

Every evening, when the sky over Riya's lane turned the colour of mango juice, a small brown dog came to sit by the blue gate. She had one floppy ear and one ear that stood straight up, as if it was listening for something far away. Nobody knew where she slept.

Today, Riya had saved half of her biscuit from snack time. She wrapped it in a paper napkin and held it carefully all the way home in the auto, even when the auto bumped over the speed breaker near the temple.

"You need a name," Riya told the dog. She thought hard. The dog was brown and a little crinkly at the edges, just like the pakodas Nani made whenever it rained. "Pakoda," said Riya. "Your name is Pakoda." The dog sneezed, and Riya decided that meant yes.

Riya unwrapped the biscuit and held it out. Pakoda waited, patient as a statue. Patient means she didn't push or snatch or hurry at all. She sniffed the biscuit. She sniffed Riya's fingers. Then she did something Riya did not expect. She took the biscuit gently in her mouth, the way you might carry an egg, and she did not eat it.

Pakoda walked three steps down the lane and looked back at Riya. She walked three more steps and looked back again. Her straight ear twitched.

CROWD type:
Wh- prompt (prediction)
Why it passes:
The child has to combine two clues (she won't eat it, she keeps looking back) to make a guess. Nothing before the question says where Pakoda is going.
Expand tip for the parent:
Whatever your child guesses, ask "What made you think that?" and say their idea back in a full sentence.

"Amma!" Riya called. Amma came out, wiping her hands on her dupatta. "Can we see where she's going? Just to the corner?" Amma looked down the lane, then took Riya's hand. "Just to the corner," she said. Together they followed Pakoda, past the ironing cart and the bicycle with the flat tyre. Pakoda waited for them at every lamppost.

At the corner, under the big neem tree, sat a cardboard box. Inside the box, three tiny puppies were squeaking. Every one of them had a floppy ear. Pakoda set the biscuit down in the middle of them, and the puppies nibbled and nudged until there was nothing left but crumbs on their noses.

"You weren't hungry at all," Riya whispered. "You were bringing it home to share." The smallest puppy wobbled over and licked Riya's shoe, and she laughed so softly that only Amma heard. Amma squeezed her hand, and they watched the puppies tumble over each other until the streetlights came on.

On the way back, Riya asked if tomorrow she could save a whole biscuit, one half for Pakoda and one half for the puppies. "And one for me," she added. "We'll see," said Amma, which Riya knew usually meant yes.

CROWD type:
Distancing prompt
Why it passes:
It links the story to the child's own life, has no single right answer, and a five-year-old can answer it from their own memory.
Expand tip for the parent:
Add a feeling word to their answer: "So you felt proud when you gave Dada the last ladoo."

That night, Riya lay in bed and listened. Far down the lane, under the neem tree, everything was quiet. Four dogs were asleep in a cardboard box, and one of them had a straight ear, still listening, just in case. Goodnight, Pakoda.

New word: patient, waiting calmly without pushing or hurrying, explained in the story itself.

Pre-registered hypothesis

Registered on 7 November 2026, before any family has joined.

In aggregate, among families who opt in to research, the share of "Got it alone" answers rises from week one to week four.

We will publish the result whatever it shows, including no change or a fall.

Limitations

  • Synthetic notes are not real children. They test the writing, not what happens at bedtime.
  • LLM judges carry bias, which is why we check the judge against human labels and publish the agreement.
  • Parent taps ("Got it alone," "With a hint," "Not yet") will be self-reported, and parents may be generous.
  • There is no outcome evidence for Twicetold yet. The research behind dialogic reading is about the method, not this product.

External review and open repo

We're asking an early-childhood educator to review the rubric before v1.0 is final. Status: not yet reviewed.

The rubric, prompts, synthetic notes, and stored run outputs will be public on GitHub with the first run, along with one command that rebuilds every table on this page.