Three unrelated records, looked at more briefly than a full dossier. Each gets its own verdict.
The record
Repair video. Independent shops fix phones and laptop circuit boards by replacing parts a few millimetres across under a microscope, and a camera is often already fitted to it. What gets recorded is the footage, the technician's diagnosis, the parts used, and the outcome on the job ticket: the device starts or it does not.
An illustration, not a real file: a phone will not charge. The technician finds a short circuit, locates a faulty part by the heat it gives off, removes it and tests again. The phone starts. About 30 minutes of video and one yes-or-no result.
Fire predictions. During a bushfire, an analyst draws a map of where the fire front will be in the hours ahead, by hand or with a simulator. The truth arrives later as the mapped edge of the burnt ground. A real example: for New Year's Eve 2019 the New South Wales Rural Fire Service published a spread map for fires around Sydney showing "the communities that may come under threat from embers or fire fronts" 1. Prediction and outcome are both shapes on a map, so the check is how far they overlap.
Film cost reports. Each shooting day a film production issues a plan and a progress report saying what was shot and what caused delays. Each week its accountant reports spending against budget. All of these go to the completion guarantor, an insurer whose promise to the financiers is that the film will be delivered.
An illustration, not a real file: a 30-day shoot is budgeted at A$6 million. By day 12 the unit is a day behind after bad weather and the forecast final cost is A$6.4 million. After a visit from the guarantor, a location is cut and two scenes become one. The film finishes at A$6.1 million. The record shows the plan, each day's slippage, each intervention and the final number.
Why it might matter for training AI
Repair video. Data vendors' pages show robot builders paying for recordings of skilled manual work. No lab was found buying microscope repair video, and no robot was found doing this work.
Fire predictions. An objective check, but the more likely customers are fire-technology companies and insurers.
Film cost reports. A long project with a forecast that can be marked against the final cost. No buyer was found.
Who holds it and who may share it
Repair video. The shop owns its footage. The device's owner owns any personal data that appears when the device starts.
Fire predictions. State fire agencies hold them. They are public bodies. Whether they keep the predictions they issue was not established.
Film cost reports. Film Finances, a guarantor, says its archive "contains production material relating to over 600 films that the company guaranteed before 1980" 2, and that it welcomes ideas for using it. A historian who worked in the archive describes "budgets, scripts, financing and distribution contracts, production schedules, progress and cost reports" 3. Its Australasian arm says: "We are emailed each day's Call Sheet and Daily Progress Report. We are copied on weekly Cost Reports" 4.
The production company owns the reports and the guarantor keeps copies. Whether the guarantor may license them for training AI was not established.
Who is already there
Repair video. Nobody was found pairing microscope video with the job outcome.
Fire predictions. No collected set of analysts' predictions against outcomes was found. A 2020 review says checking of both hand-drawn and simulated predictions "is not systematic" 5.
Film cost reports. Film historians have used the archive. Nobody else was found using guarantors' files: no dataset, no test, no AI company, no data broker.
What would have to be stripped
Repair video. Lock screens, notifications, photos and serial numbers that show when a device starts.
Fire predictions. Nothing personal. The difficulty is institutional: a scored archive would show where predictions fell short on days when people died.
Film cost reports. Cast and crew names, pay, injuries and notes on conduct. For the oldest files the sensitivity has faded; for current productions it is acute. Scripts in the files belong to others.
What stands in the way
Repair video
- Similar video is already collected and sold at low cost overseas.
- Robots do not do this work, and so no lab is known to want the footage.
Fire predictions
- Timed maps of the true fire edge exist for few fires.
- Agencies have reason to refuse.
- No lab buyer was found.
- Hand-drawn prediction is giving way to simulators.
Film cost reports
- No buyer for estimate-against-actual project records was found.
- The guarantor may refuse, or may lack the right to grant a licence.
- The open archive is decades old and on paper.