# Delimiter-before-word typing: R03/R04 comparison

September 15, 2026. Review draft; selection and publication remain with the parent.

## Finding and its limits

**R04 default and R04 adaptation retain an empty-delimiter aperture closer to the source than their respective R03 variants.** This is the spatial preference established before revision names were reconciled. It is not an overall-quality recommendation or a moving-playback winner.

Both revisions preserve the observed order: left delimiter alone, empty pair, then interior letters. The two empty-pair dwell intervals match the source. The native witnesses show narrower empty pairs in R03 default and R03 adaptation; no temporal regression was demonstrated within either same-caption comparison.

The aperture finding is supported by the previously inspected native regions. Its perceptual importance at real-time playback speed was not verified. Caption, glyph, color, and layout differences also limit claims of overall source fidelity. Revision-name reconciliation adds no perception, viewing credit, criterion, scoring weight, or confidence claim.

## Criteria retained unchanged

These criteria were fixed before the comparison movies were inspected and remain unchanged:

1. **Delimiter/content order.** Check both groups independently: left marker alone, closed empty pair, then interior letters. Preserve the distinction between a pair existing before content and punctuation merely appearing somewhere during typing. Measure the two empty-pair dwell intervals and marker-to-word leads rather than accepting a matching final still as proof.
2. **Character sequence and cadence.** Check the prefix, `idea`, connector, and `deployment` onsets. Distinguish discrete character addition from whole-word replacement, fades, or a spatial wipe. Compare actual elapsed times as well as order; report any shifted event and duration instead of hiding it inside an overall length match. Do not infer experienced smoothness from stills.
3. **Spatial coupling during construction.** Track the early leftward displacement, the stable vertical placement and apparent size, the left delimiter's relation to the prefix, and the closing delimiter's advance with word growth. Check both intermediate and late states. Do not impose constant inner clearance, a fixed full-word-width empty box, or permanent horizontal centering of the partial phrase; those are not the observed source behavior.
4. **Glyph form, color, and legibility.** Check heavy proportional black lettering, violet-purple square versus round markers, marker height relative to letters, and the white field. Use native detail for apparent merging, clipping, outline differences, or tiny spacing defects. Separate a projected x-overlap from an actual contour collision. Do not claim an exact font or color match from these references alone.
5. **Composition and placement.** Check the full frame as well as the text band: final horizontal centering, whitespace, no unintended wrapping, crop, extra marks, or size changes. Distinguish intermediate travel from a camera zoom or an unrelated framing shift. The given ink bounds are measurement anchors, not a license to infer unseen master geometry.
6. **Opening, completed hold, and end.** Check the initial blank interval, when the complete statement becomes readable, and how long it remains before the cut. Preserve the absence of an observed exit animation within this clip. Do not mistake decoded pixel variation in a static-looking hold for a new visible event without checking the images.
7. **Evidence and technical limits.** Bind each anonymous movie by hash; probe its actual stream, all retained PTS, and duration. Assess the movie itself, not a role description, builder, key, or still supplied in place of the movie. State the exact indices, scales, and regions viewed. If real moving playback is not actually inspected, leave that claim unmade. A timing disagreement, a mechanism disagreement, and a minor visual rendering difference must be reported separately.

Source fidelity is the comparison target. A source departure and its likely perceptual consequence are separate judgments. There is no preselected expected gap, candidate ranking, numerical pass threshold, or forced count of failures. An unresolved native relationship stays unresolved rather than becoming a confident defect report.

## Empty-pair aperture

The common comparison mask selects decoded RGB pixels where the channel maximum minus minimum is greater than 50 and the channel minimum is below 180. Gaps are differences between projected horizontal ink bounds. They are not contour distances, font advances, or recovered design tokens.

| Movie | Square-pair gap, indices 29–37 | Round-pair gap, indices 78–89 | Pair widths at indices 29 / 78 |
| --- | ---: | ---: | ---: |
| Source reference | 11–12 px | 12–13 px | 61 / 81 px |
| R03 default | 6 px | 5 px | 58 / 75 px |
| R04 default | 12 px | 13 px | 64 / 84 px |
| R03 adaptation | 6–7 px | 5 px | 57 / 75 px |
| R04 adaptation | 12–13 px | 13 px | 64 / 84 px |

**Mask qualification:** the initial source inspection used a purple-specific mask: R > 80, B > 150, G < 160, and B−R > 40. It reported approximately 60 px square-pair width and a 13 px round-pair gap at index 78. The later common mask moves some fringe bounds by one pixel. That qualification remains unchanged: both masks support the distinction between approximately 12–13 px separation and approximately 5–7 px separation. The table is not an assertion that every edge pixel is invariant under threshold or compression changes.

At index 29, PTS 7424, or 0.483333… s, the closing square bracket begins at x=861 in R03 default and x=867 in R04 default; both opening brackets end at 855. The corresponding adaptation positions are x=862 and x=868, with both opening brackets ending at 855. At index 78, PTS 19968, or 1.300000 s, both variants' opening parentheses end at 1099. Their closing parentheses begin at 1104 in R03 and 1112 in R04.

Previously inspected native regions at 29/37 and 78/79/89 show these smaller breaks throughout the R03 empty dwells. R03's round pair looks more like a nearly closed oval, but a break remains visible: it is not a single glyph or a filled shape. Once interior letters arrive at 38 and 90, corresponding marker positions converge within each same-caption comparison, apart from small threshold-dependent fringe variation. The aperture difference is not a delayed word, missing delimiter, or changed dwell duration.

Index 79 had the largest above-threshold whole-frame difference count in each same-caption comparison. Those counts only helped localize the difference. Compression-related differences also occur elsewhere; no claim is made that the empty-pair regions contain the only unequal pixels.

## Timing, placement, and filled-word limits

All six input movies had verified actual 186-frame clocks: 1920 × 1080 H.264 video, reported nominal and average rates 60/1, no audio stream, time base 1/15360, PTS equal to index × 256, and 256-tick frame durations. The last PTS is 47360; the end-exclusive boundary is 47616 ticks, exactly 3.1 s. Revision reconciliation reuses that evidence through verified byte identity; it adds no new playback or stream inspection. Inherited source-master mapping gains no new provenance credit.

All retain five opening white frames, indices 0–4. Corresponding prefix letters arrive at 5, 8, 13, 15, 17, and 20. The square group's left marker appears at 27, closing marker at 29, and first interior letter at 38. The round group's corresponding indices are 74, 78, and 90. Empty square and round dwells are nine frames (150 ms) and twelve frames (200 ms); left-marker-to-letter leads are eleven and sixteen frames, approximately 183.333 and 266.667 ms. Four-letter first words add letters at 38, 43, 48, and 53.

The following already observed departures remain; the aperture preference does not cancel them:

- **Connector:** all four R03/R04 variant movies first show the connector's `t` at index 65, PTS 16640, or 1.083333… s. The source does so at 66, PTS 16896, or 1.100000 s. All reach `to` at 69. This changes the first-word-only interval from thirteen frames to twelve: a local 16.667 ms change, not a global clock shift. Its real-time pacing effect was not observed.
- **Completion:** both defaults reveal the ten letters of `wildflower` at 90, 94, 97, 99, 102, 104, 107, 109, 112, and 116, matching the ten ordinal onsets of source `deployment`. Their completed hold is seventy frames, 1.166667 s. Both adaptations reveal nine-letter `stoneware` at the first nine onsets and complete at 112, holding for seventy-four frames, 1.233333 s. Different word length prevents attributing the four-frame completion difference to an unintended timing defect. No exit animation was observed within these retained intervals.
- **Early travel:** the variants preserve early leftward displacement without a visible overall scale or vertical-placement change. Their thresholded black left edge settles at x=279 from index 77; the source's settles at x=277 from 88. These are pixel-bound measurements, not recovered easing functions or proof that all subpixel transforms have stopped. Different initial glyphs limit causal attribution.
- **Filled-word spacing:** the square groups around `seed` and `clay` are larger than the source group around `idea`. In both defaults, projected clearance before `)` is 97 px at `wildfl` (104) and 94 px at `wildflo` (107), compared with source values 27 px at `deploy` and −3 px at `deploym` under the source-stage mask. The openings are visibly larger, but the words contain different proportional glyphs; this does not establish a particular width-calculation bug. Negative horizontal projection is not a contour collision, and the source itself has variable spacing.
- **Final marker step:** in both defaults, the closing parenthesis has its final measured position at 112 and does not advance when the final `r` appears at 116; the word grows into the remaining space. Source `deployment` and its closing marker both advance at 116. Both adaptations have the whole word and final marker position at 112. These are ordered-position observations, not experienced motion-quality judgments.
- **Composition:** the variants' final horizontal ink envelope is `[279,1615)`, midpoint 947, versus source `[277,1643)`, midpoint 960. It is about 13 native pixels left of the source center, with a different total width. Each remains one line with broad whitespace and no observed crop or wrap. Caption and glyph changes limit attribution solely to a framing implementation.

## Caption and visual identity

| Variant | Completed text | Marker color category |
| --- | --- | --- |
| Source reference | `Go from [idea] to (deployment)` | Violet-purple |
| R03 default and R04 default | `We turn [seed] to (wildflower)` | Purple |
| R03 adaptation and R04 adaptation | `We turn [clay] to (stoneware)` | Teal |

Heavy black proportional lettering, a white field, tall square brackets, and round parentheses are retained. Exact glyph outlines, font family, brand palette, rendering engine, and color identity were not established. Teal is an explicit departure from the source's violet-purple category. The revision/variant names do not establish whether every caption, identity, or layout departure was intended; those remain comparison limits rather than newly diagnosed authoring errors.

The source was recognizable through its byte identity, and captions/colors made same-content groupings recognizable. The sealed comparison therefore did not establish perfect perceptual identity blinding. Names were reconciled only afterward; no visual assessment was rerun.

## Exact retained viewing scope

The six input files represent four revision/variant movies and two byte-identical source instances. Every one of their 1,116 movie-index assignments has new or explicitly reused coverage from the completed visual review:

- **Source reuse:** sixteen chronological source-overview sheets covered all 186 indices in 480 × 270 tiles, reduced 25% per dimension from 1920 × 1080; their canvases were 1440 × 1160. Native-regional inspection covered 51 unique indices in unresampled 1440 × 160 regions, rectangle `[240,460,1680,620)`: 4, 5, 8, 20, 26, 27, 28, 29, 30, 37, 38, 42, 43, 47, 48, 52, 53, 60, 65, 66, 69, 73, 74, 77, 78, 83, 89, 90, 93, 94, 96, 97, 98, 99, 100, 101, 102, 103, 104, 106, 107, 108, 109, 111, 112, 115, 116, 117, 120, 150, and 185. Additional 480 × 140 endpoint regions, rectangle `[1180,470,1660,610)`, covered 99, 104, 107, 109, 112, and 116; these were subsets, not additional unique indices. There were 26 source-stage sheet views. Reusing that baseline for the two identical input files created no additional viewing credit.
- **Variant overviews:** 48 new chronological sheets, twelve per variant, covered all 186 indices per variant: 744 new movie-index observations. Tiles were 480 × 270 Lanczos reductions, 25% per dimension; canvases were 1920 × 1160, row-major. Each final sheet had ten real frames. Gray labels and unused cells were inspection decoration, not movie content.
- **Variant regional details:** twenty new strip sheets, five per variant, contained 120 unresampled 1440 × 160 regions: thirty indices per variant. Their source rectangle was `[240,460,1680,620)`, with exclusive right and bottom bounds, covering 1/9 of each frame. The indices were 26, 27, 28, 29, 37, 38, 43, 48, 53, 64, 65, 66, 73, 74, 77, 78, 79, 89, 90, 99, 102, 104, 107, 109, 111, 112, 115, 116, 120, and 185. The 120 saved crops were viewed within strip sheets, not separately opened 120 times.
- **Tight comparison:** one further 640 × 924 sheet contained ten unresampled 140 × 140 tiles: index 29 crops `[800,470,940,610)` and index 78 crops `[1030,470,1170,610)`. All were subsets of already inspected regions, including two reused source tiles. It added no unique-index credit.
- **Totals and limits:** 69 newly viewed sheets in the variant-review stage, with 744 new overview observations and 120 native-regional checks. This name-reconciliation stage added **zero** image views, decoded frames, stream probes, perception, or viewing credit. No moving playback, full-master, full-native-frame, native-pixel-complete, or audio-listening claim is made. Unresampled artifacts do not establish calibrated one-device-pixel-per-source-pixel display. Computational hashes, masks, and differences add no visual credit.

The completed holds were stable in visible content at that scope, not proven pixel-identical. From index 116 onward, full-frame decoded RGB hashes had 14 distinct values for R03 default, 10 for R04 default, and 8 for each adaptation. Those differences were not treated as visible flicker without perceptual evidence.

The existing aperture-only preference, its mask qualification, and its viewing limits are unchanged. No overall or moving-playback winner is established. The parent retains selection and publication authority.

## Packaging scope

This packet selects the two named current content exports. Earlier comparison variants remain in the sealed research evidence; they are not extra examples or selected-export credits. Original builder snapshots stay dated historical artifacts. Delivery checks, moving perception and native Resolve execution remain separate.