Source text
AI documentary pipeline
Frameroo reads a book, checks its facts against sources, writes the narration, plans every shot as a director would, and reconstructs the scenes in Blender — the ones no camera can film any more. A human editor approves each stage.
A factual series starts with months of reading, verification, rights clearance and scripting, and then needs footage that often does not exist: a 1920s trading floor, a shipwreck, a battle, a laboratory that burned down. Networks either pay for a reconstruction shoot or settle for stock, stills and slow pans.
The expensive parts of that work are structured. Reading a source and pulling out what matters, checking a date against a record, deciding which shot carries a sentence, and building a scene to a shot list are all tasks a model can carry, under human review.
Frameroo takes a book and returns a finished documentary cut, with every claim traceable to a source and every scene rebuilt in 3D rather than filmed.
The stages run in order because each one needs what the last decided. Numbers below match the stage order in the running system.
The book is split into passages and read for the narrative spine: people, places, dates, turning points, and the claims that carry the story. The model searches for sources behind each claim rather than relying on what it already knows.
The script is cut into parts of about 550 words and each part is checked at the same time by its own model call with web search. Only real problems come back — a wrong date, a number, a name, an attribution, a quote that is not word for word — each with the exact words to replace. Fixes are applied by exact match; anything that cannot be found verbatim is reported, never guessed.
One director reads the whole narration and decides, sentence by sentence, what the viewer sees: an animated map, a real archive photo, the article behind a claim, a chart or dateline, a label over the footage. It hands each tool exactly its moments, so no two tools claim the same beat and a date graphic lands on the date.
Where no footage exists, the plan becomes 3D. Each moment is written as one to three shots in a fixed vocabulary — place, time of day, action, camera move, extras, shot length — and rendered in Blender, then graded to the look of the series and cut to the exact length of its line of narration.
Narration, scenes, graphics and sound are laid on one timeline and exported as an edit the team reviews, reorders and finishes in their own suite. Every stage writes its plan to disk, so a producer can see why each shot is there and change it.
Frameroo is being built on a working internal pipeline. Below is the fact-check stage as it is implemented: the gate that decides whether a script is allowed to reach the voice recording.
"""factcheck.py — the script is checked before a second of it is voiced. A documentary channel lives on its facts, and a model writes confident, plausible, wrong sentences: a date a day off, the famous version of a story instead of the true one, a quote that was never said in those words. Once the voice is recorded, every fix costs a new voiceover and a new render. """ PART_WORDS = 550 WORKERS = int(os.environ.get("FACTCHECK_WORKERS", "6")) # 1. the narration is cut into parts of about 550 words # 2. every part is checked at the same time by its own model call with web search # 3. each call returns only real problems, with the exact words to replace # 4. fixes are applied by exact replacement; what cannot be matched is reported def check_script(path, language): parts = split_narration(path, PART_WORDS) with ThreadPoolExecutor(max_workers=WORKERS) as pool: reports = pool.map(check_part, parts) return apply_exact(path, merge(reports)) # original kept alongside
Excerpt from the internal codebase. Model names, prompts and provider keys are omitted.
| Stage | Decision the model makes | What it returns |
|---|---|---|
| research | Which claims in this passage carry the story, and what is the source behind each | Claims with sources |
| factcheck | Is this statement false, is this quote word for word, is this certainty earned | Exact words to replace, and the fix |
| director | What does the viewer see while this sentence is spoken | A moment assigned to one tool |
| shots | Place, camera, action, time of day, length, and why this shot | 1–3 shots per moment, as JSON |
The pipeline is model-agnostic: each stage calls a leading large language model through one provider interface, so a stage can be moved between providers without touching the rest of the system.
The scenes a history series cannot shoot: a 1920s trading floor, a siege, a factory that no longer stands. Built to the script instead of staged with actors, and costed per shot rather than per shooting day.
Move from a book or an archive to a reviewable rough cut, with the research trail attached to every claim.
Turn a backlist title into a screen adaptation that stays faithful to the text it came from.
Builds and runs the pipeline. Background in video production and automated editing systems.
Owns the Blender side: scene construction, shot rendering and the look of the finished picture.
Owns source material and verification: what the script claims, and what each claim rests on.
An independent, privately held, for-profit team. No outside funding to date.
| When | Milestone | How we will know it worked |
|---|---|---|
| Q4 2026 | Reference film from a single public-domain title, end to end | A watchable cut a producer signs off on |
| Q1 2027 | Private pilot with a first group of production partners | Hours delivered, and notes from their editors |
| Q2 2027 | Producer console: review and change the plan at every stage | Partners run a film without us in the room |
Dates are targets for a pre-seed team, not commitments.
Frameroo takes a book or other source text and produces a documentary cut: it researches and fact-checks the material, writes the narration, plans every shot, builds the scenes that cannot be filmed in Blender, and assembles a timeline for a human editor to finish.
Yes. Every stage writes a plan a producer can read and change, and the final cut goes through a human edit before release. Nothing publishes itself.
The script is split into parts of about 550 words and each part is checked in parallel by a model call with web search. Only real problems are returned, with the exact words to replace. Fixes are applied by exact match and the report ships with the script.
Broadcasters and streamers commissioning history and factual series, documentary studios, and publishers or education companies adapting existing titles for screen.
We are selecting a small number of broadcasters and studios for the first pilots. Tell us what you would put through it.
Or write to email.