skills/ai-film-pipeline-master/references/phase-01-story-generation/cat-info-4.md

# Information Control — Part 4

Category `info`, continued. Parts 1 to 3 are in `cat-info.md`, `cat-info-2.md` and `cat-info-3.md`. Nothing here repeats them.

**Part 4 of 4.** 2 cards.

These cards close gaps the database reported about itself: each was named as missing by an agent that needed it and could not cite it.

---

### The Fact-Check Pass

**Also called:** the check, checking, the annotated draft, the source-marked script, verification pass, the read-back, running the piece against the notes
**What it is:** A separate pass over a finished draft in which every checkable claim is marked and made to carry its own source at the point it is made, run against notes, recordings and documents rather than against the writer's memory of them.
**Effect on the audience:** Nothing, when it works — which is why it gets cut. What it prevents is the specific collapse where one wrong particular takes the whole piece down with it: an audience that catches a checkable error stops grading claims and starts grading you, and every unfalsifiable thing in the film — the interpretation, the tone, the argument you actually care about — is re-read as the same quality of work as the number that was wrong.
**Used for and where it works best:** In this pipeline it runs where the brief asks for a checked factual piece or on a history piece the user wants realistic, and there it follows root `SKILL.md` §0 — a claim with no source found is attributed or given the closest supported wording, recorded, and the work continues; it is never a stop. Everything non-fiction, and the heritage documentary hardest of all, because its authority is built out of exactly the particulars that break (`expo.heritage_specificity`). Run it as a procedure, not as a mood. **One: mark the draft.** Go line by line and mark every claim that could in principle be wrong — every number, date, name, title, place, quotation, attribution, causal statement and superlative. Mark the ones inside quotes too; a contributor's error published unmarked is your error. The physical act matters: magazine checking departments put a mark on each verified word precisely so that an unverified one is *visible* rather than merely unremembered. **Two: attach a source to each mark, at the claim, not to the piece.** A bibliography is not a check. Each marked claim gets the specific thing that supports it — document, page, timecode, the recording at 41:18 — and a claim whose "source" is another sentence in your own draft is unsourced. **Three, the step writers skip: check the ones you are most confident about.** Errors do not live where you felt uncertain; you already looked those up. They live in the facts you have known for years, which is why checkers routinely kill things authors would have sworn to. If a claim went in without a lookup because you simply knew it, it is not checked, and it is where your check should start. **Four: know the five classes that fail most** — the number that drifted while being retold (round it once and the round version becomes the fact); the date converted between calendars, which is off by a day roughly whenever anyone uses a formula instead of the actual calendar, and off by a year whenever a regnal or era date is silently treated as an absolute one; a name's spelling in transliteration, where there is often no single right answer and the job is to pick one, source it and never drift; the quotation that has been paraphrased so long that the famous wording appears in no primary text at all, only in secondary repetitions of it; and the causal claim that was a correlation in the source and became a cause somewhere in your drafting, usually at the sentence where you compressed. **Five, the craft rule that decides the hard cases: a claim you cannot source does not get softened, it gets cut or attributed.** Hedging keeps the assertion and removes the accountability, which is worse than either option — "widely believed to be", "some say", "possibly the first" all put the claim in the audience's head while leaving you nothing to stand on when it is challenged. Either it goes, or it becomes somebody's named position: not *the fire saved the tablets* but *the excavator's explanation was that the fire saved them*, which is a different, smaller and completely defensible sentence. **For anyone working on the ancient world from secondary sources, hold three tiers apart and never let a claim move up one.** A primary source is the tablet, the inscription, the chronicle, the excavation report. A scholarly reading is what a named specialist concluded from it, and it is evidence of a reading, not of the event. A popular repetition is a documentary, a museum caption, a coffee-table book or a Wikipedia line quoting the reading — and it is not a source at all, it is a rumour with a bibliography, because circular citation means the popular version often ends up cited back to a "reliable source" that was itself repeating the popular version. **A claim repeated in ten videos, including your own previous ten, is not a checked claim; it is one unchecked claim with nine echoes.** Two-thirds of a heritage check is walking a familiar sentence back down the tiers to see whether anything is underneath it. **And when the check breaks the story, the check wins — but say what that costs, because it is the real test.** If the fact that anchors your best sequence does not stand up, the available moves are, in order: restructure so the piece is about what is actually known; make the uncertainty itself the material (`info.disputed_history`, `info.contradictory_sources`); publish the claim with the state of the checking attached, which is a deliberate and narrow choice (`info.unverified_allegation`); or drop the sequence. What is not available is keeping the sequence with a softer verb. This is why the pass belongs at the end of a draft and not at the end of the schedule — checked after lock, the only fix left is a hedge, and a piece full of hedges is a piece that fact-checked itself into being unbelievable. **The boundary:** this is what you do before publishing. `info.the_correction` is what you publish after, when this pass missed something. `info.source_on_screen` is the separate decision about what the *audience* is told about a source you have already verified — a checked claim can carry no super, and an unchecked one is not rescued by a citation. Where the brief names a broadcaster or distributor that asks for an annotated script, this pass produces it.
**Best in:** formats: Documentary, Heritage Documentary, Reportage, Long-Form Non-Fiction, Podcast, Essay | genres: Journalism, Heritage, Historical, Science, Biography, Advocacy
**Avoid when:** The claim is a judgement rather than a fact. You cannot check an interpretation, and running one through this procedure produces false precision — a sourced-looking sentence that is really your reading wearing a citation, which is the more dangerous error because it is now armoured. Separate the two lists before you start and check only the first. Also honest about the limit: this pass catches errors, not the absence of context, and a piece can survive a perfect check and still be untrue in the way that matters. And it is not a legal read, which this pipeline does not run.
**Example:**
> Heritage VO, marked. Every claim numbered, each one needing its own source:
>
> "Ashurbanipal's library at Nineveh held more than thirty thousand tablets ⟦1⟧. It was the first library in the world ⟦2⟧. The king wrote that he could read the obscure Sumerian and the difficult Akkadian ⟦3⟧. When the city fell in 612 BC ⟦4⟧, the fire that destroyed the palace baked the clay and saved the collection ⟦5⟧ — which is why we can read Gilgamesh today ⟦6⟧."
>
> **⟦1⟧ Corrected — number drift.** The catalogue count is roughly thirty thousand *tablets and fragments*; the whole-tablet figure is far lower. I had been saying "tablets" for three years because that is how I first heard it. New line: "around thirty thousand pieces — most of them broken."
> **⟦2⟧ Cut — superlative, unsourceable.** Older palace and temple archives exist. There is no authority I can name for "first", and the softened versions — "one of the first", "possibly the oldest" — keep the claim while removing my accountability for it, so they are worse. Cut entirely. The paragraph loses nothing; it never needed a record.
> **⟦3⟧ Kept, with the translation named.** The words are from the king's own inscription, but "obscure" and "difficult" are a translator's English, not his. Attribute the rendering on screen rather than presenting a translation as a quotation.
> **⟦4⟧ Kept — sourced to the chronicle**, not to the twelve other films that say 612. Note in the script which tablet, so the next writer does not have to redo this.
> **⟦5⟧ Attributed — correlation promoted to cause.** The excavation account offers this; it is a reading, and the counter-reading is that the fire shattered as much as it preserved. New line: "The men who dug it out believed the fire had baked the clay and saved it. Others think the fire is why so little is whole."
> **⟦6⟧ Kept.**
>
> Three claims, three different fates, and not one of them softened.

`info.fact_check_pass`

---

### The Source You Cannot Name

**Also called:** granting anonymity, the unnamed source, the protected contributor, blind quoting, the useful description, writing around a name
**What it is:** The decision not to name a person, treated as a writing problem rather than a permission: the name is withheld, so a substitute has to be built out of whatever else about that person makes the claim believable, at the finest granularity safety allows.
**Effect on the audience:** A named source is asserted; an unnamed one is discounted, and the reader does the discounting whether or not you want them to — the measured finding is that stories carrying only unnamed sourcing are rated less credible, and that specific attribution rates higher than vague attribution on the same sentence. That is the cost and it does not go away. What the description recovers is the *grounds*: a reader told nothing tests the claim against the publisher's reputation, a reader told how this person came to know it tests the claim against the claim, which is the exchange you want. `info.source_on_screen` is the same reading applied to documents and figures; this is the version where the attribution is a human being and the strongest word in it has been removed.
**Used for and where it works best:** Investigative reportage, workplace and institutional stories, testimony documentary, and any community or diaspora piece where naming a person reaches their relatives before it reaches your audience. Six rules, in order. **Describe the person by what makes them credible for this exact claim, not by their proximity to it.** "A source close to the matter" is a hole in the sentence; "a woman who had worked in the building for nineteen years" is an argument. The test is mechanical — your description answers *how would they know?* and nothing else. Rank, tenure and duty answer it; adjectives never do. **Decide granularity claim by claim, and spend it where the writing lives.** Anonymity is not one switch: you withhold a name, and possibly a role, a date, a department, a gender, and each withheld item is separately costed. Audit in both directions — strip what identifies (the jigsaw problem is real: initials, a location, a unique duty, an unusual detail of access, and the organisation's own logs will finish the identification you started), then keep and *use* everything that survives. A writer who redacts by instinct deletes the nineteen years along with the name and has nothing left to write. **Give the reason for anonymity in the sentence, and make it a reason rather than a formula.** "Who was not authorised to speak" is a category; "who still works there and is a witness in the pending case" is a reason, and it hands the reader the motive to weigh, which they will look for anyway. **Put the claim to the subject without handing over the source.** An unnamed account still has to survive `scene.right_of_reply`, and an approach specific enough to get an answer can be specific enough to identify who gave it to you — that collision rewrites the letter, it does not excuse skipping it. **The screen case is a different craft problem.** A silhouette, a back-of-head, a pixelated face, a re-voiced line or an AI-generated stand-in is not a quiet convention like italics; it is a visible admission, on screen, for as long as the contributor is, and it costs the thing a documentary interview runs on — a face doing something while it speaks. So earn it: say in words, early, why this person is disguised; keep the disguise consistent so the audience stops re-solving it; and shoot something that does the work the face was doing — hands, the room, the pause before the answer. Prominent disguise is the worst of both, which the 2026 AI-avatar experiments proved: viewers reported unease rather than empathy, and the better the mask the harder it pulled focus. If the disguise is synthetic, label it, or you have imported `scene.reconstruction_unmarked` into your interview. **None of these disguises is a protection method, and whether one is enough is not this card's question**: a silhouette still carries build, voice and gait, and `doc.anonymous_interview` withdraws the method and hands the protection brief to a person whose profession it is (root `SKILL.md` §0). **When the reason is danger rather than preference, the writing changes too.** Preference gives you a description problem; danger gives you a structural one. You lose the right to be interesting about the details — the specific duty, the memorable incident, the quotable turn of phrase are precisely what identifies them — so the weight moves off the person and onto corroboration: the document, the second account, the thing this source's claim predicts that you can check by yourself. Build the piece so the anonymous account confirms something you can stand up another way, never so it is the load-bearing beam; and say plainly that you have done this, because a reader who watches you work around a hole trusts the hole more than one who trips into it. **The failure to name is anonymity used to make a claim unfalsifiable.** One diagnostic question — *if this claim were solid, would I have kept the name?* When the answer is yes, the anonymity is not protecting a person, it is protecting the claim from being checked; a reader who cannot scrutinise a source cannot dispute one. That is the version editors have complained about for twenty years: anonymity as routine rather than exception, granted by habit, and handed to powerful people who use it to attack from behind a screen the piece provides for free. The second-order damage is worse than the first — the honest use gets discounted along with the lazy one.
**Best in:** formats: Reportage, Documentary, Documentary Interview, Long-Form Non-Fiction, Podcast, Documentary Series | genres: Journalism, Testimony, Crime, Political, Workplace, Advocacy
**Avoid when:** A name is available and the only thing anonymity buys is a better quote. A person who will say something sharper unattributed than attributed is telling you exactly what they will stand behind, and the sharper version is usually the one that is not true. Avoid it too where the description you can safely publish identifies the person anyway to the only audience that matters — a village, a department of six, a congregation — because you have then paid the full cost of anonymity and delivered none of the protection; there the choice is publish with consent, publish without the claim, or do not publish. Do not reach for compositing as the way out: merging two informants into one "source" is `char.real_person_composite` applied to non-fiction, and in reportage that is fabrication. This card is also not the neighbouring problems — `info.unverified_allegation` is the claim you cannot prove, this is the person you cannot name, and a well-sourced claim from an unnamed person is not an unverified one; `char.living_subject_reader` is the inverse pressure, a named person who will certainly read it; `inter.consent_without_a_cut` is consent granularity where there is no final version, which is a different reason to negotiate at the clip rather than the piece. And never let a blanket "sources say" carry a film — `info.credibility_ladder` is the discipline that stops every unnamed account being weighted the same.
**Example:**
> **Named, and therefore finished:** *Rana Toma, the ward's senior night sister from 2011 to 2019, said the second incubator had been out of service since March.*
>
> **Unnamed, described by what makes her credible:** *A nurse who ran the ward at night for eight years, and who still works in the trust, said the second incubator had been out of service since March. She is not named because she is a witness in the pending case.*
>
> **Unnamed, hidden behind a label — the same fact, weighing nothing:** *A source close to the hospital said the second incubator had been out of service since March.*
>
> One claim, three weights. The middle sentence costs eleven words more than the last and is the only one a hostile reader can argue with, which is the whole of the difference.

**Yields to:** `info.unverified_allegation` — The problem is the claim, not the person.  
`info.anonymity_decision`

---

## Coverage

**2 cards** in this file: `info.fact_check_pass`, `info.anonymity_decision`.

## Sources

Verified 2026-09-17.

- [Fact-checking at The New Yorker — Columbia Journalism Review](https://www.cjr.org/critical_eye/fact-checking_at_the_new_yorker.php)
- [The Three Models of Fact-Checking — KSJ Handbook](https://ksjhandbook.org/fact-checking-science-journalism-how-to-make-sure-your-stories-are-true/the-three-models-of-fact-checking/)
- [How to Fact-Check — The TiJ Fact Checking Guide](https://thetijproject.ca/guide/how-to-fact-check/)
- [A Guide to Fact-Checking Investigative Stories — GIJN](https://gijn.org/resource/guide-to-fact-checking-investigative-stories/)
- [We Stand Corrected: When Good Journalists Make Stupid Mistakes — Poynter](https://www.poynter.org/reporting-editing/2007/we-stand-corrected-when-good-journalists-make-stupid-mistakes/)
- [The Expert's Guide to Fact-Checking Your Nonfiction Book — Narratively Academy](https://www.narrativelyacademy.com/p/the-experts-guide-to-fact-checking)
- [Errors & Omissions & Rights, Oh My! — International Documentary Association](https://www.documentary.org/feature/errors-omissions-rights-oh-my-guide-protecting-your-film)
- [List of citogenesis incidents — Wikipedia](https://en.wikipedia.org/wiki/Wikipedia:List_of_citogenesis_incidents)
- [Einstein didn't say that: how viral misquotes evolve and replicate — Big Think](https://bigthink.com/high-culture/how-viral-misquotes-evolve-and-replicate/)
- [Evaluating primary and secondary sources — Hofstra University Libraries](https://libguides.hofstra.edu/c.php?g=323882&p=2171901)
- [SPJ Ethics Committee Position Paper — Anonymous Sources — Society of Professional Journalists](https://www.spj.org/spj-ethics-committee-position-papers-anonymous-sources/)
- [Confidential sources — ONA Ethics](https://ethics.journalists.org/topics/confidential-sources/)
- [Anonymous sources: leaving journalism's false god behind — Poynter](https://www.poynter.org/reporting-editing/2014/anonymous-sources-leaving-journalisms-false-god-behind/)
- [Evaluating unnamed sources in news reports — News Literacy Project](https://newslit.org/news-and-research/evaluating-unnamed-sources-in-news-reports/)
- [Anonymous sources hurt credibility of news stories across cultures — International Communication Gazette](https://journals.sagepub.com/doi/10.1177/1748048514548534)
- [Source Attribution and Perceptual Effects — Bournemouth University Research Online](https://eprints.bournemouth.ac.uk/20891/1/Source%20Attribution%20and%20Perceptual%20Effects.pdf)
- [Digital and Physical Safety: Protecting Confidential Sources — Committee to Protect Journalists](https://cpj.org/2021/11/digital-physical-safety-protecting-confidential-sources/)
- [12 principles journalists should follow to protect their sources — Nieman Journalism Lab](https://www.niemanlab.org/2019/01/here-are-12-principles-journalists-should-follow-to-make-sure-theyre-protecting-their-sources/)
- [How Data Journalists Can Use Anonymization to Protect Privacy — GIJN](https://gijn.org/stories/how-data-journalists-can-use-anonymization-to-protect-privacy/)
- [3 Creative Ways To Hide An Anonymous Source On Camera — Storyhunter](https://medium.com/storyhunter/3-creative-ways-to-hide-an-anonymous-source-on-camera-cf2d1da3fd04)
- [Why Netflix Documentaries Are Turning To AI To Mask Contributors — Deadline](https://deadline.com/2026/03/ai-documentary-mask-netflix-lucy-letby-1236752359/)
- [Anonymity in the New York Times: By the Numbers — FAIR](https://fair.org/home/anonymity-in-the-new-york-times-by-the-numbers/)