← Back to motion comparisons

Approved benchmark 01

September 5, 2026. The user approved publication of the reviewed pass-7 benchmark. Five original treatments and the 28-second FEN reuse demonstration are published as an additive chapter on the existing Cloudflare Pages fieldnotes site. Existing chapters and downloads are preserved; navigation links point to this benchmark. This is not the final all-hands presentation.

Source map

Frame indices are zero-based; endpoints are inclusive. Source titles link to the official films. Each comparison contains one numbered source/output pair per encoded frame, not independently synchronized players.

Observations and implementation choices

Opening

Observed: first word f3; second word f6; third word f9; final word's partials f12/f13 and completion f14. The title persists through f61. The complete window replaces it at f62 without a delivered dissolve. Source title ink at f14 spans approximately x358–1560 and y490–616 in 1920 × 1080 coordinates.

Implementation: the same native-frame construction and cut, original wording, Geist at 136 px/600 weight with fitted tracking. The complete original phrase is measured before rendering, centered as a whole, and its left edge stays fixed while it builds. Oversized replacement copy is rejected rather than silently shrinking the treatment. The desktop artwork and archive specimen are original.

Camera

Observed: established wide context, a pointer approach, intermediate blurred magnifications around f95–102, then the selector at f103. The source does not identify an authoring camera, easing curve or shutter setting.

Implementation: a piecewise scale/translation fit with a low-speed tail through menu entry and 25 closely spaced temporal samples during the fast travel. The menu belongs to the screen, not the camera or an annotation layer. Original menu construction, pointer and selected state remain visible. Selector dimming starts at source f127 and reaches gray at f128, independently of the first-row highlight at f129; the menu disappears at f138. The first pass placed the menu too low and too large; the revised geometry is checked in the synchronized comparison rather than excused as different content.

The screen is an original illustrated specimen, not a Poe product demo. No product capability is claimed. A real all-hands demo still requires authentic Poe captures and the separate recording/storyboard workflow.

Collapse

Observed: heterogeneous overlapping outputs, sustained relative travel, rapid contraction and complete replacement at source f2348 (local f248). The first bracket appears at that cut; the second appears at f2356. The pair travels left before opening. Five source letter onsets are f2387, f2389, f2390, f2392, f2396.

Implementation: thirty placements of twelve original SVG portfolio designs; four successive foreground handoffs; a contraction fit; measured bracket-position anchors; and the same letter-onset cadence. The first pass's oversized, off-center collapse was rejected and replaced. Original copy is “Tints,” not source product copy. Replacement title width changes the final bracket spacing without changing the onset frames or source type size.

For the shrinking group, these are thresholded visible-pixel envelopes from the 960-pixel source inspection frames, doubled to presentation coordinates. They are measurements of the encoded image, not recovered layer or camera bounds:

Individual original-art trajectories and overlap order are fitted choices. A matched envelope alone cannot prove that the field's motion matches the source. That remains a playback review item, not an automatic pass.

Pass 5's first two foreground handoffs ran ahead of the source and were rejected. Consecutive source frames establish lime-over-blue through local f163 and blue-over-lime at f164; blue-over-white through f197 and white-over-blue at f198. Pass 6 uses those depth changes and refits the near-field window positions and sizes. For example, the blue window at local f180 is approximately (231,93), 443 × 257 in the 960-pixel source frame. The original blue window uses that envelope. The reviewer's consecutive inspection identifies the later dark card overtaking the white finance/dashboard card at local f217; pass 7 uses that frame for its depth change. The retest does not find the same visible corner reveal in the original, because the corresponding background overlap is absent; this remains a disclosed difference, not a verified visual match. The source promotion is not a depth switch with the separate blue-bar card. Its original workbench replaces the previous overly light object-catalog hero. Foreground pages use a landscape viewport with vertical cropping, not stretched type. Other background placements remain fitted, not recovered source layers.

The added multi-exposure effect in pass 3 was rejected: it produced ghosted headings and borders where the source contraction remains sharply bounded. The revised field uses one opaque sample per frame, not a guessed shutter blur. The footer begins after the clear, with its fade fitted from consecutive source contrast measurements; it does not appear at full opacity on the clear frame.

Layered icon

Observed: outer tile at source f2699, blue core at f2710, white directional glyph at f2719. The layers do not arrive together. Inner growth and rotation continue after glyph appearance; later scale and shading changes preclude calling the source tail a bit-identical freeze. The settled outer tile is roughly 256 square pixels, centered on the 1920 × 1080 frame.

Implementation: those same discrete onsets, independently fitted core growth and foreground rotation, the observed late whole-tile contraction, an original warm four-lobe material and an original F glyph. Layered SVG gradients supply depth. The material and geometry are an original identity, not a copied Atlas icon or a claim to have recovered its renderer. The rejected pass 5 glyph turned roughly twice as far during its early arrival. Pass 6 holds its initial attitude over local f24–25, then turns 39° through f32; this fits the source's within-glyph change, not its absolute arrow orientation. The source core and the original four-lobe material still have different silhouettes, highlights and internal transparency.

Lower third

Observed: both fields switch on at source f117 and are absent at f219, the shot change. Approximate source ink bounds: name (108,943)–(329,979); role (436,943)–(1025,978). There is no backing panel, divider, logo, slide or fade.

Implementation: fictional name/role at the same field anchors, a shared baseline, the same hard-on and cut removal. A width check prevents long replacement names from crossing the role column. The original still-life background changes crop at the exit. It is not a generated presenter or a replacement talking head. Background identity is intentionally different; overlay geometry and contrast must still be judged against the reference.

Baseline audit

The earlier library passed mechanics-oriented checks, not visual fidelity. A fresh-context audit inspected 84 decoded export samples across thirteen recipes and their sequence, plus 48 archived source samples. This was bounded sampling, not a claim to watch the complete research corpus again.

• The title, type and proof framing differed from the selected Atlas example; the arriving browser's continued enlargement was missing.
• A montage substituted 0.8-second palette swaps and a long hold for rapid changes of treatment, angle and crop.
• The camera push omitted observed blur and the closely following selector action.
• The selection study skipped separate action construction, typing and reading.
• Six repeated cards replaced a heterogeneous overlapping moving field.
• A backed two-line artifact panel replaced a one-baseline person identifier.
• Growing bars replaced the observed wordmark departure and mark transformation.

The last item is the separate Blossom treatment; it is not silently declared fixed by this benchmark's Atlas layered product icon. That wordmark remains outside this representative set. Existing published studies remain available; their earlier simplifications are not relabeled as fidelity acceptance.

Independent visual review

This is a publication summary of the fresh-context pass-7 review, not a new reviewer verdict. The reviewer decoded and played all five originals, source excerpts and synchronized comparisons at 1×. Consecutive frames and intermediate states were checked around construction, cuts, occlusion changes and exits. Playback was automated; the reviewer inspected captured images and metadata, not an uninterrupted human-viewed video stream. Not every decoded source frame was manually inspected. Playback completion alone is not the visual verdict. No remaining material P1/P2 mismatch was confirmed in the inspected final set. That bounded judgment does not establish pixel equivalence.

• Collapse handoffs at local 164 and 198 now follow the inspected source depth changes. Its blue-card intermediate fit, contraction, clear at 248, and irregular title onsets at 287/289/290/292/296 were retested.
• Icon glyph arrival at 24 and its first rotation no longer show the rejected excessive early spin. Core arrival at 15 and late contraction at 102–109 remain.
• Camera dimming at 45, gray state at 46, hover at 47, menu removal at 56 and replacement label at 57 match the observed boundaries. Stable open direction is corrected; the transient poses below remain different.
• Opening construction, full-line anchoring, hold through 61 and cut at 62 pass the bounded comparison. Lower-third appearance at 12 and removal at 114 retain simultaneous paired fields without a panel or invented entrance animation.

Remaining differences

These were disclosed in the review preceding the user's publication approval. They remain visible rather than being described as exact matches.

1. Dark-card geometry: at Collapse local 217, the original is approximately 8 px left, 16 px low and 4–5% wider at the 960 × 540 comparison scale.
2. Dark-card corner reveal: source local 216→217 exposes a corner over a separate white card. The original corner is already unobscured; a code-level depth change does not prove the same visible occlusion.
3. Chevron posture: Camera local 21 and 56 differ in transient direction. Stable open direction, menu/hover boundaries and replacement-label timing match.

The gallery links directly to these exact frame pairs. Original text, licensed Geist, artwork, icon material and glyph are intentional identity substitutions, not permission to change measured choreography. The source font, easing, shutter, hidden layer geometry and 3D materials were not recovered.

Render and comparison verification

The canonical project passed clean installation, TypeScript checking, timing tests and all five renders. Output hashes are bound to the source-file and asset hashes in the render manifest. The download ZIPs passed exact membership, CRC and byte-equality checks against the verified files.

Originals are silent 1920 × 1080 H.264/yuv420p MP4s at their native reference cadences. Comparisons are 1920 × 626, with two complete 960 × 540 frames below labels. No source frame is dropped or synthesized. Nominal-CFR packaging differs from the source container timestamps by at most approximately 1.267 ms here. Source URLs, full-file hashes and each decoded PTS are in the gallery's ledgers. Measurements came from 960 × 540 inspection frames; doubled coordinates do not make them full-resolution measurements. The exports report a BT.709 matrix, but not verified color primaries/transfer metadata: these are not certified color-managed broadcast masters.

An earlier comparison build was rejected because independent video input timebases mispaired some cuts. The final builder instead joins numbered raw source frame n and decoded original frame n before encoding a single sequence. It passed 172 decoded-panel checks around cuts and moving intermediate states.

Headless playback can drop displayed frames even when the file decodes fully. The independent review recorded two original Collapse drops and one comparison drop; other reviewed videos reported zero. The final local gallery check recorded zero across its five comparisons and one in FEN; the independent final FEN run recorded zero. These observations are retained, not erased by later runs.

All 687 frame-inspector images use lossless WebP for this publication. Every decoded RGB image is checked against the reviewed PNG; videos and ZIPs are byte-identical to the reviewed deliveries. Publication checks are separate from the source-fidelity verdict.

Fresh-context reuse

A fresh-context agent extracted the frozen original project and used all five families to create FEN / Waterline, a fictional coastal-materials lab. It uses new text, twelve original portfolio designs, its own screen data/mark, a new glyph, and a vector-relief background. No shipped component was edited.

The final demo is 840 frames / 28.000 seconds at 1080p30, silent H.264/yuv420p, with a 2.8-second quiet tail whose first and last decoded frames match exactly. Native treatment seconds are mapped into this 30 fps edit; the 60 fps Collapse is sampled at 30 fps. Additional holds and transitions belong to this original edit, not to the synchronized source comparisons.

The reuse trial passed clean installation, typecheck, shipped tests, 41 public contract assertions, a rendered README example and the complete new sequence. It detected a one-frame floating-point boundary error in its own timing adapter. A regression failed before repair, then 62 timing assertions and typecheck passed. The repaired sequence was rerendered and fully decoded, played and hash-checked: the tile is absent at output 381 and present at 382. All twenty shipped files remain unchanged. Broad installation/contract checks were retained from the unchanged baseline, not falsely claimed rerun for that two-file repair. The main packaging check additionally confirms the editable reuse ZIP matches the final verified working files.

Limits: fixed violet Collapse brackets, small peripheral screen text, visible camera blur, fitted geometry, and the three disclosed canonical differences. This original screen specimen is not a capture of Poe or a claim of real product behavior. No presenter, voice or final all-hands presentation was generated.

Downloads and historical status

• Original asset bundle: five original MP4s, runnable project, manifest, guide and the optional host-specific loopback helper.
• Runnable original project: the reviewed twenty-file project, including licensed Geist and its OFL notice.
• Editable FEN sequence: the unchanged twenty shipped files, eleven authored reuse files and the reuse guide.
• Release file checksums: SHA-256 and size for every staged file except this self-referential manifest.

These ZIPs preserve the exact reviewed bytes. Their pre-approval wording records the earlier review gate; the user's September 5 publication approval supersedes that wording for this release. No code/render was changed to rewrite history. Normal hosts should follow the bundled README. The verification host needed a guarded os.networkInterfaces() errno-97 loopback preload and a working Chromium installation; dependencies were not patched. The optional helper is included in the asset bundle, not required on normal machines.

Preservation and scope

The user explicitly selected the existing Cloudflare Pages fieldnotes site. This release preserves its current deployed file inventory, adds this benchmark chapter and navigation links, and does not replace any existing download or research content. The old benchmark's local preservation check still records seven concurrently changed project inputs out of twenty-six; nineteen retained their hashes. Those edits were left intact and the failed baseline was not rewritten. The full live manifest is the release baseline. Unmodified files retain their hashes; removing only the tagged new navigation and benchmark callout reproduces each modified existing page exactly. Full served-file checks accompany publication.

Full source films, private presentation material, rejected renders, runtime logs, repository paths and dependency/cache directories are not published. Short source excerpts and frame pairs are credited research comparisons only; they are not permitted presentation assets or part of either original-only ZIP. Source footage © OpenAI. Original code/artwork and the licensed font remain separate. This is a benchmark publication, not completed library-wide propagation or storyboard approval for the final presentation.