Subtitle Fact Audit

Required key.

Add this in your Lovable project under Settings → Secrets before pasting the prompt below.

Paste into a fresh Lovable project. Make sure the key above is set first. read the build strategy →

Build "Subtitle Fact Audit" as a ONE-SHOT Lovable build (~5 credits, no follow-ups).
Single-page TanStack Start app, one route, cut scope ruthlessly.

THIS PROMPT IS SELF-CONTAINED
Do not assume the participant knows Jev or has read any other page. TypeSafe Jev is a decision model,
not a text generator. It receives `state` plus a map of typed `questions` and returns JSON that code can
use directly. Use Choice to select exactly one defined option, Noul for a yes/no probability from 0 to 1,
and Score for a probability-weighted position across ordered levels. Keep rules, thresholds, arithmetic,
and actions in code. Use a generative model only if the product separately needs prose.

CONCEPT
Paste subtitle files, source dialogue notes, and approved names or terms, and Jev verifies flagged subtitle fields against evidence, escalating uncertain corrections.
Discipline: Videography & Film (Caption quality control).
Jev pattern: Verify & Escalate (check then hand off) — a Jev `noul` checks each claim/field against its evidence; low or mid probabilities are escalated to a human review queue instead of auto-acting.
Why typed decisions: Typed mismatch decisions identify review-worthy caption errors without rewriting dialogue or silently changing meaning.

ENGINE: TypeSafe Jev (typed decisions, not text)
- Before writing code, ask the participant for `TYPESAFE_API_KEY` with Lovable's secure secret form. Never
  ask them to paste it into chat, never prefix it with `VITE_`, and never expose it to browser code.
- Put the callable server function in `src/lib/jev.functions.ts`; do not import a server-only module from UI code.
- Validate the user's input before the handler and read `process.env['TYPESAFE_API_KEY']` inside the handler.
- From the server, call:
  POST https://api.typesafe.ai/v1/systemone
  Authorization: Bearer <TYPESAFE_API_KEY>, Content-Type: application/json
  body: { "model": "jev-latest", "state": <Subtitle cues, source notes, and approved terminology>, "questions": QUESTIONS }
- Ask all questions in ONE request (they run in parallel and cannot see each other).
- Read Choice from `answers[id].choice`, `.probabilities`, and `.confidence`; Score from `.score`,
  `.probabilities`, `.legend`, and `.confidence`; Noul from `.noul` only (it has no separate confidence).

SERVER IMPLEMENTATION (adapt the input name and result mapping to this concept)
```ts
import { createServerFn } from "@tanstack/react-start";
import { z } from "zod";

const Input = z.object({ text: z.string().trim().min(1).max(20000) });

export const decide = createServerFn({ method: "POST" })
  .inputValidator((input: unknown) => Input.parse(input))
  .handler(async ({ data }) => {
    const apiKey = process.env['TYPESAFE_API_KEY'];
    if (!apiKey) return { ok: false as const, status: 500, message: "TYPESAFE_API_KEY is not configured." };

    const requestBody = {
      model: "jev-latest",
      state: { text: data.text },
      questions: QUESTIONS,
    };

    let response: Response | undefined;
    for (let attempt = 0; attempt < 3; attempt += 1) {
      response = await fetch("https://api.typesafe.ai/v1/systemone", {
        method: "POST",
        headers: { Authorization: `Bearer ${apiKey}`, "Content-Type": "application/json" },
        body: JSON.stringify(requestBody),
      });
      if (response.ok || (response.status !== 429 && response.status < 500)) break;
      if (attempt < 2) await new Promise((resolve) => setTimeout(resolve, 750 * 2 ** attempt));
    }

    if (!response) return { ok: false as const, status: 503, message: "Jev did not respond." };
    if (!response.ok) {
      const message = (await response.text()).slice(0, 500) || `Jev request failed (${response.status}).`;
      return { ok: false as const, status: response.status, message };
    }
    const result = await response.json();
    return { ok: true as const, answers: result.answers };
  });
```

In the page component, call `decide` with `useServerFn`, preserve the participant's input while loading or
on error, disable duplicate submissions, and render the returned typed values rather than parsing prose.

QUESTIONS (start here, refine wording if needed)
{
  "field_verified": {
    "type": "noul",
    "instructions": "Does the evidence verify this subtitle field as supplied?",
    "criteria": {
      "true": "The field matches the source notes or approved reference.",
      "false": "It conflicts with, or is not supported by, the supplied evidence."
    }
  },
  "mismatch_type": {
    "type": "choice",
    "instructions": "What kind of field mismatch is present?",
    "criteria": {
      "none": "No mismatch is supported by the evidence.",
      "proper_name": "A person, place, or organization name differs.",
      "number_or_date": "A number, date, or measured value differs.",
      "meaning": "The wording changes a documented factual meaning."
    }
  },
  "source_match": {
    "type": "score",
    "instructions": "How directly can this field be compared to its source?",
    "criteria": [
      "No usable source",
      "Partial or ambiguous source",
      "Clear source with minor ambiguity",
      "Exact corresponding source available"
    ]
  }
}

DECISION RULES (in code, not in the model)
- choice/score: if confidence < 0.6 show a "needs review" badge instead of acting.
- noul: >= 0.75 = yes, <= 0.25 = no, otherwise "review".
- score: normalise with score / (levels - 1) before combining with weights.
- Missing or malformed answers go to review. Show safe HTTP status and message in the UI. Retry only 429/5xx,
  at most twice with increasing delay; never retry 400/401/402/403/404 and never invent results.
- If the product needs prose (a summary, a message), use Lovable AI for that text only; Jev makes the decisions.

UI
claim cards marked supported / unsupported / needs review, with a review queue column. Include 2 sample inputs as one-click chips so the demo works instantly.
Tailwind + shadcn, editorial look, footer: "Built during the Creative AI & Quantum Hackathon organised by StreetKode Fam during Indian Krump Festival 14".

NO database, NO auth, NO extra pages, NO tests.

BEFORE YOU FINISH
- Run the app and test both sample inputs through the real server function.
- Confirm the browser receives no API key and no fabricated fallback result.
- Confirm Choice/Score confidence and Noul probability drive the stated review rules.
- Confirm loading, success, review, empty-input, and service-error states are readable on mobile.

Reference only if needed: https://docs.typesafe.ai/api and https://docs.typesafe.ai/primitives

Market sizing.

Indicative figures for hackathon pitches — refine with your own research before raising.