EP_SPEEDMASTER, the reverse-engineered episode

A 3-scene vertical anime episode, deconstructed so a brand can see exactly what they're paying for. Three shots, one QA gate per shot, one full frame extraction set per shot.

1. The shape of the work

This episode shipped 4 July 2026 as the first end-to-end run of the seed blueprint (this page) that became the luxury blueprint (see the blueprint page). Three scenes, each with three frames extracted at t=0%, t=50%, t=98% of duration. One QA gate per scene. ~3 minute total runtime. One writer, one QA system, one set of kie.ai generation calls, visible below.

Output: 3 MP4 scenes (s1_establish, s2_deepen, s3_exit), 9 frames total (3 per scene), 1 QA JSON file with all 9 vs-seed and vs-hero similarity scores.

Source files: /home/newtazer/Desktop/NCI/2ND/newtazer/products/ep_speedmaster/

Scene 1: s1_establish , 9.2s

Establishes the setting: empty street, rain beginning, character mid-stride. Sets the visual baseline for the rest of the episode.

QA gate passes. vs-seed: 0.325, 0.324, 0.329. vs-hero: 0.317, 0.316, 0.320. Both similarity metrics show stable alignment across all three frames. No regeneration triggered.

Frame	vs-seed	vs-hero	Status
f00 (t=0%)	0.3248	0.3175	PASS
f50 (t=50%)	0.3244	0.3158	PASS
f98 (t=98%)	0.3292	0.3202	PASS

Scene 2: s2_deepen , 7.6s

Deepens the rain and the character's posture. The most dramatic frame in the episode.

QA gate passes, drift detected. vs-seed climbs from 0.335 (entry) to 0.369 (exit), vs-hero from 0.327 to 0.362. This ~10% drift over the 7.6s scene is intentional (the scene's job is to deepen, not stay static), the gate catches catastrophic failures, not stylistic drift.

Frame	vs-seed	vs-hero	Status
f00	0.3352	0.3266	PASS
f50	0.3531	0.3460	PASS (drift within tolerance)
f98	0.3692	0.3617	PASS (drift within tolerance)

Scene 3: s3_exit , 8.0s

The character leaves the frame. Lowest similarity scores of the three scenes, that's correct: exit means departing from the reference, not staying close to it.

QA gate passes (lowest of the three, as expected). vs-seed 0.231 → 0.211, vs-hero 0.230 → 0.213. The lowest scores in the episode are the correct ones, a character walking away SHOULD diverge from the reference. The QA gate is calibrated to flag catastrophic divergence, not stylistic movement.

Frame	vs-seed	vs-hero	Status
f00	0.2314	0.2302	PASS
f50	0.2276	0.2260	PASS
f98	0.2111	0.2130	PASS

The QA gate that ran

Every scene was measured against two reference images using the same gate that runs on every frame in every batch, qa_selector.py (ResNet50 cosine similarity, threshold M16.8, 16.9 calibrated on a reference bank held out from the active training run).

The pass/fail criterion is: both vs-seed and vs-hero stay within the calibrated band for the scene type. Catastrophic divergence triggers an automatic regenerate; stylistic drift within tolerance does not. See the blueprint page for the full gate logic, the regeneration budget, and the calibrated thresholds.

Raw QA JSON

{
 "s1_establish": [
 {"pos": 0.00, "vs_seed": 0.3248, "vs_hero": 0.3175},
 {"pos": 0.50, "vs_seed": 0.3244, "vs_hero": 0.3158},
 {"pos": 0.98, "vs_seed": 0.3292, "vs_hero": 0.3202}
 ],
 "s2_deepen": [
 {"pos": 0.00, "vs_seed": 0.3352, "vs_hero": 0.3266},
 {"pos": 0.50, "vs_seed": 0.3531, "vs_hero": 0.3460},
 {"pos": 0.98, "vs_seed": 0.3692, "vs_hero": 0.3617}
 ],
 "s3_exit": [
 {"pos": 0.00, "vs_seed": 0.2314, "vs_hero": 0.2302},
 {"pos": 0.50, "vs_seed": 0.2276, "vs_hero": 0.2260},
 {"pos": 0.98, "vs_seed": 0.2111, "vs_hero": 0.2130}
 ]
}

What the brand sees vs what the agent does

The brand receives three video files and a single QA report. They do not see the prompts, the regeneration logic, or the threshold tuning. That's the whole point of the gate, the brand signs off on "the character looks like the character", not on the mechanism that ensures it.

For brands that want to see the mechanism, to verify the work is auditable, not just trusted, this page is the entry. The full production system is documented in the blueprint; the higher-end version (the luxury tier that adds a four-ward expert routing panel, an AIDM Thompson-sampling credit allocator, and an EEAI leakage audit) is on the same blueprint at the luxury tier.

All episodes See the blueprint How I make an episode