SKIP TO CONTENT ↓
Model · RCM 1EARLY ACCESS · IN TRAINING

RCM 1

Real Cut’s editing model — in training to do what an editor does: take raw footage, decide what matters, and hand back a finished cut with a reason behind every decision. This page is its public record: version, status, benchmarks, methodology, and the model card.

SEC 01 · VERSION

RCM 1 · PRE-RELEASE · IN TRAINING

A point release for every training run. A major version for every new capability. Versions date the past — they never promise the future.

No training run has shipped a version yet. When the first one does, its note lands on the release ledger.

Release notes →

SEC 02 · BENCHMARKS

The battery, before the numbers.

Six arms — one of them a deliberate control — and five axes, pre-registered before any number exists. Measurement cells stay em dashes until the first judged battery completes; the statuses in the table are the only claims.

The RCM 1 benchmark arms and their current status. No results exist yet; every measurement cell is empty by design.
ARMSPEEDACCURACYPRICEINTEGRITY CONTROLREFINEMENT OBEDIENCESTATUS
RCM 1QUEUED — FIRST JUDGED BATTERY AFTER THE CURRENT TRAINING ROUNDS
Fable 5ACCESS PENDING — API KEYS
Opus 4.8ACCESS PENDING — API KEYS
GPT-5.6ACCESS PENDING — VETTED-PARTNER PREVIEW
Human editorsQUEUED — BATTERY PRE-REGISTERED
Wrong-footage controlQUEUED
RCM 1
The system under test — our own editing model, run on the same battery as every other arm. No home-field briefs, no cherry-picked footage.
Fable 5
A frontier general model behind its public API — the strongest general-purpose baseline available, run on the same footage and briefs, never a strawman configuration.
Opus 4.8
The second frontier baseline. Same footage, same briefs, same judging — the comparison is against the general models people actually use.
GPT-5.6
The third frontier baseline, pending vetted-partner preview access. Every baseline runs the identical battery, or its row stays empty.
Human editors
Professional editors, same briefs, same footage. The reference arm — the bar every other row is judged against, not a marketing foil.
Wrong-footage control
This arm is deliberately handed the wrong footage. A system that scores well anyway is measuring nothing — if the control ever scores, the battery is broken and nothing publishes.

NO NUMBER ON THIS PAGE IS INVENTED.

Benchmarks publish with their methodology — including the runs we lose.

Read the methodology →

SEC 03 · METHODOLOGY

How the numbers will be made.

Written down before the first run, so the results cannot quietly choose their own rules. There is one honest benchmark for an editing model: the cut a professional would have made from the same footage.

THE BATTERY

Every arm receives the same held-out raw footage and the same brief — material RCM 1 has never seen in training. Each arm returns a finished cut, and the cuts are judged against professional editors’ work on the same material.

THE AXES

SPEED
Wall-clock time from footage handed over to finished cut returned.
ACCURACY
How close the cut lands to the professional reference: selects, structure, and frame-level cut placement.
PRICE
What the run costs at each arm’s public rates, per finished cut.
INTEGRITY CONTROL
Whether the cut stays inside the footage: no invented frames, no reordering that changes meaning, nothing on screen that did not happen.
REFINEMENT OBEDIENCE
Given a typed correction, whether the next cut does what was asked — and only what was asked.

THE CONTROL

One arm is deliberately handed the wrong footage. A benchmark an arm can score on without the right footage is measuring the judge, not the editing. If the control scores, the battery is broken and nothing publishes.

PUBLICATION

Results publish with this methodology attached, including the runs we lose. Until then, every measurement cell on this site is an em dash — and the pledge under the table binds this whole page: no number here is invented.

SEC 04 · MODEL CARD

RCM 1 · MODEL CARDRCM 1 · PRE-RELEASE · IN TRAINING

MODEL

RCM 1 is an editing model. It reads raw footage and proposes the cut — which take carries, where each cut lands, what stays, what goes. It proposes; you decide: typed corrections and marked moments are the interface, and the next cut answers them.

STATUS

Pre-release, in training. RCM 1 does not serve customer cuts yet — the professional editors it learns from do, and human work is labeled HUMAN EDIT wherever it appears on this site.

TRAINED ON

TRAINED ON: NARRATED EDITS · OWNED + LICENSED FOOTAGE · 0 SCRAPED FRAMES · 0 GENERATED PIXELS

Narrated edits are the core: professional editors editing real projects with a microphone on, explaining every decision as they make it — the footage, the finished cut, and the reasoning that connects them. The footage is owned or licensed, paid for. Public-web video informs measurement only, never training pixels. Nothing scraped.

WHAT IT WILL NOT DO

It is not a video generator. It will not make footage of things that didn’t happen. Its editing pipeline never generates a frame — no AI avatars, no synthetic voices, no fake you. The person on screen is the real one, on their best take.

DATA IN, DATA OUT

It learns from narrated edits, owned and licensed footage, and opt-in edit decisions — never your footage, never silently. Training on customer material is explicit, opt-in, never retroactive; paid tiers train nothing by default. No perpetual license over your content — you grant only what is needed to process footage and return the cut.

SCOPE

Talking-to-camera first. Broader classes — voiceover over footage, your own B-roll, music-led edits — are where training expands. Capability claims arrive with benchmark numbers, not before.

EVALUATION

Pre-registered on this page: six arms — including a wrong-footage control — judged on five axes. The first battery runs after the current training rounds. Results publish with the methodology, including the runs we lose.

THE EDITORS

The taste in RCM 1 belongs to working editors — paid per session, credited by name unless they ask otherwise. The first editor in the corpus is the founder, Žan Bassanese. Credits accrue in public on the provenance ledger.

The provenance ledger →

CONTACT

SEC 05 · THE SURFACES

Where it surfaces.

API documentation →

MCP for agents →

Pricing →

The approach essay →

Release notes →

SEC 06 · THE ASK

EARLY ACCESS · IN TRAININGMake your first short