A 10-Second Cat Cartoon Taught Us Why Video Releases Need Evidence Gates

A ten-second cartoon sounds like the simplest possible content project. One office, three cats, an exploding coffee machine, two lines of dialogue. But getting that short safely into a multi-platform publishing queue exposed several engineering problems familiar from much larger systems: provenance, lossy speech recognition, reproducible builds, authentication boundaries and idempotent writes.

This is a concrete production walk-through for SolarClabs Office Cats — The Coffee Break Strikes Back. It's a fictional office comedy, not a business case study, and the numbers below are file and QA facts, not audience or performance claims.

The original AI-animated clip was approximately 10.005 seconds, 720 × 1280, at 24 fps, with H.264 video and AAC audio. It showed our established cast: an orange tabby office worker, a white kitten and a stern black-cat boss. The espresso machine explodes into harmless foam; the boss ends up with an enormous foamy moustache.

We did not replace the original footage with simulated movement or a slideshow. The post-production output was 1080 × 1920, still 24 fps, with an added editorial opening hook, a localized clean-up of an unwanted corner mark, and synchronized subtitles. Scaling changes the pixel dimensions, but it does not magically create native 1080p detail.

We kept the original AAC audio stream rather than creating a new voiceover. That matters: changing the sound and then using the script to generate captions would defeat the point of checking what the characters actually said.

Our intended lines were:

That was creative direction, not reliable subtitle evidence.

The first automated transcription attempt used the Windows offline speech recognizer. It returned a low-confidence, clearly incorrect sentence, with confidence around 0.227. A pipeline that automatically publishes whatever its first speech-to-text engine returns could have burned nonsense into the video.

We instead extracted mono 16 kHz WAV audio from the exact original file and ran an unprompted Whisper base.en speech-recognition pass through Transformers.js. We did not feed the intended dialogue into the model as a phrase hint.

That second pass recognized the actual spoken words and produced approximate word boundaries:

These timestamps let us produce an editable SRT file and burn in the dialogue captions at the right moments. Machine transcription remains a candidate until the actual result is reviewed; a matching script alone is not certification.

General lesson: separate planned_dialogue, detected_dialogue, verified_dialogue and rendered_subtitles in your release data. One boolean called has_captions is insufficient.

Our self-hosted Windows GitHub Actions runner could authenticate to Postiz, but it initially could not download the private Google Drive video. It lacked a usable Drive credential reference.

Making the file publicly accessible would have bypassed that problem, but widened access unnecessarily. Instead, the approved connector transferred the original bytes into private GitHub blobs. A small manifest recorded the ordered blob SHAs, expected byte count and expected full-file SHA-256.

The runner retrieved the chunks with its authorized GitHub token, reassembled them in order, and checked the final SHA-256 before any media processing or scheduling. We used a similar method for a decoded audio-only WAV.

A minimal integrity check in Node.js looks like this:

import { createHash } from 'node:crypto';

const digest = createHash('sha256')
  .update(videoBuffer)
  .digest('hex');

if (digest !== expectedSha256 || videoBuffer.length !== expectedBytes) {
  throw new Error('Exact release media mismatch');
}

This is not a recommendation to store every video in Git objects: Git has its own storage and object-size trade-offs. For a repeated production system, an authenticated private object store with signed, short-lived access may be a better transport. The crucial property is reproducible integrity verification without making the private asset world-readable.

Our rendered master underwent technical checks for exact duration, 9:16 dimensions, H.264/AAC streams, decode errors, original-audio preservation, spoken caption timing and sampled frames across the opening, dialogue, payoff and final expression.

A clip can pass video QA and still be not scheduled. Likewise, a Postiz item in QUEUE is not yet a public, playable post on YouTube, Instagram or any other platform.

We maintain independent states:

SOURCE_VERIFIED
  -> SPOKEN_AUDIO_REVIEWED
  -> SUBTITLES_RENDERED
  -> FINAL_MEDIA_QA_PASSED
  -> POSTIZ_MEDIA_UPLOADED
  -> PROVIDER_QUEUE_READBACK_VERIFIED
  -> PROVIDER_PUBLISHED_VERIFIED

Each transition needs a corresponding artifact: hash, subtitle timings, decode report, provider media ID, queue item ID or published URL. Skipping a stage makes incident investigation much harder.

Our content plan had twelve native platform packages, but not twelve authorized publishing destinations. For this release, seven destinations were eligible for the short video. DEV.to is a separate article destination; it needs relevant Markdown content, not an incorrectly uploaded vertical clip. Other channels have their own permissions and setup gates.

When scheduling, we first query Postiz for posts in the intended date window, match the specific episode and exact integration ID, and only create a missing post. After each creation we query again and require an item in the expected scheduled state, with the exact timestamp.

That is not perfect distributed transaction semantics—the API may fail after committing a write—but it is much safer than an unconditional retry. Treat an ambiguous write as needs readback, not “try POST again.”

The target for this episode is 12 October 2026 at 18:30 BST (17:30 UTC). A queued entry proves scheduling, not delivery, impressions or retention; those require their own later evidence.

For the next Office Cats gag, I would keep a small, stable release manifest containing:

The comedy and the engineering are connected. A strong first second earns attention. Provenance and accurate dialogue preserve trust. A hash-verified, idempotent release process stops an otherwise funny ten-second mistake from being multiplied across every channel.

SolarClabs Office Cats is an original fictional animated comedy series. This article describes the verified production workflow for its coffee-machine episode. It does not claim that queued posts are already public or that short-form reach is guaranteed.