Frame Review: Keep the Lesson

Calm a reading pause without freezing the explanation. Creator-reviewed OpenCV repairs, exact teaching-frame checks, and proof you can keep.

Comment

Calm the blink. Keep the changing explanation.

A teacher has a finished screen recording and no editable project. An idle caret keeps blinking during a reading pause, beside a package download readout that is still changing. Holding the whole frame can stop both. Re-recording is another substantial task.

Frame Review lets the creator propose one small hold and name what the lesson must keep doing. OpenCV finds a supported detail and interval. The creator inspects both observed states, marks teaching areas, checks a source-bound plan and explicitly generates a comparison. The actual encoded result must preserve the teaching timeline, prepared audio and timing, and pass cut-point checks. The original stays recoverable.

This is a tool for selected 1–12-second lesson excerpts, for creators with only the recording. Its contribution is a repair constrained by the explanation around it, with evidence the creator can keep after the job expires.

What the selected real lessons show

Across three previously selected real 12-second excerpts, whole-excerpt selected-region events fall from 37 to 21, including entry and return. Every declared teaching RGB frame and ordered sequence remains exact on each prepared 360-frame timeline. Prepared audio/timing and boundary checks pass.

Three selected real lesson regressions: 37 to 21 measured changes, while declared teaching content stays exact

Previously selected real lesson Whole excerpt: source / local edit Protected teaching frames
French upload assistant 8 / 2 All exact
Joomla installation 16 / 10 All exact
Gentoo package progress 13 / 9 All exact

Events mean adjacent selected-ROI mean RGB steps of at least eight codes. This includes cut points and codec variation; it does not measure attention, learning or preference. These are development cases, not held-out detector recall. Full vector figure.

In Gentoo, the protected download readout keeps all 32 changing adjacent pairs inside the hold, while the selected caret's interval activity falls 4 to 0. Those sample changes are not a count of semantic progress ticks. An editor may reasonably keep that caret.

We also encoded a controlled whole-frame-hold counterfactual with the same selected interval/reference. Both edits remove the chosen blink, but the whole-frame hold changes 90 of 97 French teaching frames and 90 of 96 Joomla teaching frames; the local edit changes zero declared teaching frames in those cases. This historical V12 comparison is diagnostic, never an app-accepted output. Current regressions confirm its source/local media stay identical.

Choose the moment in lesson time

Entering 1:02 or 62 fills the same 62-second start. Enter and Use timestamp never upload or approve. The creator sees the complete 1:02–1:14 range before explicitly analyzing. Invalid 1:60 preserves the previous range. Start and length stay aligned on desktop and a 375px browser viewport.

Historical V21 timestamp setup in an actual 375px Brave viewport

Historical V21 original Gentoo file selected; the same timestamp controls remain in V22. No analysis or approval is sent by timestamp entry. This is browser QA, not a participant or physical-phone study.

A sixteen-region refusal also has a concrete next step: retain the original/prepared links, Set up a 4-second excerpt, inspect the range and explicitly Analyze again. Setup keeps the selected file/start and failed-job links, clears plan/consent and sends no request or approval. A shorter range may still be refused.

Historical V21 Brave testing uses the real SMW DOCKER lesson at 18s: the 12-second excerpt hits the limit; explicit four-second retry reaches review with zero proposals, then stays unchanged. This recovers the review route, not a newly repaired lesson. Historical V19 in-app-browser checks retain both real jEdit 95/176s refusal/source-save/retry flows and their independent source/preparation/frame-count verification.

Check the evidence pack before saving

V22 adds Receive and check pack. The creator sees transfer progress, then receives a complete bounded ZIP bound to the current job, source and reviewed recipe. The client checks archive structure and SHA-256 of every manifest file before revealing Save checked ZIP. Truncated, corrupt, mismatched or cancelled transfers withhold that link. Individual downloads remain available.

Actual Brave received the untouched Gentoo original's 19.2 MiB pack and verified all 44 file hashes, plus manifest.json. A prepared repeat verified 42 file hashes. The UI explicitly says this is not a saved file: browser-memory transfer, API download and disk-save completion are separate. The manifest does not independently authenticate the server. Both current prepared and untouched-original Brave pack disk saves are independently verified below.

The Gentoo guide now covers the full 364-pixel changing progress readout without the earlier manual width correction. It still requires inspecting the teaching content, checking a fresh plan and explicit consent. The current same-frame progress crop loads at 200% and all 19 displayed teaching-image instances load. A 375x812 responsive viewport has no horizontal page overflow; this is not a physical-phone test.

Inspect the words at the same frame

The teaching comparison shows prepared source and encoded edit in one fixed-scale view. Switch manually at the same frame, inspect at 100% or 200%, and pan with buttons or arrow keys. The last visible pan survives switching while a replacement image loads. Teaching proof cards use the full row so command text has more room.

Historical V18 Joomla teaching comparison: same frame and scale, manual switching

Magnification displays prepared pixels; it does not recover labels lost during preparation. Failed images, mismatched URLs/dimensions or changed results withhold the requested view. Extra crops load only when opened. Review remains static, with no automatic alternation.

Try the creator workflow

  1. Choose the credited package-progress lesson, or upload an authorized recording and enter a lesson timestamp. Inspect the smallest prepared label.
  2. Compare both observed states. Useful typing, gestures, position indicators and progress changes should stay.
  3. Load or mark teaching areas, adjust their rectangles/names, check a fresh plan, then explicitly tick consent and generate a comparison.
  4. Inspect teaching frames and cut points. Receive and check pack, then use Save checked ZIP and confirm the actual saved file. Extract it before expiry; open review.html for embedded static proofs without a server or video playback.

Generating a comparison does not record acceptance for use. The owned workflow explains the next step and binds source, prediction, recipe and checks; it is a current snapshot, not an event history.

Test more than the positive examples

The newest screen adds 14 fixed windows from two additional original lessons, Docker and BigBlueButton. Nine reach review, verify original/prepared recovery and withhold edits; five refuse the region cap and are not recovered by that screen harness. Nine rapid-change proposals in three windows concern instructional UI changes or an uncertain logo edge. Agent contact-sheet inspection is not human semantic ground truth. Zero newly useful repairs qualified.

Together the retained cohorts cover 70 exploratory windows, nine newly selected lesson titles and one repeated bootcamp original. There are 17 proposals across seven windows and seven region-cap refusals. Windows overlap, creators are shared and cohorts selected sequentially; they are not independent trials or people. Earlier drawing transformations, code/status changes and jEdit failures remain. All selection plans/outcomes are retained. Four-second follow-ups are recovery tests, not additional discoveries.

OpenCV 5 and AWS perform the actual work

OpenCV 5 temporal differences and connected components identify supported stationary two-state details. A deterministic source-matched policy selects an observed brightness reference and bounded interval. The creator decides whether changing that activity is appropriate. Editable example guides never approve a plan.

The encoder holds selected brightness samples while keeping every original chroma sample on its timeline. Native mapping, decoded RGB outside the actual caret box, every protected teaching frame/ordered sequence, timestamps, prepared audio packets/skip metadata and decoded PCM must stay exact. Encoded entry, return and clip-loop changes cannot worsen. Failed gates withhold edited output.

AWS ECS/Fargate default-lesson-rescue:27 runs preparation, analysis, approval-bound encoding and artifacts in one 1-vCPU/2-GiB task, one worker/eight jobs. Owner sessions, CSRF/origin checks and processing bounds remain. Jobs expire after 24 hours; a restart can clear them sooner. CodeBuild qualifies the pinned ECR image; temporary scoped qualification access is removed afterward. The repair worker has no application AWS task role. The separate human pilot uses an HTTP API, Lambda and private S3.

Verify the current release

  • All 14 Linux stages, 137 backend / 139 studio tests (one platform skip), and 76 focused UI/compatibility tests pass. Eleven additional pure Node checks exercise the exact shipped ZIP verifier with valid and malformed/wrong-binding archives; these are separate from browser saves.
  • Three current untouched-original native/export and fresh HTTPS API regressions pass. Every member of the 51/45/45-file packs verifies, including self-contained offline proof pages. All 108 media/PNG members match V19 bytes. Empty plans/stale recipes refuse; unowned jobs/artifacts return 404.
  • Actual V22 Brave completes original Gentoo timestamp selection, explicit analysis, corrected teaching guide, fresh plan and generation, then complete client-side ZIP/hash verification. A prepared repeat completes same-frame 200% progress inspection. Historical V21 SMW and V19 jEdit source-save/retry evidence stays dated. Current responsive testing is separate from physical-phone qualification.
  • Both current Brave pack saves pass. Genuine completed disk files independently verify 42 manifest-file hashes in the 2.7 MiB prepared pack and 44 in the 19.2 MiB untouched-original pack, plus their ZIP integrity and whole-ZIP hashes. The original pack includes the exact 17,268,947-byte source upload and media/PNG bytes identical to native qualification. The full original upload -> timestamp -> analysis -> protected plan -> explicit generation -> checked transfer -> saved file route completes. Download-event observers still time out; independently checked disk bytes are the stronger evidence. These are selected Gentoo workflows, not every browser, setting or lesson. Task-owned Downloads duplicates are removed only after verified evidence copies are preserved.
  • One V22 API pack read was truncated after 10,792,936 bytes; its failure remains. Fresh affected-case downloads pass complete member checks. A premature CodeBuild launch before source staging failed before media checks; the staged-source retry passes without expanded ListBucket permission.
  • The first V20 timestamp layout wrapped desktop fields onto different rows; V21 fixes it. A local source-video-preview experiment failed browser codecs and two frozen static contracts, and was restored before release. No player, relaxed CSP or weakened gate ships. Earlier platform/color/boundary/image failures remain.

Current four-page report · Architecture, cases and reproduction · Evidence and retained limitations · Earlier V21 criterion map

Runtime manifest 104e5c0fd19b5bd26eef5f1fac78ed70a488edab1062eff09c2e14761dd5c210, 119 files; pinned image sha256:5863c7ed997b4a241dd3abb5964095c83e4e2953beae793c77ce9a7af095c145. V22 layers interface changes over the pinned V21/V20 media runtime; detector, encoder and thresholds stay unchanged. After retaining native/API packs and browser observations, a clean activation of the identical tested image frees the demonstration slots. Final identity/health are verified separately from the preceding full qualification. The human pilot is untouched. Startup preflight does not establish general cold-start reliability.

Current source: V22 snapshot ZIP. Required judge attachment: Frame-Review-Verified-Original-Creator-Judge-Package.zip, 34,984,591 bytes, SHA-256 7463d1ec8e8e420ced643159611e8e1b66c8c25fb22650b0de78e708525125e5. Its source and report match the pinned release. The GitHub-generated archive has identical file contents with different compression; the required attached ZIP is below 35 MB. Attachment save/readback is verified separately from server file-byte verification.

Measure the viewer experience

The prospective reading/preference protocol and analysis were published before collection. Adults manually compare three silent source/edit lesson pairs in counterbalanced order; the first reading answer locks before seeing the second version. Wrong readings, skips, negative preferences, ties and neither choices remain. The primary descriptive result counts people with an edit majority, source majority or no majority; team responses stay separate.

Authenticated export 20261026T194736Z: zero completed valid human responses, zero independent/zero team; two QA records excluded. One invitation start is recorded. Interest/start does not verify personhood, enrollment or benefit. Target: 8–12 volunteers; 96 invitation slots are capacity. Fixed close: October 27 at 06:00 UTC / 07:00 Lagos. A small convenience pilot cannot establish learning gain or population efficacy.

Adults who watch or make screen-recorded lessons can reply “interested” in the competition discussion for an individual 3–4-minute invitation. No personal contact details, mic or camera needed. Skip/stop freely; negative feedback is useful. Recruitment stays on Devpost and asks for honest feedback, not votes.

Sources and scope

Expanded manifests credit Adam Coster, Indrajit Das, GitHub, Andreas Kalt, Info-farmer, Otautz and Kenneth Francis Rodrigues. Audacity's conflicting page-license wording stays in its records. No defect was added to these lessons. Original hashes/preparation/adaptations are retained. Listed metadata is not independent rights clearance or endorsement.

Originals up to 64 MiB; selected SDR excerpts prepare to 640×360 / 30fps, with mono AAC when audio exists. Preparation can reduce small-label legibility. Exact edit checks reference prepared bytes, with originals separate. No full-lesson splice-back, original-resolution recovery or moving semantic tracking. Viewer preference, time saved, attention, comprehension, accessibility, adoption and safety are unmeasured. Static proofs show selected moments; encoded playback is not universally supported. No unsupported COOL or special-award eligibility assertion.

The new creator-narrated demo uses Joseph's supplied voice at original speed, with actual app captures, source-derived lesson playback, measured results and labelled historical counterfactual/recovery. English captions are lightly edited for clarity. Presentation encoding is separate from native media qualification; no synthetic voice or human result. The narration reflects the earlier recording-time download status; the later prepared and original pack save verification is dated in its description and current evidence. Both current saves now verify; no all-browser or human-outcome claim. The old video remains historical. Direct default-branch pushes; no PR or cron.

Share this project:

Updates

posted an update —

A new Frame Review build is live. You can now compare the prepared source and encoded edit at the same teaching frame, inspect at 100% or 200%, and keep the same scroll position when switching. The current report includes three untouched-original regression checks, a whole-frame-hold comparison, and retained failures.

I am still looking for adult lesson viewers for a 3–4 minute anonymous reading and preference pilot. Negative feedback and skips are welcome; no mic or camera. There are no valid human responses yet, so I am not claiming a learning or attention benefit. If interested, please reply in the official study discussion for an individual invitation.

Try the live app

Log in or sign up for Devpost to join the conversation.

posted an update —

Looking for adult volunteers who watch or make screen-recorded lessons. The anonymous comparison takes about 3–4 minutes: watch short lesson clips, answer what you can read, and give your own preference. You can skip or stop; negative feedback is useful too. No microphone or camera is needed.

Reply “interested” in the official competition discussion for an individual invitation: https://opencv26.devpost.com/forum_topics/45681-3-4-minute-lesson-clip-study-looking-for-adult-volunteers

This is product feedback, not a request for votes. Creator responses are reported separately and software QA never counts as human evidence. Collection closes October 27 at 06:00 UTC.

Log in or sign up for Devpost to join the conversation.

Submission history