SISYPHOS Join the waitlist
Sisyphos — Editorial Reasoning Engine Launch / 001 · Local-first · Desktop

Your first cut,
already done.

You shot for hours. Sisyphos watches every frame, finds the moments worth keeping, and lines them up as a first cut you can open and change. It decides by measured signals: motion, sound, beat, sharpness, aesthetics. Every pick comes with a reason. Every reject, too. You stay the editor.

AEvery frame watched
BEvery choice explained
CNothing leaves your Mac
01

Sixty seconds of proof

Launch film · 001
Sisyphos — launch film 16:9 · stereo
For anyone who shoots far more than they use.
DocumentaryBroadcastSports & highlightsReality & unscriptedCommercial & brandedConcert & liveCorporateMusic videoEvent & multicamYouTube & creatorsWedding
Footage watched
Every frame
Reasons given
For every keep & cut
Runs on
Your Mac, offline
AI cost
Cents per clip, your own key
Re-run
Same cut, every time
02

Not a black box

Guessing vs. knowing why

Other AI editors guess.

CUT 01 · 00:07 — why? no answer
CUT 02 · 00:19 — why? no answer
CUT 03 · 00:24 — why? no answer

pattern: random · intention: none

Footage in, finished video out — and no way to ask why. They cut by transcript and template, not by what’s on screen. A strong moment with no words gets lost, and when the result is wrong, your only option is to start over.

Sisyphos knows why.

Event0.82
Emotion0.64
Rhythm0.77
Technique0.71

It measures what’s actually in the frame — motion, sound, beat, sharpness, composition — and rates every moment the way an editor would: what happens, how it feels, how it flows, how it’s shot. Ratings with a reason in plain words. A first cut you can question, adjust, and trust.

03

How it sees

Four layers · L1–L4
L1
Perception

Sisyphos measures your footage on your Mac: where the cuts are, how much moves, how loud it gets, where the beat lands, what’s sharp, what looks good. Ten times a second, every clip.

On-device
L2
Evidence graph

Every measurement is stored and searchable. When Sisyphos explains a decision, the answer comes from stored measurements — not from guesses.

SQLite
L3
Judgment

The strongest moments go to the AI judge. It rates them and writes down why, in plain words.

Sparse
L4
Taste

Coming next: Sisyphos learns from what you keep and what you throw out — so the first cut starts looking like your cut.

In development
interest(t) — measured, fused, snapped clip A007 · 00:00–00:36 · preset event/sport
IN 08.480 OUT 12.960 · onset IN 22.840 · shot OUT 28.360 · valley below threshold
Curve — how interesting each second is
Bands — the moments it picked
Marks — cuts placed on real events: a beat, a shot change, a quiet dip
Never — cuts at fixed intervals
04

The ledger

Every candidate · every reject reason · on the record
CandidateWindowStageRankReject reason
b0108.480 – 12.960FINAL1
b0422.840 – 28.360FINAL2
b0209.120 – 13.480dropped·too similar to b01
b0315.200 – 18.040dropped·too little happening
b0530.680 – 33.120dropped·didn’t make the shortlist
b0633.560 – 35.900dropped·no strong second to trim to

Every clip Sisyphos considered is on the record — including the ones it dropped, and why: too similar to a better take, too little happening, didn’t make the shortlist. Ask “why isn’t this moment in my cut?” and you get an answer, not a shrug. And when a clip gets trimmed, it’s trimmed toward its best second — never just cropped to the middle.

05

Same footage, same cut

Run it twice · compare the timelines

Run it today.
Run it next year.
Same cut, to the frame.

Run 001 c6d4b29578e006af9be39c820f51fa6ce74d12be5318255393dbeb1d2f60aea1 sha-256
Run 002 c6d4b29578e006af9be39c820f51fa6ce74d12be5318255393dbeb1d2f60aea1 identical
diff run-001 run-002 → ∅ two runs · one result · verified byte for byte

Sisyphos doesn’t roll dice. The same footage with the same settings produces the same cut — every time, and you can verify it. That also means re-runs are free: anything already analyzed is never paid for or waited on twice.

06

Local-first, end to end

Raw footage → DaVinci Resolve
Input
Raw footage
Your cards, your drive.
L1
Perception
Measured on-device. Footage never leaves the machine.
L2
Evidence graph
A local SQLite index of everything measured.
Fusion
Beat-fusion
Windows snapped to real transitions.
L3
Sparse judgment
Top-K stills only. Bounded. Cached.
Cut
Timeline
Segments with rationale attached.
Deliver
DaVinci Resolve
Pushed straight to a timeline. Or EDL / FCPXML.
Perception: 100% on-device Judgment input: keyframe stills + signals — never raw video Exports: Resolve push · EDL · FCPXML
07

One license, yours

Perpetual · runs on your machine
Sisyphos License
€129 €189 after launch
Launch price
APerpetual license. Pay once, use forever — includes 12 months of updates.
BAfter year one, keep using your version forever.
CBring your own AI key. The judging step uses your Google account: cents per clip, and nothing routed through us.

“Perpetual, because it runs on your machine.”

A subscription for local software is a leash. Sisyphos is a tool: it works on your footage, on your hardware, offline — so you own it, the way you own your lenses.

Footage you’ve already analyzed never costs you again — Sisyphos remembers its answers instead of buying them twice.

In development. Join the waitlist and you hear it first when it ships.

08

Fair questions

The skeptical-editor section
Q1Does my footage leave my machine?

No. Your video files stay on your Mac — all analysis runs locally. Only the final rating step sends a few still images plus the measurements to the AI, and the app technically can’t send more: every outgoing file is checked, and anything that isn’t a still image is blocked before it leaves.

With your own Gemini key, those stills go to your Google account under your data terms — not through our servers. We run no cloud.

Q2Which NLEs does it feed?

DaVinci Resolve gets a one-click push — Sisyphos builds the timeline in your running Resolve Studio via the scripting bridge. Everyone else gets standard FCPXML and EDL (CMX3600) exports. All three link your original files at the right timecode — never proxies.

Q3What hardware do I need?

A Mac. Apple silicon is recommended — perception is compute-heavy and caches aggressively, so the first pass over a card is the slow one and everything after is instant. Windows is on the roadmap; tell us on the waitlist if that’s you.

Q4What does analysis cost to run?

Cents per clip, with a hard cap. Footage you’ve already analyzed is never billed again — Sisyphos remembers. You see every cent in your own Google console.

Q5Is this an AI auto-editor?

No. An auto-editor hands you a finished opinion; Sisyphos hands you an argument — every pick has axes, a rationale, and a ledger row, and every reject has a reason. It assembles the first cut; you stay the editor.