A working system, not a demo page

Videos made in stages.
Checked at every stage.

FrameForge is a complete video production system that runs on one rule: every stage produces one artifact, every artifact passes a gate, and no video is called done until a self review proves it. The video below was made by the system itself, end to end, at a total cost of $0.

8stages, all gates passed
82 secfinished video, 3.4 MB
$0.00total production cost
65 secfull pipeline run time

Watch the proof: How GPS Finds You

Researched from four named sources, narrated by a local voice, designed cards, a music bed, burned in captions, and a passed self review. Press play.

0:00 How GPS finds you
0:09 Satellites that shout the time
0:23 Timing is the measuring tape
0:38 Spheres, then a point
0:51 Why a fourth satellite?
1:04 It is a clock, not a map

Inspect the real run files yourself: checkpoints - final review - decision log - research brief - script - subtitles - metadata

The run board

Every stage of the GPS run, with the gate it had to pass. The proposal stage is a hard stop: without recorded approval, the run waits there and nothing after it starts. That was tested first, before the approved run.

PASS1. Research6 facts, each sourced and rated, 4 sources, 3 angles
PASS2. Proposal3 concepts, itemised cost $0 of $2 cap, approval recorded
PASS3. Script226 words against a 225 word budget, hook first
PASS4. Scene plan6 scenes, 5 scene types, full coverage
PASS5. Assets6 narration files, 6 cards, music bed, all verified on disk
PASS6. EditTimeline built from measured voice lengths, 36 caption lines
PASS7. ComposeZoom and fades, voice over ducked music, captions burned in
PASS8. PublishVideo, thumbnail, subtitles and metadata exported
82.5 secprobed duration, video and audio streams present
4 frameschecked at 10, 35, 65 and 90 percent: none black, none blank
-16.8 dBmean audio level, peak -1.0 dB, no silence, no clipping
36subtitle entries covering the full video

How it works

The engine is small on purpose. Four Python parts, one pipeline file, and four short director files that tell the agent how to run each stage.

ENGINE

engine.py

Runs the stages in order, enforces every gate, writes checkpoints and the decision log, and stops the run when a gate fails.

CHECKS

validate.py

The gate rules: sourced facts, at least two concepts, script word budget, scene coverage, files on disk, timeline completeness.

TOOLS

tools.py

Local voice narration, designed scene cards, a music bed, FFmpeg compose with zoom and fades, and the self review.

DIRECTORS

directors/

One short file per stage pair, written in plain words: what good looks like, and the common failure to avoid.

Decisions, logged honestly

Every production choice is recorded with what was rejected and why. No silent swaps.

DecisionChosenRejectedWhy
Narration voicePiper, local and freeCloud voice APIsNo key needed, cost stays at zero
Composition engineFFmpeg directRemotion, HyperFramesDesigned cards plus zoom and fades cover this pipeline, with fewer moving parts
Music bedMade in the run itselfLibrary musicNo licence questions, no download needed

What is real, and what is not yet

Real and proven above: the full explainer pipeline, research to publish, with gates that were seen stopping a run without approval, and a visual check that caught and fixed two design faults before delivery, cropped edges from an over strong zoom, and text running under the diagram.

Not yet built: AI generated footage and images, which need paid provider keys, and the documentary and clip pipelines, which exist as a studied design, not as code. FrameForge v0.1 makes designed card explainers. It does not pretend to be more than that.

FrameForge was designed after a full study of the OpenMontage open source project: its pipeline pattern, approval gates and self review ideas were studied in depth, then written fresh here. All FrameForge code and text are original.