Ritu shoots and her brother Ankit handles sound and the laptop. They own one mirrorless camera, one lavalier into a phone, and two 64 GB cards. The brief is a twelve-minute film plus six short clips for the centre's social feeds. Three days of sessions, four teachers interviewed between them, and one closing address by the centre's founder that happens once.

On the morning of day one the arithmetic is already tight. Sixty-four gigabytes at the camera's 4K bitrate is a little over an hour and a half of recording. A full day of sessions, interviews and cutaways is four to five hours. Both cards will be full by the afternoon.

The laptop is a five-year-old machine with a 512 GB internal drive and one USB port that works reliably. They have a 2 TB portable hard disk. A straight drag-and-drop copy of one full card to the hard disk takes about twelve minutes over that port. Ankit knows, because he read about it the week before, that a drag-and-drop copy is not verified: the operating system writes the bytes and reports success without ever reading them back. Copying with rsync --checksum reads everything twice and takes closer to forty minutes a card.

So the decision arrives at lunchtime on day one, and it is not a technical question. It is a scheduling one.

Option one: verify, and lose the afternoon's first session. Forty minutes a card, twice a day, on the only laptop, which is also the machine Ankit uses to keep an eye on the audio recordings. The sessions do not pause for them. Whatever happens in the hall between two and half past two is not filmed.

Option two: copy fast and format. Twelve minutes a card, no verification, and they are back in the hall before the session restarts. If a file is silently bad they will find out in the edit, three weeks later, when the card has been formatted eight more times and the session cannot be repeated.

Option three: stop formatting. Buy four more 64 GB cards. In Kota in March that is roughly eleven hundred rupees each, so forty-four hundred rupees against a thirty-two thousand rupee fee. The cards stay unformatted until the film is delivered, which means every card is a second copy of itself, and the copy on the hard disk is a third. Nothing needs verifying, because nothing is being destroyed.

Ritu's objection to option three is straightforward and correct: the fee already has no margin in it, the camera is theirs, the drive is theirs, and forty-four hundred rupees is most of what they will clear on the job.

Ankit's objection to option two is that the founder's closing address happens at four o'clock on day three, it lasts nine minutes, and it is the only thing in the brief that cannot be re-created. If that file is the one that comes back as green blocks, there is no film.

They are also aware, as they argue about it on the stairs, that both of them have done option two on every job they have ever taken, and nothing has ever gone wrong.

The question is what the material is worth, not what the process costs. The sessions can be filmed again next month, because the centre runs the programme every term. The founder's address cannot: he is a guest, he is not local, and he will not be in the building again.

That distinction is the one that resolves it, and it resolves it unevenly. The correct answer is not the same for every card on the shoot.

The material is forty-one minutes of interview with Sunita, who has been a health worker in the same three villages for nine years. It is good. She is funny about the first year, precise about what changed, and quiet about a child she could not get to a hospital in time. The camera was locked off in her front room, one angle, mid-shot, from the moment she sat down to the moment she stood up.

Beside it in the folder are two cutaways, both filmed in the last four minutes before the light went: her register of home visits, and a wide of the lane outside.

The radio edit comes down to five minutes forty of speech that holds together end to end with the screen off. That part goes well. The problem is what it looks like.

To get from forty-one minutes to five and a half, the editor has made about thirty removals. Each one is a jump cut: same framing, same chair, Sunita's head in a slightly different position on either side of the join. Two of the thirty are covered by the register shot, and one by the lane, because a two-second cutaway cannot be used more than about once every ninety seconds without drawing attention to itself.

The NGO's communications lead watches the rough cut and says it looks cheap. She is not wrong, and she is not able to say why. What she is reacting to is that Sunita moves quite a lot while she talks, so the jumps are large, and that after twenty of them the front room has stopped existing as a place.

The editor lays out three options and what each one costs.

Leave the jump cuts and punch in on alternate joins. The source is 4K and the delivery is 1080p, so scaling the incoming clip to 115 per cent and reframing from centred to a third is free in quality. Cost: about two hours. It converts roughly half the jumps into something that reads as a second camera. It does not fix the other half, and it does not give the film a single shot of anything other than Sunita's face.

Go back and shoot pickups. Ranchi to the village is two hours each way on a good day. Sunita does not need to be there for most of the list: the register, her hands, the doorway, the lane, the bicycle, the health sub-centre, the two women she visits, the medicines box. That is a shot list of about eleven items, all of them ten seconds each, all of them achievable in ninety minutes on site. Cost: one full day of the editor's time, about eighteen hundred rupees in travel, and a week of delay, in a month where the annual report has a printing deadline behind it.

Use stills and graphics. The NGO has photographs from the last four years. A held photograph with a slow push is a real cutaway and costs nothing. Cost: the photographs were taken by three different people, two of whom no longer work there, and nobody has a written note of who owns them.

The communications lead's instinct is the first option, because it costs nothing and does not move the deadline. The editor's instinct is the second, because he has counted: a six-minute film wants roughly twelve usable cutaways, he has two, and no amount of scaling manufactures the other ten.

What neither of them can do is invent a reverse angle. Nobody filmed the interviewer, and nobody filmed the women Sunita visits. Those shots are not recoverable by any technique in the edit.

The interview is with a junior engineer who has worked on the scheme for six years. Asked whether the remaining villages in the block will be connected, he says this:

"In most cases, yes. If the current funding continues into the next financial year, and if the two contractors who left are replaced, we can extend it to the remaining villages. That is the plan."

Twelve seconds. The film has a four-minute limit set by the slot it is going into, which is a screen in the reception areas of district offices, and the section it belongs to is already thirty seconds over.

The communications officer sends a note with four changes. Three are ordinary: a name spelling, a section that runs long, a request to open on the reservoir rather than the pipes. The fourth asks the editor to use the engineer's answer as "we can extend it to the remaining villages" — four seconds, clean, and it fits the section exactly.

The join is easy. The engineer pauses before "we", and there is a breath after "villages". Cut at the two breaths and the sentence is grammatical, the audio is continuous, and nobody watching could tell. The editor could do it in ninety seconds.

What the cut removes is not filler. It removes two conditions and a hedge. The full answer is a plan contingent on funding and staffing. The short answer is a commitment. A viewer in a district office who has not been connected yet would take one of those as information and the other as a promise, and the engineer did not make a promise.

The editor works this out by applying the test he was taught and has mostly not needed: if the speaker watched this cut, would he be able to point at a proposition in it that he does not hold? The answer is yes. The engineer holds "we can extend it if the funding continues and the contractors are replaced". He does not hold "we can extend it".

The cost of refusing is not abstract. This is the editor's third job for this department and the fourth is under discussion. The communications officer is not acting in bad faith; from where she sits, she has asked for a shorter version of something the engineer said, which is what editing is. She has a slot length and a department that would prefer the film to sound confident. If the editor writes back explaining fabrication and broadcast codes, he is telling a client she has asked him to lie, which is both true and unlikely to be well received.

The cost of agreeing is also not abstract, and it is worth being precise about what it is rather than treating it as a general feeling. The rushes exist. The engineer will see the film, because it plays in the building he works in. And the shape of the sentence — a conditional turned into an assurance about a public service — is exactly the shape that gets repeated back at somebody in a meeting.

The interesting part of the problem is that the editor has three ways to give the client the four seconds she needs, and only one of them involves the cut she asked for.

The recording is unusable as it stands, and everybody agrees on that within the first thirty seconds of playback.

The hall is a rectangular room with a tiled floor, plaster walls and a high ceiling. Six ceiling fans were running. The camera was eleven metres from the lectern. What the camera microphone captured is therefore mostly the room: a low continuous roar from the fans, a long reverberant tail on every word, and the lecturer's voice arriving at roughly the same level as the air conditioning in the corridor.

The department's media assistant, Farhan, measures what he has. The noise floor sits around minus thirty-eight dBFS. The lecturer's speech peaks around minus twenty-six. That is a gap of about twelve decibels between the voice and the room, and a good part of the voice's energy is reflection rather than direct sound.

He tries the obvious things, in the order he was taught. A high-pass at 100 Hz removes some fan rumble and changes very little, because the fans also produce broadband noise well inside the speech range. Six decibels of spectral noise reduction helps. Twelve decibels produces the watery, fluttering artefact where quiet harmonics of the voice fall below the noise estimate and are removed in some frames and kept in others. A de-reverb pass, pushed hard enough to matter, leaves the ends of words clipped and gives the lecturer a thin, gated quality.

Compression makes it worse in a way that surprises him until he thinks about it: turning the loud words down and adding make-up gain raises the fans by exactly as much as it raises the voice.

That leaves three real options, and a deadline: the repository page is meant to go live before the new term.

Publish it with the light cleanup. Six decibels of noise reduction, a high-pass, gentle compression, captions burnt in and a sidecar file. Honest, quick, and hard work to listen to for ninety minutes. Farhan tests it on his phone with a fan running, which is how a student in a hostel will hear it, and loses whole clauses.

Resynthesise the voice. One of the newer speech tools will take the recording, run it through a model and return something close to studio quality. Farhan tries thirty seconds of it and it is genuinely startling. It is also generating audio: it can shift timbre and invent detail that was never recorded. This is a lecture by a named economist that will be cited, quoted and possibly assessed. What comes back is not a recording of what happened in the room; it is a plausible reconstruction of it.

Ask for a re-record. The lecturer has left Pune. The department could ask her to read the lecture again into a phone with a cheap lavalier at her own desk, and the picture could be held as slides and stills. That is thirty minutes of her time, which she may not give, and it produces a different artefact: a reading, not a lecture, with none of the questions from the hall in it.

The decision is not really between three audio processes. It is about what the file is for.

The scene is a two-minute sequence in a consulting room. A doctor explains what the clinic does about follow-up appointments, and a patient — who has consented, on paper, for this one film — describes what changed for her.

Most of it matches without much trouble. The room was lit by a large window on the left and the shots were filmed within forty minutes of each other. The exception is one shot: the patient's answer about the phone call that brought her back, which is the emotional centre of the sequence and the only take of it. It was filmed twenty minutes later, after the sun had gone behind the block opposite, with the overhead tube lights switched on.

On the parade the difference is obvious. The tube light is strongly green in the mid-tones, so the patient's skin sits well off the reference line, and the green is not uniform across the frame: the window behind her still carries the last of the daylight and reads cold. The editor corrects for the green in the skin and the window goes magenta. He corrects for the window and the patient goes green again. There is no single primary correction that fixes both, because there are genuinely two different light sources in the frame.

The technical options are clear and so are their costs.

A qualifier and a tracked window. Draw a soft window around the patient's head and shoulders, track it across the shot, and qualify skin inside it so the correction touches only her face. Then treat the window behind her separately. This is the textbook answer and it works. The complication is the footage: the camera recorded 4:2:0, so the qualifier is building its mask from a quarter of the picture's resolution. Blurring the matte enough to hide the stepped edges means the correction bleeds onto the wall behind her head, and the patient moves, so the window has to track through about forty seconds of small movements. The editor's estimate is an evening, and he has two other scenes to grade.

Accept a compromise primary. Split the difference: leave the skin slightly green and the window slightly magenta, and reduce saturation in both so neither is loud. Twenty minutes. The scene reads as competent, and the one shot the film is built around is the least attractive in it.

Do not let the shots touch. Put a cutaway between the daylight shots and the tube-light shot, so the viewer never sees a direct join. The clinic's own waiting-room signage, the appointment book, a wide of the corridor. Cutting away also means the mismatch is compared across two seconds of something else rather than across a single frame, and the eye is considerably more forgiving of that.

The third option is an editorial answer to a colour problem, and it is the one that keeps getting forgotten by people who are enjoying the colour page.

There is also a decision that is not about this shot at all. The whole project was set up with colour management off, the timeline at Rec.709 gamma 2.4, and every log clip carrying its own input transform. The editor is tempted, standing in front of this mess, to switch the project to managed colour and let it sort the sources out. Doing that would change the starting image under every grade he has already built on the other seven minutes.

The brief is specific and reasonable. The film is about a free eye-screening camp, it will be watched almost entirely on phones in a muted feed, and it must carry Marathi captions, because the audience it is for reads Marathi and a good part of it does not read English.

The cut is finished. The mix is finished. The Marathi translation came back from the corporation's own office, checked, in a document. Sneha pastes the first line into her editor's title tool and it comes out wrong.

Not missing-glyph boxes, which would have been obvious. The letters are all there and they are in the wrong shapes. The conjuncts have broken apart, with the halant visible as a separate stroke between letters. In two words the vowel sign that should be drawn to the left of its consonant is sitting on the right. It reads, to somebody who reads Marathi, as nonsense wearing the right letters.

She tries three fonts, including a Noto face, and gets the same result each time. That is the tell: it is not a font problem. The editor's text engine is not shaping. Complex scripts need a shaping engine that reads the font's OpenType tables and applies its substitution and reordering rules, and this tool either does not use one or uses an old one. She confirms it with a two-character test: कि renders with the vowel sign on the right.

It is four in the afternoon. The options are these.

Ship it in English. The English captions render perfectly and the film goes out on time. It also fails the brief, and it fails the audience the campaign exists for.

Burn the Marathi captions in with FFmpeg. The subtitles filter uses libass, which uses HarfBuzz, which shapes correctly. It is one command with a force_style string, and it re-encodes the video, so it has to run from the master rather than from the delivery file. Sneha has never used FFmpeg. Her partner Vikram has, twice. Neither of them has done it under a deadline, and if the command produces something subtly wrong at eleven at night, neither of them reads Marathi well enough to catch it.

Set every caption as an image. Type the Marathi in Inkscape, which shapes correctly, export each cue as a transparent PNG at four times the size, and place them on the timeline. This definitely works and it is entirely mechanical. The film has thirty-one cues. Sneha estimates four minutes a cue with positioning, so about two hours, plus the risk of a typo in any one of thirty-one separate pieces of artwork.

Sidecar only. Upload the video with a .srt and let the platform's own text engine shape it. This renders correctly, costs nothing, takes ten minutes, and produces captions that do not exist in a muted autoplay preview — which is precisely where this film lives.

The fourth option is the one that looks cheapest and is the one that most directly defeats the purpose of the job. The first is the one the deadline wants. The decision is really about who the film is for, and there is a fifth thing nobody has thought of yet, which is that they could ask.

The shot is eleven seconds long and it is the one that opens the film. A handheld move down the length of the market at six in the morning, the light coming in sideways off the water, boxes being carried across frame. It establishes the place in a way nothing else in the rushes does, and every version of the cut has opened with it.

Four days before picture lock, a man who appears in it for about five of those eleven seconds sends a message. He had signed the release on the day. He now works somewhere that he would rather not be associated with the market, he has thought about it, and he asks not to be in the film.

The director's position is that the legal question and the right answer are not the same question. She has a signed release. She also has somebody who has told her plainly that he does not want to be in it, and a film that will be shown to an audience in the city he lives in.

So the shot has to change. There are four ways, and each costs something different.

Cut the shot. Eleven seconds, gone, and with it the only wide that establishes the market as a place. The next best opening is a close-up of ice being poured, which is a good shot and tells the viewer nothing about where they are. Cost: the opening, and about a day of re-cutting the first ninety seconds around it.

Reframe. The source is 4K and the delivery is 1080p, so there is room to scale and reposition. The man walks through the middle third of the frame from left to right, and the camera is moving with him for part of it. A reframe that keeps him out would have to move continuously, would end up at roughly 190 per cent for two seconds, and would lose the two boxes being carried in the foreground that give the shot its depth. Cost: the shot becomes a different and weaker shot, in about two hours.

Remove him. Cover him with pixels that plausibly belong. The camera moves, so a static patch slides off: the surface behind him has to be tracked, the patch attached to the track, masked with a soft edge and matched for grain. Behind him is a row of crates, a wet floor with reflections, and two other people who cross behind him at second six. That last detail is the expensive one — a patch that has to respect the edge of a moving person is rotoscoping, at somewhere between one and four hours per five seconds of footage. Cost: realistically a day and a half, with no guarantee, four days before lock.

Blur or box him. A tracked solid box or a heavy mosaic. It is reliable and it takes an hour. It also announces itself in the opening shot of a documentary, which tells the audience that something has been hidden in the first eleven seconds of the film.

There is a fifth option she almost misses, which is that the market runs every week and the light at six in the morning is the same light.

The grade took most of a day and it is good. Deep blacks in the warehouse, clean whites on the fabric, skin on the two machinists sitting where it should on the vectorscope. The editor, Deepak, exports an H.264 review copy and sends the link.

The reply comes back in forty minutes: the colour has gone flat, the blacks are grey, and it looks worse than the version they saw last week. Could he put the contrast back.

Deepak opens the same file on his own machine and it looks correct. He opens it in VLC and it looks correct. He opens it in his browser and it is slightly lifted. The client's screenshot, taken on the office laptop, shows a noticeably washed-out image with grey blacks.

Two mechanisms could be doing this and he has to tell them apart, because the fixes are opposite.

The first is a gamma difference. Rec.709 is defined for a display gamma of 2.4 in a dim room. The web standard is approximately 2.2. A grade that is correct at 2.4 looks slightly washed in a browser, and the browser is applying the standard it is supposed to apply. That difference is real and it is small, and it lives in the shadows.

The second is a data-level mismatch, and it produces a large, obvious error rather than a subtle one. A file can be written full range, with 8-bit values running 0 to 255, or video range, with black at 16 and white at 235. If it is written as one and read as the other, everything shifts: video-range read as full crushes, full-range read as video lifts the blacks to grey and dulls the whites. The file carries a flag; some encoders write it wrongly and some players ignore it.

ffprobe on the export returns a blank color_range field. Nothing was tagged. Every player is guessing, and the client's player is guessing differently from his.

So he has a choice, and the shape of it matters more than this one job.

Regrade to match the client's screen. Add contrast until the screenshot looks right. It would take twenty minutes and the client would be happy today. It would also make the film wrong in VLC, wrong on his own machine, wrong on the exporter's website, and wrong on any player that reads the tag correctly. He would be chasing one player for ever, and the next client would have a different one.

Fix the tagging and prove it. Set the colour tag explicitly on the render rather than leaving it automatic, re-export, run ffprobe to confirm the field is populated, and check the result in two different players and a browser. It costs an export and an hour. The client did not ask about tagging and will not care about the explanation.

The second cost is the awkward part. The client's complaint was "put the contrast back", and answering it with a paragraph about transfer functions is a way of being right and being ignored.