# Timeline cursor repair: published review history The actual outputs close the early interior pointer pop. The source remains preferred: later pointer departure at local243–275 and the missing final caret remain open. This is one revised kit, not another example or a moving blind win. The frozen reports below bind successive private documentation checkpoints. After R03, the parent updated only current-status descriptions in the guide and README to reflect these findings. Recipe and movies remain unchanged; seven regression tests still pass. This summary is not a fresh critic acceptance of those prose changes. Final ZIP checks are a separate delivery gate. --- # Timeline cursor R01 — frozen label review Frozen: **2026-09-14T00:02:12.357109+00:00**. ID: `sora-timeline-insertion`. **Static per-native-frame, known-source comparison only. B.mp4 = C.mp4 (tie). Zero quality wins; zero new examples.** Neither output earns source-quality acceptance. The role key, runtime code, parent measurements/images, repair notes, and guide remain unread. ## Verdict **Single largest remaining gap — B.mp4 and C.mp4 leave the timeline too early after insertion.** At f242 the hand remains beside the selected cell in all three. At f243 B/C's arrow appears above-left of the cell; A's arrow stays near the preceding hand location. Through f244-251 A remains beside the descending timeline while B/C already rise into the card. The separation is conspicuous at f252-259, diminishes through f264-275, and is mostly closed around f276-280. This changes the overlap between inserted-result confirmation, the reframe, and the next prompt approach. It is a remaining source mismatch, not an established newly introduced regression. PRIVATE evidence: `details-r01/cursor-001.png` through `cursor-011.png`, especially `cursor-002.png` (f244-247), `cursor-004.png` (f252-255), and `cursor-005.png` (f256-259). Columns A/B/C; four consecutive frames per sheet; common-frame crop x275-549, y210-539. The mismatch survives lossless PNG reinspection and comparison against the cell/ruler/footer anchors; it is not explained by JPEG noise, cursor outline alone, or a whole-camera displacement. Code cause is deferred until unblind. No defensible choreography/composition advantage separates B from C at this display scale. A is the already-known source, not a blind competitor. Confidence: high for the late cursor mismatch; moderate for the output tie; no moving judgment. The previous largest gap—the f26 interior cursor pop—is **not observed** in B/C. They remain cursor-free through f42 and enter from the top around f43-44. That narrow improvement is not quality acceptance. ## Coverage and limits - Visually inspected all **360 A + 360 B + 360 C = 1,080 frame instances**, including every quiet/overlap/end frame: **720 new candidate frames + 360 repeated source frames**, none omitted. Native cadence is 60 fps; frame indices are zero-based. - All 90 consecutive triplet sheets were viewed. Primary usable full-frame coverage is f0-303 in sheets 001-076, plus every f304-359 in all 42 PRIVATE `tail-columns/{A,B,C}-304-307.png` through `{A,B,C}-356-359.png` strips. All 16 PRIVATE cursor crops f240-303 were repeat-viewed. Reinspection adds no unique frames. - Wide tail-sheet displays showed apparent black missing regions. A direct labeled-movie decode and single-column PNG views refuted a movie blackout; all tail frames were re-viewed, not inferred from one diagnostic frame. - Common comparison frames: 960x540. A is native 960x540; B/C are native 1920x1080, downscaled with Lanczos. No temporal resampling. Initial sheets use JPEG q6; detail/tail PNGs avoid added JPEG loss but retain MP4 compression, source smear, and spatial rescaling. The viewer fits large sheets/strips, so literal 1:1 display is not guaranteed. Fine glyph/blur/caret equality and moving smoothness are not certified. - Same frozen source interval **[2880,3240)** and matching excerpt SHA; no new master-interval measurement. Source-first unique coverage remains **1,104 frames across five pieces, including 360 timeline frames**. This repair contributes **0 new source frames, 0 examples, 0 whole-video credit, 0 accepted pieces, 0 wins**. ## Boundary, hold, and regression check | Native frames | Observation | |---|---| | 0-42 | Opening camera and cursor absence: B/C do not repeat the f26 interior cursor pop. No f4 or f7-11 whole-UI camera jump observed. | | 43-81 | Top-edge entry and approach: Visible entry around f43-44 in A/B/C, followed by the approach to the insertion control. | | 82-97 | Hover gate: Hand and black plus-cell appear at f82; descent remains staged before the next cell position. | | 98-119 | Cell reposition and settle: Plus-cell shifts at f98; the approach settles without a second inserted card or extra cut. | | 120-171 | Quiet hover: Sustained hand/plus-cell hold inspected, including stationary background and neighboring card. | | 172-179 | Pre-insertion anticipation: Small source hand-shape/scale changes are not certified equivalent under source blur and unequal native resolution. | | 180-191 | Insertion boundary and hold: Number/rule replaces plus and tall card appears at f180 in all three. No substituted typing/fade event. | | 192-224 | Card slide and overlap: Old/new card overlap preserved; clear selected outline at f208; slide settles by the end of this interval. | | 225-242 | Reframe with hand: Timeline/footer descend while tall cards become the focus; no recurrence of the opening camera discontinuity. | | 243-275 | Arrow departure mismatch: B/C arrow lifts above-left of the selected cell at f243 and leads the source upward through the reframe. Largest retained gap. | | 276-303 | Convergence and final approach: Large vertical separation has mostly closed around f276-280; B/C retain small approach drift. Source caret is visible by f287; fine caret parity is not endorsed. | | 304-359 | Tail hold and last frame: All frames re-viewed in single-column PNGs. Stable card composition and cursor hold; no black wipe, extra insertion, typing, or end cut observed. | ## Bindings | Label | Native size | SHA-256 | |---|---|---| | `A.mp4` | 960x540 | `257597627572418755095e1f5454a907a8a87d5e3f6fb124f3b0e78571276a15` | | `B.mp4` | 1920x1080 | `6acbd1d82a31384cad8ff952d1634392ce58be751e1a4dc5fe3168b803efcbca` | | `C.mp4` | 1920x1080 | `7677826821d627579772f1fab01f02e2eab2cea38ceacf35994eea14a23812de` | `pairs.json` SHA-256: `085db1cd90fdb0b8aa349e6edff1141fa06c9a869432c166fe46eb2862da7fac`. Hashes and cadence checks bind the inputs; they do not establish visual quality. Immutable prior source-first freeze: `../batch02-source/source-freeze.md` and `.json`, frozen 2026-09-13T21:33:49.308797+00:00. JSON SHA-256: `bf5dd8034cd37a0d2c822f7ba8606c0ec18cfe00688fb9caf83a66b1a677ee4c`. `../batch02-source/TIMELINE-R02.md` and `../batch02-source/REPORT-R03.md` remain unchanged; all six prior record hashes are bound in this review's JSON. PRIVATE evidence manifests and their hashes are listed in the JSON. New private worker: **24.77 MiB / 96 MiB**. The 100 GiB run guard passed before writes; the old critic worker remains read-only. No builder edit, movie encode, dependency installation, or deployment. **Await parent unblinding and guide authorization. Source quality remains failed, with 0 quality credit.** --- # Timeline cursor R02 — unblinded repair and guide audit Recorded UTC: **2026-09-14T00:17:08.633405+00:00**. Same ID: `sora-timeline-insertion`. **Hold the current guide for one portable-instruction mismatch. Actual exports bind the cursor repair; this is not quality acceptance.** Earlier source, camera, doc and label freezes remain unchanged. ## Reconciliation - Unblind: **A = source; B = default; C = adaptation. B/C remain tied.** Both raw exports exactly match their frozen label bytes. Each has 360 silent 1920x1080 frames at60fps, the corrected recipe, unchanged respective content and zero added holds. Runtime verification passes for both; all19 bound files match. These unsigned manifests do not attest installed dependency bytes or visual fidelity. - Recipe chain: core `b96642f5…` → camera parent `57ee5407…` → cursor revision `abd94ed2…`. Core-to-camera has exactly16 scalar camera changes. Relative to that parent, **only `approach-arrow.start`, `.x`, `.y` differ**; restoring those three properties restores the entire parent JSON. Source identity/[2880,3240), defaults/adaptation, camera, hover, insertion and later focus-arrow are unchanged. Six parent regression tests pass. - The f26 interior pop is closed in the frozen actual outputs; top entry begins aroundf43–44. **Largest remaining gap: both variants leave the selected timeline too early atf243–275**, particularlyf248–263 (clearestf252–259). This later pointer track is unchanged by the repair. PRIVATE evidence: `details-r01/cursor-002.png`, `cursor-004.png`, `cursor-005.png`. ## Portable-instruction blocker `../../repairs/sora-timeline-cursor-r01/guide.json:3645` and `:3649` still give `approach-arrow` active/effective gates **[26,82)**. The construction step at6698 and checkpoint at6708 repeat the old entry. The actual recipe, README hierarchy and regenerated CSV require **[43,82)**. The bounded reproduction finds **17 contradictory visibility frames, f26–42**: JSON construction says active; recipe/CSV says inactive. This can misdirect a manual Fusion reconstruction. It is not evidence that the checked Remotion movies regressed, and no native Resolve pixel effect was tested. Two local counterchecks do not close it: (1) the correct README supersedes historical command/path examples, not these active construction gates; (2) following CSV active flags avoids the error, but `handoff.py verify` does not compare Fusion gates or checkpoint prose. The six tests cover the recipe/correction, not these fields. Isolated refutation agents were unavailable; the mismatch is locally reproduced, not independently confirmed. Before packaging, reconcile active gate fields, construction step and checkpoint with43; keep26 only as explicit history and rebind the guide hash. **No new render is needed for a guide-only correction.** ## Handoff and package checks - README correctly requires core first, mandatory `active-assets.zip` second, then exact recipe SHA assertion. The inspected core really contains `b96642f5…`, not the repaired recipe. Current active recipe SHA: `abd94ed2003a52e9807145deb560db6af8c0fc85d8174717b03133b4decc3054`. - README/correction record distinguish thresholded retained960px support from inferred1920px anchors, disclose the rejected ROI and edge clipping, and do not call the fit recovered authoring coordinates. Parent measurements were not re-performed. - Handoff validates **422 controls /12 animated**. Streamed CSV: **151,920 data rows**, SHA `9a1c9c67849815d8bc7e3edddb4ae5fdf49d74c379ebad27cab0bfd2719ae8b6`; no full CSV stored. Passing these checks does not validate the stale construction gates. - Required `extraFiles`: **`camera-correction.json`, `cursor-correction.json`, `BUILDER_NOTES.md`**. All exist and are hashed in the JSON. Packaging includes them only when explicitly listed. After the guide fix, parent must unpack the current ZIP, use its own handoff, reproduce its CSV and check these documents/current guide binding. **Final ZIP delivery remains untested here.** Audited guide SHA: `a2ea44e0b8d57bb3d3e9bf130a6b77c127f7340de44989a814692df93e38ee2f`. README SHA: `9c1a73b9bbc07c586672fccf7b68cbc3346774588794cd732f922ec9aa5335e3`. Full recipe, export, manifest, correction and immutable-freeze hashes are in `REPORT-R02.json`. ## Coverage and limits No new visual inspection: carry forward all720 candidate +360 repeated source frames from `LABEL_REVIEW_R01`; source-first coverage remains1104 across five pieces, including360 timeline frames. Known-source static inspection at common960x540; candidates were downscaled from1920x1080. MP4/JPEG compression, viewer fitting and lack of moving/native-resolution judgment remain. Private worker24.77MiB/96MiB; run100GiB guard passed. No repairs, new images, encode, dependency install, deployment or old-worker writes. **Source quality remains failed: 0 wins, 0 accepted pieces, 0 new examples, 0 moving/source-blind or native Resolve credit.** --- # Timeline cursor R03 — exact-delta closure Frozen UTC: **2026-09-14T00:24:07.095362+00:00**. Same ID: `sora-timeline-insertion`. **R02's portable guide-gate blocker is closed. No remaining issue found in this exact delta. Final ZIP delivery and source-quality acceptance are not granted.** - Exactly **seven guide fields** changed: active/effective start43, construction step43, checkpoint43/text, later-gap wording, and mandatory core-then-active recipe instructions. Reverting only those fields reproduces R02's exact guide SHA. Removing only the new regression method reproduces the previous test SHA; the six prior tests are unchanged. - Both guide gates now match the actual recipe **[43,82)** at every local index: **0 contradictory frames**, formerly17. All **7 tests pass**. The new test also rejects the old R02 guide in memory. Actual-runtime handoff passes: **422 controls /12 animated**. - Recipe, raw exports, label copies, render manifests and all19 runtime bindings per export remain byte-identical. README, both correction records and builder notes are unchanged. **R02 and all prior freezes remain immutable.** - Current guide SHA: `c0ba037979e2d154ffbf15ca48723e3f9348bca6f1fe7c95e8820d9e2730e557`. Test SHA: `a2a3cb06088b8ace8afbab409adde3dc4577cad26c4dc72f38beb27966f3faf8`. Recipe remains `abd94ed2003a52e9807145deb560db6af8c0fc85d8174717b03133b4decc3054`. Full bindings and delta pointers are in `REPORT-R03.json`. - **Largest visual gap remains open atf243–275, especiallyf248–263**: both variants leave the selected timeline too early. This is carried from the frozen static review, not newly inspected. The earlier f26 interior pop remains closed in those unchanged exports. - **Final ZIP still pending.** Parent must bind this guide, include `camera-correction.json`, `cursor-correction.json`, `BUILDER_NOTES.md`, unpack core then mandatory active, run the delivered handoff and reproduce its CSV exactly. No new visual frames, source credit, images, renders, repairs or full CSV storage. Prior720 candidate/360 repeated-source frame coverage is unchanged. Private worker24.77MiB/96MiB; run100GiB guard passed. **Source quality remains failed: 0 wins, 0 accepted pieces, 0 new examples, no moving/source-blind or native Resolve credit.** Critic scope complete; stopping here.