CUTPROOF CONTEXT-FIRST CLIP STUDIO / V1.3

A creator tool that does the work

Keep the context.
Check the source.

From an English recording to captioned clips and a source-bound editing handoff.

Local speech recognitionExplicit reviewReal video exports

Joseph Ayanda · Solo builder · September 2026
Working open-source prototype, not a subscription service.

CUTPROOF PROBLEM / TARGET USERS

Shorter is not automatically faithful

Three errors can survive a polished edit.

01

The caption is wrong.

A negation, number or qualification disappears from the supplied transcript.

Listen and compare.

02

The context is outside.

The excerpt is real, but a related correction sits elsewhere in the source recording.

Inspect beyond the cut.

03

The file has changed.

An editor receives a same-name video that is not the exact media used for the edit plan.

Verify the handoff.

Initial user hypothesis: solo creators and small editing teams making source-faithful interview, tutorial and product-explanation clips.
02 / 08These failure modes are exercised in internal fixtures. Customer prevalence has not been measured.
CUTPROOF SOLUTION / MEANINGFUL AI

Evidence Desk

Ask the audio,
not another prompt.

Whisper-tiny.en processes the actual selected audio in a browser worker. The supplied caption is not passed to the model as its answer.

01 / Generate timestamped captions from a recording, or import SRT, VTT or JSON.

02 / Compare supplied captions with recognized speech. Highlight disagreements.

03 / Review the source, edit explicitly, and keep the decision visible.

Optional model downloads require consent. Media is not uploaded. ASR can be wrong; agreement is not certification.

Actual Evidence Desk output highlighting speech: not in the missing-word test
Actual browser inference on the fictional missing-word fixture.
CUTPROOF PRODUCT / SOURCE-BOUND HANDOFF

From review to deliverable

The edit plan belongs with the right video.

REVIEW

Select a contiguous source range.
Inspect related context.
Approve explicitly.

→
BIND

Compute media SHA-256
and byte count.
Capture the editing state.

→
DELIVER

Export captioned WebM or an edit ZIP with an H.264/AAC renderer.

Source Lock rejects a same-name changed source file
Actual same-name, metadata-changed video rejection.

Changing state invalidates old work.

Editing during hashing aborts the export. Replacing media resets reviews. Restoring matching bytes never restores approval.

The native renderer checks the source lock before output creation. A digest identifies bytes, not truth, authorship or consent.

CUTPROOF TECHNICAL ARCHITECTURE

A browser tool, not a remote inference wrapper

AI proposes. The workflow preserves control.

INPUT

Local recording + captions

Web Audio decoding · SRT/VTT/JSON import · editable timed cues

ANALYZE

Speech worker

Whisper-tiny.en · pinned model revision
Transformers.js 3.8.1 + ONNX Runtime Web

Source retrieval

TF-IDF/diversity clip ranking
BM25 + lexical framing cues for context

DECIDE

Explicit review and stale-state checks

Changed transcript, selection or media invalidates affected approvals and checks.

OUTPUT

Browser rendering

Captioned portrait WebM with source audio

Source-bound edit ZIP

Timed captions · review report · manifest
Python + FFmpeg H.264/AAC route

05 / 08Retrieval is lexical, not semantic contradiction detection. No publishing or machine-control agent is claimed.
CUTPROOF EXECUTED EVIDENCE / KNOWN LIMITS

Reproducible internal checks

A working system, with the misses visible.

153JavaScript tests
125 inherited + 28 new
13Native source-identity
and renderer checks
32Browser workflows
also passed on public HTTPS

The public checks run the real app.

Actual Whisper inference, caption editing, reviewed-only export, source mismatch rejection and decoded video output. Published assets match the verified release.

The boundary detector still misses cases.

10 / 15 risky cuts flagged; 5 missed.
2 / 9 safe cases falsely flagged.

Small internal diagnostic, not representative accuracy. Source Lock does not improve this separate score.

CUTPROOF VALUE / FEASIBILITY / ROADMAP

Make review easier, not invisible

Validate usefulness before claiming impact.

NOW

A complete local workflow

Caption preparation, source-context inspection, explicit review and actual exports in one tool.

English input: up to 10 minutes / 120 MB. Browser encoding is real-time, not frame-exact.

NEXT

Authorized creator pilots

Compare manual editing with CutProof on consented recordings. Measure review time, missed corrections, false alarms and usable exports.

Proposed test, not completed user research.

THEN

Improve the bottleneck

Better word alignment and retrieval; verify desktop-editor imports; test longer recordings only after memory and latency checks.

Commercial hypothesis: paid setup or team workflow support. No revenue claimed.

The value proposition is a shorter path from recording to a reviewable handoff. Measured time savings remain unverified.
07 / 08No customers, market-size figures, clinical outcomes or carbon savings are invented.
CUTPROOF REVIEW / SOURCE / PROVENANCE

Try the failure cases, not just the happy path

One recording. Visible evidence.
A handoff you can recheck.

01 / WORKING APPLICATIONOpen CutProof 1.302 / RECORDED DEMONSTRATIONWatch the 3:13 walkthrough03 / CODE AND REPRODUCTIONRead source, setup and test evidence04 / PORTABLE SOURCEGet the tested release ZIP

Joseph Ayanda · Solo builder. Original application created September 6–8, 2026 with substantial AI assistance. Original code: MIT; third-party licenses retained. Demonstration narration: disclosed stock Kokoro synthetic voice. This is the same CutProof project also entered in AI Content Engine, not a separately invented product. No award or cash receipt is claimed.