Production evidence tool · local-first
Log the miss.Price the keeper.
A controlled AI video test is only credible when failed requests and rejected footage stay visible. Record every attempt, attach the evidence location and let usable seconds—not submitted duration—drive the cost.
Evidence locker · Local 01
Controlled
test ledger
No seeded results · no paid ranking
Only recorded runs move the numbers.
Log one row for every queued request, provider failure, rejected return and accepted take. The dashboard derives retry rate and cost per usable second from those rows.
Stored only in this browser.
Back up with JSON before clearing browser data.
Media stays in your own rights-cleared archive.
Submitted
0
0 scheduled / queued
Provider fails
0
Returned no usable file
Creative rejects
0
Returned, not accepted
Accepted
0
Takes entering the edit
Usable seconds
0
0s returned
Usable yield
—
Usable ÷ returned
Retry multiplier
—
Submitted ÷ accepted
USD / usable sec
—
Credits tracked separately
Publication airlock
Private evidence in.
Approved sequence out.
A batch becomes exportable only when A, B and C each have an accepted take, evidence, cleared rights, explicit publication approval and one shared sequence-reel URL.
Publication-ready
controlled batches
G-1
A/B/C accepted
G-2
Evidence linked
G-3
Rights cleared
G-4
Release approved
G-5
Sequence reel
Chain of evidence
Attempt register
Recording protocol
Same brief in.
Every outcome out.
A fair comparison keeps the creative brief stable and lets the tools vary. If you change the prompt or reference pack, increment its version instead of silently overwriting the test.
Gate 1
Lock the run
Keep the shot brief, reference pack, prompt version, aspect ratio, duration and review gate fixed before comparing tools.
Gate 2
Log every request
A queued request, provider failure, creative reject and accepted take each needs its own row. Never hide failed attempts inside an average.
Gate 3
Review returned footage
Measure only the seconds that can survive identity, continuity, contact, motion, audio and edit review—not the duration requested.
Gate 4
Attach evidence
Store the clip or screenshot in a rights-cleared archive and place its filename, archive ID or URL in the matching ledger row.
Derived metrics
No hidden denominator.
| Metric | Formula | What it reveals |
|---|---|---|
| Usable yield | usable seconds ÷ returned generated seconds | Shows how much returned footage survives review. |
| Retry multiplier | submitted attempts ÷ accepted takes | Shows how many paid or completed attempts were needed per accepted output. |
| USD per usable second | total recorded USD ÷ usable seconds | Keeps retries and rejects inside the production cost. |
| Credits per usable second | total recorded credits ÷ usable seconds | Preserves app-credit accounting without inventing a dollar conversion. |
Review taxonomy
Name the failure once.
Choose one primary reason for reporting, then use review notes for secondary problems. Stable labels make provider and model comparisons possible later.
F-01
Identity drift
F-02
Hands or object contact
F-03
Wardrobe or prop continuity
F-04
Screen direction or geography
F-05
Motion or physics artifact
F-06
Audio or lip-sync failure
F-07
Text or visual artifact
F-08
Safety, moderation or provider error
FAQ
Before opening the locker
Does ReelPulse upload or publish my test data?
No. This version stores ledger rows in the current browser's local storage. Export JSON for a restorable backup or CSV for analysis. Clearing site data can remove the local copy.
Why are there no sample attempts when the ledger opens?
The empty state is deliberate. ReelPulse does not seed placeholder generations, retry counts or acceptance rates because they could be mistaken for controlled evidence.
Should a provider error count as an attempt?
Record it as a provider failure. The submitted-attempt metric counts completed requests and provider failures, while queued requests remain outside the attempt total until their outcome is known.
Why keep credits and USD separate?
Creator-app credits and API dollars are different billing units. The ledger reports both per usable second but never guesses a conversion rate between them.
Can I attach a video or screenshot?
The ledger stores an evidence label and optional URL, not the media file itself. Keep the original file in an archive you control and only link material you have the right to store and review.
Does creating an A/B/C batch claim that the videos were generated?
No. A locked work order creates three scheduled rows marked queued. They remain outside submitted-attempt metrics until each row is updated with a provider failure, creative reject or accepted return.
When can a test result enter the public comparison?
Only a complete batch can be exported: A, B and C each need an accepted take, evidence link, cleared rights, explicit publication approval and the same final sequence-reel URL. The public benchmark then reads a separately validated data file.