Drop-Frame, ARRIRAW and -24 LKFS: The Assistant Editor's Log Day a General AI Model Still Gets Wrong

Eleven hours to log a two-camera day

Nine hundred clips. That is a normal two-camera corporate shoot day — interviews plus b-roll, an Alexa 35 on A camera and an FX9 on B, plus a sound bag running double-system. Your assistant editor spends the next day and a bit offloading cards, checking checksums, matching clip A007C012_260714_R1AB to scene 14 take 3, reconciling the sound reports against the camera reports, reading the script supervisor's lined script for circle takes, building selects bins, and writing the notes that will let the offline editor start Monday without asking questions.

It is not skilled work in the sense that it is hard. It is skilled in the sense that being wrong is expensive: a mislabelled take means the wrong performance ends up in v1, and the client's note comes back as "the read felt off" rather than "you used take 2."

Plenty of shops tried to hand pieces of this to AI in 2024 and 2025 and quietly stopped. The reason was always the same. The tools did not speak the language.

The words a general model gets wrong

Ask a general-purpose model to interpret a camera report and a sound report together and watch where it falls over. MOS is not a typo. Sticks means the slate clap, not tripod legs, unless it means tripod legs, which depends entirely on whether a camera assistant or a grip said it. A circle take is the one the script supervisor marked as printable. The martini is the last shot of the day. Abby Singer is the second to last. Apple boxes come in full, half, quarter and pancake.

Then the technical vocabulary, where being approximately right is the same as being wrong. ProRes 422 HQ and ProRes 4444 XQ are not interchangeable and one of them is your graphics delivery. ARRIRAW, BRAW and R3D are three different camera formats with three different handling requirements. 23.976 and 24 frames per second look identical on a spec sheet and will drift your audio apart over a twelve-minute piece. Drop-frame and non-drop timecode differ by a punctuation mark in the display and by real seconds in the count. Broadcast deliverables in the US have to hit roughly -24 LKFS to satisfy the CALM Act rules that stations enforce, while the same cut for social is mastered nearer -14 LUFS. AAF, OMF, EDL and FCPXML are four different handoff files with different capabilities, and handles are the extra frames on either side of a cut that the colourist needs and someone always forgets.

And the money vocabulary, which is worse. A kit fee is not a rental. A 10-hour day plus an hour for lunch is the standard, overtime starts after the tenth hour, and meal penalties accrue in increments. On an AICP bid form, expendables and grip truck are different numbered line items and putting gaff tape in the wrong one is how an actualisation comes back messy.

flowchart TD
  A["Card offload: 900 clips, two cameras"] --> B["Read camera reports and sound reports"]
  A --> C["Read the lined script and circle takes"]
  A --> D["Read the DIT log and LUT names"]
  B --> E["Match A007C012 to scene 14, take 3"]
  C --> E
  D --> E
  E --> F["Build selects bins and the AAF handoff note"]
  E --> G["Flag drop-frame and non-drop timecode mismatches"]

What "trained on the trade" actually means now

Vertical AI — models tuned on one industry's own material rather than the general internet — became a distinct category this year. The idea is old; what changed is that it stopped requiring a research team. A tuned model for post production is one that has been taught on camera reports, sound reports, lined scripts, deliverable specs, AICP bid forms and edit-suite shorthand, so that it treats MOS as a recording state, reads A007C012 as camera A, roll 7, clip 12, and knows that a request for "the ProRes for broadcast" means a different loudness target than "the ProRes for the website."

A general model can be told all of this, every single time, and it will still lose the thread over a nine-hundred-clip day. A tuned one starts from your vocabulary. The difference shows up not in the impressive demonstration but in the boring middle of the job, which is where all of your money actually goes.

Wednesday morning, the offline handoff

Cards came off Tuesday night. Wednesday at 8am your assistant editor points the tuned tool at three things: the DIT's offload log, the two camera reports as photographed on set, and the script supervisor's lined script scanned at the wrap.

Forty minutes later there is a draft. Clips grouped by scene and take, circle takes marked, sound roll matched to picture, the two clips where the slate was called wrong flagged rather than guessed at, a note that B camera timecode ran non-drop while A ran drop-frame, and a plain-English summary of which interview questions were covered on which card. The assistant editor then does the job that is actually worth their rate: opens the flagged items, watches the questionable takes, fixes the two the tool got wrong, and hands the editor a bin that is right.

Same output as before, minus most of the typing. The editor starts Monday at 9am instead of noon.

The arithmetic: error rate, not hours saved

The honest way to measure this is not "hours saved" — it is how many labels you have to fix, because fixing is what eats the day. Illustrative numbers for a two-camera shoot day; run your own on one real project before believing any of it.

General tool: 108 × 3.5 min = 6.3 hours of correction — which is why shops abandoned it, because that is most of the day you were trying to save. Tuned tool: 27 × 3.5 min = 1.6 hours. Difference: 4.7 hours per shoot day logged. At 60 shoot days a year, that is 282 hours, or about $16,900 of assistant editor time, plus the harder-to-price benefit that the offline editor starts with a bin they trust.

Note what the arithmetic says: at a 12% error rate the technology is worthless here, and at 3% it is transformative. That gap is entirely vocabulary, and it is why this became its own category in 2026.

What the script supervisor and the colourist still own

Performance is not a metadata field. Which take is the good one is a judgement about a person's eyes at the end of a line, and no tuned model will make that call for your editor. Marking circle takes is a script supervisor's job on the day and it should stay that way — the tool reads the marks, it does not make them.

Two more places to keep hands on. First, the colour workflow: LUT names, ACES settings and what the DP actually intended are worth a two-minute phone call and no amount of automated log reading substitutes for it. Second, anything where the paperwork on set was wrong. A tuned tool will faithfully reproduce a mis-slated take, and it should — flagging is honest, guessing is not. If your camera reports are sloppy, fix the reports before you automate reading them.

And do not let a tuned model near client-facing deliverable specs unsupervised. Getting -24 LKFS versus -14 LUFS wrong on a broadcast master is a rejected delivery, and rejections in the last week of a flush-season schedule cost more than the whole year of savings.

Frequently asked questions

Do I need to train something on my own footage?

No. Tuning on this trade's material is what the tool vendor does. What you supply is your own house conventions — your clip naming, your bin structure, your delivery spec sheet — as a short reference document the tool reads every time. That takes an afternoon and it does most of the work.

How is this different from the auto-transcription we already run?

Transcription turns speech into words. This turns a shoot day's paperwork into a correctly organised project. They overlap only at the interview bay, and the transcription tools still mispell product names and technical terms unless you give them a word list — which you should, on every job.

Will this let me cut an assistant editor?

If you have one good AE, no — you get them back for work that is worth more. If you are paying a freelance AE for two days of logging on every job, you will likely buy one day instead of two. That is where the money shows up, and it shows up as fewer freelance days, not as a layoff.

What is the smallest way to test it?

Take one finished project where you already know the truth, run the camera reports and lined script through, and count the wrong labels yourself. One afternoon, one honest number, and you will know whether the 12% or the 3% is your reality.

Start with your own house style guide. One page: how you name clips, how bins are structured, what your standard deliverable specs are, and the twenty terms your shop uses that an outsider would misread. That document is useful whether or not you ever automate anything, and it is the difference between a tool that speaks your language and one that guesses.

The same vocabulary problem shows up on your phone line. When a client calls and says they need a :15 cutdown in ProRes by Tuesday with the logo lockup from the last campaign, whoever answers has to understand that request well enough to write it down correctly. CallSphere builds AI voice and chat agents for business lines and web chat that can be set up with your studio's own terms and services, so after-hours requests get captured accurately, meetings get booked, and new-business enquiries reach your executive producer rather than a voicemail box. It handles the intake, not the edit.