Method, limits & data

100 distinct source films. Specific examples, not a claim to have reverse-engineered OpenAI's production system.

This public edition contains 100 official OpenAI videos, 1,075 timecoded shot groups, 637 app/browser occurrences, and 100 close motion studies. It expands an earlier 36-film collection with 64 additional films. The study date is September 4, 2026; the library is not a complete channel archive or a list of the latest products.

What was inspected

All 100 films have source metadata, a full-duration visual sampling pass, an authored shot-group timeline, representative stills, framing observations, adaptations, and limits.
The first 36 retain their earlier two-second overview samples and selected closer inspections. The added 64 use four-second full-duration contact sheets. Short shots between samples can be missed.
63 added motion studies were inspected using the nearest decoded source frame to each 125 ms target across the excerpt. Each has an individual observation, choreography, adaptation, limits, and three sampled phase anchors. Their intervals do not overlap the earlier 37 excerpts. A sampled state is not an exact event onset.
12 selected short windows were additionally checked in every consecutive decoded frame. Their expandable frame strips document those checks. One example distinguishes an Atlas cut between source frames at 72.300 and 72.333 seconds from a Codex push with visible intermediate frames. Decreasing displacement supports a qualitative settle, not an exact easing curve.
The earlier 37 retain their stated inspection methods, including the prior 16 half-second studies. They have new native-cadence players, but are not relabeled as newly inspected at 125 ms. All 100 excerpts now preserve consecutive source frames and their presentation timestamps, checked within 2 ms of the local source file. Re-encoding at 960 pixels wide changes compression and resolution, not the intended frame cadence.
The source metadata was checked against the official OpenAI channel identifier, and source file hashes were checked against the local analysis records. The public source index includes the identifiers and hashes, not downloader logs or media-server credentials.
This is visual analysis. No listening pass, soundtrack analysis, frame-exact edit reconstruction, independent performance benchmark, or current feature-availability audit is claimed.

What the counts mean

A film is one distinct official source video, not another excerpt from the same film. A shot group is an editorial beat that may contain several cuts. Groups cover the source timeline, but their boundaries are approximate; this is not an edit decision list.

An app/browser occurrence is a group with identifiable interface content: a complete screen, a staged or cropped composer, a menu, an embedded artifact viewer, or a readable filmed-device insert. A group may also contain a presenter cutaway. Occurrences are not unique features or unique UI components.

An isolated generated image, an output reel, a title, an advertising chat-text overlay, or an unreadable background laptop is not counted as an app-screen example. Films with zero screen examples remain useful references for human storytelling, output presentation, or typography. Filter the shot list by type to see those differences.

A motion study is a short excerpt with its own narrow observation and adaptation. It is not another source film and is not added to the shot-group total.

The 100 motion studies span 54 of the 100 films. The new 63 cover six descriptive families: layered composition, camera and framing, UI state and attention, typography and marks, continuity and cuts, and live action with overlays. Earlier studies remain a separate filter group rather than receiving an unverified new classification. Counts are coverage aids, not proof of complete comprehension.

The 6 original motion experiments are separate teaching diagrams. They do not increase the source-film, shot-group, screen-example, or motion-study counts.

Inspecting motion precisely

Use the motion index to loop an excerpt, play it at 1×, ½×, or ¼×, or step through its decoded frames. Controls use each file's actual frame timestamps, including 24, 30, and 60 fps families where present, rather than assuming every source is 30 fps. The player reports a decoded frame when the browser supplies a video-frame callback; a target frame is the fallback, not a claim of confirmed display. Seeking targets the middle of a frame's display interval to avoid boundary ambiguity.

Clicking a phase image seeks to that sampled state. The expanded consecutive-frame strips cover only their labeled short windows. The full excerpt may contain other motion that was not checked frame by frame. Browser playback and source compression limit what can be seen; this is not optical-flow measurement, camera tracking, or a recovered compositor project.

The first frame-control action fetches the complete short excerpt into a local browser object URL. This avoids relying on server byte-range support for frame seeking. Normal Play still uses the original media URL until a frame-control action is used. The browser copy is not uploaded anywhere and is released when the page is discarded.

The motion lab lets you play, pause, slow, and scrub six neutral diagrams. Their durations, staggering, transforms, and cubic-bezier curves are explicitly chosen demonstration settings. Each links to observed evidence and names where the adaptation differs. Nothing autoplays, including when reduced motion is enabled. Native media controls and all written observations remain available without JavaScript; custom frame controls and the lab require it.

Observation is not a production preset

The guide separates three things:

Observed evidence: a credited frame or excerpt and a description of what is visible at the sampled source time.
Editorial interpretation: why that framing or sequence may help an audience understand the task. The techniques page compares examples; its conclusions are this study's interpretations.
Proposed adaptation: a way to use the idea in another production while retaining the real product's behavior and that presentation's identity.

No inferred duration, crop, font size, animation curve, or UI token is presented as an official OpenAI film specification. Automatic scene-change candidates are not treated as a confirmed cut list. Visible processing labels and quick edits do not establish latency.

How to use the library

Start with the cross-film techniques, then choose a film format. Use the screen index for an app moment and the motion index to inspect a transition. The shot list gives the full narrative context around each example.

Each still can be enlarged and has a link to the corresponding official source moment. Film pages include a “Do not infer” note to distinguish an attractive example from a verified product claim. Filters and pagination keep large indexes readable; with JavaScript disabled, the underlying entries remain available.

Applying this to a paired presentation

These are proposed workflow safeguards, not observations about how the source films were produced.

Treat the supplied talking-head recording—including its actual speech and pauses—as the timing master. Derive the transcript timing from the recording, not script length.
Keep the talking-head video and generated presentation video as two separate assets. The presentation should not duplicate the spoken narration or add music by default.
Start the presentation with a title or logo. A recognizable content transition can cue a manual switch to the OBS playback scene. Record both cue time and the expected human-response offset; do not assume simultaneous playback.
Compare both assets on one shared timeline. If the talking head begins at presentation time offset, its shared ending is offset + talking_head_duration. The presentation ending must precede that ending by the agreed buffer. An opening lead-in and an ending buffer are separate requirements.
Confirm that the actual supplied footage contains the ending wave or pause. Do not invent or alter the speaker's performance to fill a missing buffer.
Plan one consolidated storyboard and demo review. Capture genuine product behavior using approved demo data; do not recreate source-film capabilities that the target product lacks.
Set final aspect ratio, resolution, frame rate, codec, and playback constraints against the supplied assets and meeting setup. The study clips' 960-pixel width and source-frame cadence are research browsing choices, not final export settings.

Source footage and publication boundary

This is independent research, not an official OpenAI publication or an endorsement. Source footage, brand marks, and example data belong to their respective owners. The stills and brief silent excerpts are presented with specific visual commentary and official source links. Their presence here does not grant permission to reuse them in a new presentation.

Full source movies are not hosted here. This standalone edition also excludes private component catalogs, internal product source code, private-site files, local machine details, and raw downloader metadata. For a new production, borrow the editorial idea and create original, faithful footage of the real product.

Lower thirds and screen overlays

The lower-thirds chapter records nine paired presenter-ID examples and one split-field identity treatment in the existing collection. Seven close readings cover 342 consecutive source frames in 15 separate entry, camera-cut and exit windows. The frames distinguish simultaneous entrances, three-frame name/role staggers, cut-based exits and stepped topic callouts.

Discovery used 116 existing live-action samples from 47 films, followed by 322 half-second-target samples across ten candidate intervals. This is a positive inventory, not an exhaustive count of all IDs or speakers across the 100 films. It does not increase the film or motion-study totals. An absent ID in a sampled frame does not establish its absence elsewhere in the film.

Frame timestamps, source hashes, approximate visible-ink measurements and limits are in the evidence JSON and 342-frame index. Four editable HTML recipes are independent adaptations with original copy and no source footage. Their timing defaults and fonts are not official OpenAI templates or production rules.

Research downloads

Complete study data — JSON: all films, methods, shot groups, observations, and adaptations.
100 official sources — JSON: source titles, channel identifiers, dates, links, durations, and reference hashes.
Timecoded shot list — CSV: all 1,075 editorial groups.
App/browser inventory — CSV: the 637 interface occurrences only.
Close motion studies — JSON: excerpt bounds, observations, choreography, phase anchors, native-frame windows, source hashes, decoded timestamps, and linked shot groups.
Original motion lab — JSON: the six teaching diagrams, chosen settings, source references, and limits. Animation implementations are in the site's lab script.
Publication manifest — JSON: the built public files and checksums.

The data uses seconds from the start of each source film. A source URL is provenance, not a claim that the historical interface or advertised service is available unchanged today.

Paired-video acceptance: reuse the OBS worksheet. Require C+Rmax+max(T,L) ≤ P ≤ C+Rmin+min(H,E)−B, plus completed visual proof/reading holds. E is observed usable performance end, not file duration. Unknown calibration or unusable tail blocks readiness. A whole post-presentation wave is a separate observed constraint. Executable packet and checks.