Independent visual research / Public editionReviewed 05 Sep 2026
Motion fieldnotesOpenAI film study
About this study

Evidence before implementation

The source audit.

Inclusive publication dates: 2025-09-05 through 2026-09-05. Collection occurred during September 5, 2026; later uploads that day remain unknown.

Exact rolling-year cross-check: 2025-09-05T02:54:45.007714+00:00 inclusive to 2026-09-05T02:54:45.007714+00:00 exclusive. Every observed record has identical membership under the inclusive date filter and the exact rolling UTC year. No boundary record is silently added or excluded.

What was counted, and what was watched

All 764 inventoried records have official-channel-checked watch-page publication timestamps in the newer writing corpus. Of those, 421 are inside the requested date window and 343 are older. The existing 100-film moving-image study contributes 80 in-period films and 20 older references. Listing and date verification do not mean every upload was visually inspected.

Fresh anonymous Videos, Shorts and Live listings add no IDs to that inventory. One previously documented uploads-only item is absent from the three tabs; this is not treated as deletion. Private, unavailable and unlisted uploads absent from public listings are outside the inventory.

The thirteen primary technique windows cover ten existing in-period films, with 404 consecutive frames in bounded subsets. One additional recent tutorial, not in the 100-film selection, was investigated with a four-second full-film overview and a 12-second moving excerpt. Its extra native subset contains 60 frames. There remain 410 in-period uploads without a new close inspection in this task.

All 764 source records, timestamps, collection hashes and recipe decisions / JSON · All authored observations / JSON

The existing twelve recipes

Six have only in-period citations, three have mixed dates and three have none in period. Historical references remain available in the original playbook; their date status is not silently rewritten.

Existing recipePeriod supportReuse decisionOriginal references
title-interfaceallRetain premise/hold/cut, but compare source post-cut scale motion with the fixed-window default.codex-first-look: in-period
atlas-intro: in-period
chapter-resetallRetain task-state contract; exact snapshot identity is our invariant, not a source pixel-match measurement.workspace-risk: in-period
codex-first-look: in-period
context-detailallA direct crop and a traveling camera are different edits and now have separate code and source cut evidence.codex-first-look: in-period
atlas-intro: in-period
pointer-intentallRetain target-first order. No submission or inferred click efficacy is demonstrated.codex-first-look: in-period
atlas-intro: in-period
type-submitnoneBoth references are out of period. New Codex evidence supports focus before typing, not submission in the inspected window.agent-look: out-of-period
canvas-writing: out-of-period
selection-actionmixedKeep the old historical reading. Use the in-period document/table window for a separate scoped adaptation; no completed-edit claim.canvas-writing: out-of-period
atlas-intro: in-period
scroll-landmarkmixedRetain fixed mask/content separation; do not carry the older canvas citation into the rolling-year claim.information-visual: in-period
canvas-prototyping: out-of-period
progress-resultnoneBoth original references are out of period. New in-period prototype evidence supports empty pane before loaded content, not the older chart-build timing.canvas-prototyping: out-of-period
shipping-ide: out-of-period
result-holdmixedKeep reading-time budget explicitly chosen. Neither film samples nor word counts prove universal comprehension.canvas-prototyping: out-of-period
codex-first-look: in-period
object-handoffnoneBoth original references are out of period. The new in-period example preserves request identity across a cut but NOT size or screen coordinates; do not claim pixel-matched continuity.study-mode: out-of-period
agent-slideshows: out-of-period
quiet-endingallRetain quiet-tail structure. The neutral bars do not recreate the source letter deletion or knot geometry.codex-first-look: in-period
atlas-intro: in-period
codex-first-look: in-period
anchored-captionallRetain frame-space ownership, using an artifact identifier instead of adding a presenter to the separate presentation.agent-plugins: in-period
computer-use: in-period

Investigating outside the old selection

Use ChatGPT Work to build custom creative tools · published 2026-09-04T05:27:16-07:00.

The delivered-site hero holds, the page scrolls through two result sections, then the image-editor context replaces it. The native subset is a hero hold; it does not locate the later cut. This corroborates result holds, scrolling and context cuts rather than establishing a new transition mechanism.

One additional recent tutorial cannot characterize all 341 in-period uploads outside the existing 100-film study. No soundtrack, generated-output reliability or product availability is evaluated.

Limits of the supported vocabulary

This is not a channel-wide cut census or recovered house-style specification. Scene scores are candidates, not cut confirmations. No fitted easing, font identification, soundtrack analysis, product latency or output-reliability benchmark is claimed. Fine timing belongs only to the specifically inspected consecutive windows. Sources outside those windows retain their sampling limits.

The quiet-ending example does not recreate branded letter deletion or knot geometry. FocusType does not claim to demonstrate a source submission. ObjectContext preserves semantic identity, not the older playbook’s exact same-position artifact contract. These are explicit differences, not missing claims disguised as matches.

Compressed waits: coverage gap

The inspected windows establish container arrival, visible progress and result reveals, but do not establish the unedited elapsed time of an application task or a recovered speed-up factor. This chapter therefore does not publish a source-backed compressed-wait recipe. If an original production omits waiting, retain genuine before/after states, label the time compression, and document the capture timestamps during storyboard review. That is a production safeguard, not a measured OpenAI technique here.

Reading holds and pacing

Holds are implemented in every recipe, with completion and reading kept separate from motion. Their frame budgets are chosen defaults, not measured comprehension or a universal source-film cadence. The sequence guide explains where to extend a hold, replace a push with a cut, or simplify a beat against the actual recording. The new recent tutorial corroborates result inspection and scroll/cut structure; it does not close the wait-compression evidence gap.

Additional recent-film inspection

OpenAI, Use ChatGPT Work to build custom creative tools, 133–145 s. The held result scrolls through the delivered page before the editing context returns. The separately decoded 133.5–134.5 s consecutive subset covers the initial hold, not the later cut.

Reference frame

Enlarged reference frame

Watch this moment on the official channel ↗