.claude/skills/video-script-production/references/research-protocol.md

# Research protocol: building the fact base (Step S3)

This is how the Moser Baer facts were gathered and verified, as a procedure. The thesis of v4 and the whole shape of v5 came from this stage; both changed because of what it found. `research-log.md` has the actual queries and pages. Output: `FACTS.md` (template from `scaffold.js`); the script's ledger is derived from it.

## 1. The rules that earned their place

1. **The person before the company.** The Moser Baer first fact base was built from rival YouTube videos and company results; the two Business Today profiles of the founder (2003, 2011), which hold his origin, the Swiss partner, the cocktail party and the floppy bet, were never opened. Opening them turned a story about a bond into a story about a man who always left what was running out. Search the founder, the family, the early years and the first deal before anything else.
2. **Verify the link, not just the numbers.** List every "X caused Y" the script will lean on and find a source for the link itself. v3 said the solar bond turned decline into debt; the sources said the debt was already there in 2005 and none says where the bond's money went. The thesis changed to the weaker, supportable one, and the script says aloud what it cannot show.
3. **Cross-check the magnitude of every key figure in a second source.** v4 used one article's "In 2004, 3,500 crore, largely funded through debt". Another magazine had total debt at 1,620 crore in 2005 and 3,500 crore in 2011: the same number, a different year. It was a garble, and it was in v4 until v5 dropped it.
4. **Prefer primary or near-primary sources:** rating-agency notes (CARE), liquidator and tribunal documents (IBBI claims, NCLT orders as reported), the company's results as reported by a wire (Business Standard, Storage Newsletter), the press conference as reported by the trade press (pv magazine), court reports, and the long-form business magazines (Business Today, Outlook Business, Moneylife). Wikipedia is a map, not a source: use it to find the cited page.
5. **Mark every row.** `[V]` checked against a non-YouTube page you opened; `[V-search]` seen only in a search-result summary (the page could not be opened; confirm before filming); `[CHECK]` unverified; `derived` computed (show the arithmetic); `analysis` your reasoning, which the voice-over must flag as such.
6. **Rival-video claims you cannot verify go in the do-not-use table** with the reason. Moser Baer: "solar unit sold for 72 crore", "787 crore", "8,000 crore", market shares of 17, 70 and 80 percent, workers burning effigies. Dramatised or single-source claims about private individuals (deaths, self-immolation) stay out on purpose.
7. **Sources disagree; say so, and use the safer one.** The gate closing was 3 November in one report and 4 November in another, so the script says "November 2017". Warburg's holding after May 2012 differed between two magazines, so the script does not say what it kept.
8. **A person named in the script needs a record**: on-record quotes in a national outlet, not a rumour. A private person is a decision for the user.

## 2. The order of work

1. **Read the rival brief** (`report.md`, `synthesis.json` from S1): the timeline with support counts, `contested`, `verification_queries`. These are your first search list. Names and numbers in the captions are unreliable.
2. **The person and the early company**: founder profile, family, first jobs, first company, where, why they left, partners, the first big bet, and who doubted it. Quote them (use the fetch prompt below).
3. **The rise**: who financed it (names, amounts, dates), the cost advantage, capacity, the peak numbers, the rival that was ahead. Ask: what was the company's own explanation then?
4. **The turn**: the first bad result (a results announcement), what the company blamed, what changed in the market (a price series from an industry body). Write the prediction your thesis makes and look for the figure that tests it.
5. **The responses**: each diversification or rescue (new business, funding, deals), with dates and amounts, and what the company said in public at the time (press conference quotes).
6. **The debt and the date**: rating-agency notes, bond terms, due dates, restructuring agreements, missed payments.
7. **The end**: the insolvency petition (who filed, when), the buyers who walked away, the liquidation order, the claims and the liquidation value.
8. **The people**: workers' accounts (named, on record), when the signs began (benefits, salaries), what happened to them afterwards.
9. **The blame question**: criminal cases (who alleged what, the amount, the agency, the court decisions, appeals), and whether any rival claim is in any record. Compute the scale of the allegations against the claims.
10. **The rival claims you will refute**: for each number a rival video uses and you do not, search for it; record "not found in any court record or news report" only after you looked.

## 3. How to ask the fetch tool so it does not paraphrase

Fetch tools summarise. Ask for what you need in the prompt: "Give the exact sentences, quoted verbatim, about X"; "What is the publication date or issue date of this article, and the year?"; "Return the article text verbatim as far as available; say if it is paywalled or truncated". If the page cannot be opened (an app-wrapped magazine page failed twice), search for the article's title and key names, accept the summary, and mark the row `[V-search]`.

## 4. Query templates (replace the entity)

- `<founder> founder profile early life how he started <first product> <city>`
- `<company> history founded <year> <partner> <first plant>`
- `<company> <year> results net loss revenue <business> Business Standard OR Storage Newsletter OR pv-magazine`
- `<company> bond OR FCCB OR debenture <amount> due <year> recast OR restructuring`
- `<company> CARE OR ICRA OR CRISIL rating <year> continuing losses negative net worth`
- `<company> workers plant closed <year> salary unpaid <city> union president`
- `<company> insolvency NCLT petition <lender> liquidation order claims <year>`
- `<company> CBI OR ED case <bank> crore discharged OR chargesheet`
- `<industry> price fall <years> <industry body>` (for the market that moved)
- `"<company>" <exact figure a rival video used>` (to refute it)

## 5. Per row, note the stake

For each fact, note whose loss it is, which belief of the viewer's it breaks, or which contrast it makes. A figure with no stake goes in the ledger and not in the voice-over (Lesson A). The same note becomes the stake line and the landing in S5.

## 7. Harvest what already exists (Step S1b)

Written after the Videocon run (2026-11-01). The Moser Baer facts came from pages found one at a time during S3. For a topic with a long paper trail this is too slow and too easy to bias, so collect the existing writing first, save it, and read it from disk.

1. **Search wide, in rounds.** Use the section 4 templates plus: `<company> NCLT OR NCLAT OR court order <year>`, `<company> annual report`, `<company> rise and fall` (finshots, Outlook Business, Business Today, Open, Moneycontrol), `<founder> profile`, `<company> trivia`, and the name of every regulator, agency and rival that touched the story. Record the URLs of every promising page in `sources/urls-roundN.txt` with a one-line note.
2. **Save the pages.** `node .claude/skills/video-script-production/scripts/fetch-articles.mjs --dir research/<slug>/sources --urls <file>`. It saves readable text, saves PDFs as they are, and lists blocked pages (403, paywall, consent wall) in `INDEX.md`. For a blocked page you still need, open it with the browser or WebFetch, or keep the fact at `[V-search]`. PDFs: `pdftotext -layout file.pdf file.txt`; a scanned PDF gives no text (a CBI FIR did), so read it with the Read tool's `pages` option or leave it as an open item.
3. **Google AI Mode as a link finder.** Ask for a dated timeline, cited, with the figures that sources disagree on; and separately for trivia. Read with `javascript_tool`; strip URLs from the text before returning it (the tool blocks text with query strings), read in slices of about 900 characters, and read the `a[href]` list by host and path. Lessons: (a) the timeline answer matched the saved articles; (b) the trivia answer named "archives" and magazines without links, and its colourful items (a founder waiting days at a ministry, quoting philosophy to a Japanese partner) could not be traced, so they are leads only; (c) it cannot be saved whole, so keep your own notes in `sources/ai-mode-notes.md` and say what was not read; (d) a page posting data to localhost does not work from the AI Mode page (blocked), so do not build on that.
4. **The primary document first.** A tribunal or court order lists dates, parties, numbers and the parties' arguments in one place (the NCLAT order for Videocon gave the claims, the haircut, the consolidation dates and the company's own table). Read its background and submissions sections with `Grep` for the numbers you need, not the whole text.
5. **Read news pages through the noise.** Saved pages carry menus and adverts. Filter lines longer than about 80 characters before reading; keep the article text; write the date next to each fact.
6. **Measure the rival videos against the harvest.** Grep the rival transcripts for the facts the harvest found most important (a tribunal name, a late-stage event, a number). Where none of the eight says it, that is a gap a new video can fill; write it as a measurement, not an impression.
7. **Write `TIMELINE-AND-TRIVIA.md`**: the shape of the story as the sources give it, a dated timeline (status on every row), verified trivia, unverified leads (kept apart), a table of figures that disagree with the safer reading, what the rival videos get wrong and miss, open items, and what only the user can judge.
8. **Stop** when the primary document is read, the end of the story is dated to this year (search for the latest hearing, ruling or sale, and say plainly when you could not find one), and each beat you expect has at least one saved source. Do not mine a long annual report or a scanned PDF unless a beat needs it; list it under open items.

## 8. What to produce, and when to stop

`FACTS.md` with the five tables (facts, causal claims, magnitude cross-check, people, do-not-use). Stop when each beat you plan has at least one sourced fact for its claim, every causal link has a source or has been softened, and every key figure appears in two places. Do not stop at the first plausible account. Do not keep researching past the point where the open items are things only a primary document can answer (the bond's use of proceeds); list them and say aloud in the script that you could not find them.