AETHEL · ONE MODEL · THREE SCENES · A GATE THAT SAID NO
The campaign looks real. So does the rejection evidence.
A Glove agent invented a fictional fashion brand, cast a recurring adult model, designed an oxidized-copper hero trench, created the identity pack, directed seven videos across two generators, and inspected every MP4. One final scene passed. The three-scene campaign did not—so the sequence delivery tool withheld it.
The clip delivers a high-fidelity execution of the premium editorial brief. The subject's likeness—specifically the facial structure, hair texture, and skin tone—aligns strongly with the provided identity references. The environment correctly captures the 'blue hour' aesthetic with appropriate atmospheric lighting and background bokeh. The camera movement is smooth and continuous, adhering to the 'very slow push-in' requirement. While visually compelling, the surface textures (coat and skin) lean slightly towards a smoother, digital finish rather than the raw realism implied by 'photorealistic' and the reference materials.
Generator
Veo 3.1 Fast
Format
1280×720 · 8s
Audio
none
Generation
$0.64
CAMPAIGN NOT PUBLISHED
A passing clip is not a passing sequence.
The rooftop passed at 88. The studio missed required living camera motion; the city coat changed length and grew an unstable trailing strap. glove_video_flow_deliver rechecked every selected shot and refused to reveal the campaign.
ALL SHOTS OR NONE
The agent created the model before it shot the model
These are not gallery decorations. Three of the generated images became identity inputs to every Veo shot and independent comparison evidence for the Qwen reviewer. The recurring definition also locked face, hair, skin tone, performance, coat color, silhouette, and failure negatives.
Each clip is playable because this page is an audit artifact, not the agent's delivery channel. The cards show the latest identity-aware verdict that controlled whether the assembled campaign could ship.
SCENE 01 · studio-hero62/100 · revise
While the subject fidelity, lighting, and texture quality perfectly match the 'premium' reference images, the clip critically fails to execute the requested 'smooth dolly' camera movement. The video is entirely static from 0.0s to 8.0s, lacking any camera parallax or subject micro-motions (like breathing). This renders it visually indistinguishable from a static photograph rather than a 'living portrait' or video clip.
1 blocking findings
SCENE 02 · city-motion72/100 · revise
The video demonstrates high technical quality with excellent lighting, color grading, and a smooth, professional camera track that fits the 'premium movement shot' brief. The model's facial identity and hair are accurately preserved from the reference. However, the clip fails a critical adherence test regarding the garment silhouette. The reference images clearly depict a structured, mid-thigh/knee-length coat with exposed black pants underneath, whereas the generated video portrays a floor-sweeping, full-length trench. This is a fundamental mismatch of the 'hero trench silhouette.' Additionally, there are distracting temporal artifacts where a stray strap or fabric tail inconsistently appears and disappears throughout the clip.
2 blocking findings
SCENE 03 · rooftop-close88/100 · pass
The clip delivers a high-fidelity execution of the premium editorial brief. The subject's likeness—specifically the facial structure, hair texture, and skin tone—aligns strongly with the provided identity references. The environment correctly captures the 'blue hour' aesthetic with appropriate atmospheric lighting and background bokeh. The camera movement is smooth and continuous, adhering to the 'very slow push-in' requirement. While visually compelling, the surface textures (coat and skin) lean slightly towards a smoother, digital finish rather than the raw realism implied by 'photorealistic' and the reference materials.
0 blocking findings
Why the agent changed models
“Best model” was treated as a capability decision, not a leaderboard slogan. Every switch followed an observed constraint and remained in the evidence ledger.
Layer
Decision
Observed reason
Inspector
Qwen 3.5 Flash
Cheap Chinese multimodal model; returned structured, timestamped frame evidence. Xiaomi MiMo V2.5 accepted the request but did not finish the smoke review in five minutes.
First generator
Seedance 2.0
Chosen for reference continuity. Its provider rejected the synthetic photoreal identity inputs as possible real-person privacy content, so the agent recorded the error and fell back.
Final generator
Veo 3.1 Fast
Accepted three person references with person generation enabled. Live execution revealed that reference-to-video requires eight seconds even though the general catalog advertises 4/6/8.
Delivery
Sequence gate
Every selected shot needs pass ≥ 82, no major/critical issue, and a latest actual-video review. Two failures held the campaign.
The final decision ledger
The reviewer saw chronological frames sampled every 0.5 seconds plus all three identity images. A score is only an index; publication is determined from the decision, threshold, and severity of concrete findings.
While the subject fidelity, lighting, and texture quality perfectly match the 'premium' reference images, the clip critically fails to execute the requested 'smooth dolly' camera movement. The video is entirely static from 0.0s to 8.0s, lacking any camera parallax or subject micro-motions (like breathing). This renders it visually indistinguishable from a static photograph rather than a 'living portrait' or video clip.
WHAT SURVIVED REVIEW
Flawless adherence to subject identity and wardrobe details (copper trench, dark bob haircut).
Consistent, high-quality lighting and color grading matching the mood board.
No distracting visual artifacts or morphing in the stationary frames.
majorAction/Camera Movement
Evidence
Comparison of frames at 0.0s and 8.0s reveals identical framing and subject position; the requested 'smooth dolly' is absent.
Required fix
Implement a slow, steady dolly-in (zoom forward) or slight parallax camera move to introduce depth and meet the brief.
minorTemporal Coherence
Evidence
Subject appears frozen with no signs of life (chest rising, hair settling); lacks the 'living portrait' quality.
Required fix
Ensure generation includes subtle breathing mechanics and fabric relaxation to simulate a live human presence.
Revision prompt produced from the evidence
Create a photorealistic 8-second video clip of a woman in a premium setting. Subject: Same woman from reference (dark wavy bob, copper double-breasted trench coat, hand in pocket). Camera: Execute a slow, smooth dolly-in movement starting from a medium shot and slowly approaching the subject's face over 8 seconds. Action: Subject holds pose but with subtle, natural breathing and slight shoulder relaxation. Lighting: Warm, soft studio lighting casting a gentle shadow on the beige wall. Output: 720p or higher, clean background, no text or watermarks.
The video demonstrates high technical quality with excellent lighting, color grading, and a smooth, professional camera track that fits the 'premium movement shot' brief. The model's facial identity and hair are accurately preserved from the reference. However, the clip fails a critical adherence test regarding the garment silhouette. The reference images clearly depict a structured, mid-thigh/knee-length coat with exposed black pants underneath, whereas the generated video portrays a floor-sweeping, full-length trench. This is a fundamental mismatch of the 'hero trench silhouette.' Additionally, there are distracting temporal artifacts where a stray strap or fabric tail inconsistently appears and disappears throughout the clip.
WHAT SURVIVED REVIEW
Perfect model facial recognition and styling (hair, skin)
High-fidelity lighting and atmospheric composition (golden hour)
Stable and smooth lateral camera tracking
criticalSubject Adherence / Silhouette
Evidence
Reference Image 2 shows a coat ending at the mid-thigh/knee exposing black pants. The video coat extends to the floor/ankles, violating the visual identity constraints.
Required fix
Strictly constrain the output to match the reference coat length and cut (short, structured).
majorTemporal Coherence
Evidence
Timestamps 0.0s, 3.0s, and 7.0s show a long strap/fabric trail; this element vanishes completely at 0.5s and 3.5s, creating a flickering morphing artifact.
Required fix
Use stronger image conditioning or negative prompts to prevent hallucinating unstable trailing elements.
Revision prompt produced from the evidence
Create a smooth 8-second lateral tracking shot of a woman with short dark wavy hair walking left to right in profile. She is wearing a structured copper-brown double-breasted coat that ends at the mid-thigh (exposing black leggings/pants), strictly matching the provided reference visual. The background is a sunlit plaza with large stone columns. Ensure the coat hem remains perfectly stable and solid throughout the movement; do not generate floating straps or fabric tails. Photorealistic, 8k, fashion commercial.
SCENE 03 · ACTUAL-MP4 REVIEWCleared as an individual shot
88/100
decision: passthreshold: metblocking issues: 0
The clip delivers a high-fidelity execution of the premium editorial brief. The subject's likeness—specifically the facial structure, hair texture, and skin tone—aligns strongly with the provided identity references. The environment correctly captures the 'blue hour' aesthetic with appropriate atmospheric lighting and background bokeh. The camera movement is smooth and continuous, adhering to the 'very slow push-in' requirement. While visually compelling, the surface textures (coat and skin) lean slightly towards a smoother, digital finish rather than the raw realism implied by 'photorealistic' and the reference materials.
WHAT SURVIVED REVIEW
Excellent facial likeness continuity with the reference model.
Accurate color palette matching the 'oxidized-copper' trench description.
Seamless temporal coherence with no flickering or morphing artifacts.
Correct implementation of natural micro-movements (blinking, breathing).
minorMaterial Fidelity
Evidence
Throughout the clip (0.0s-8.0s); the coat surface appears relatively smooth and plastic-like compared to the visible wool weave and heathered texture in the reference images.
Required fix
Instruct the generator to emphasize 'heavy wool texture', 'fabric weave details', and 'natural roughness' to better match the reference material.
minorSkin Texture
Evidence
Frames 0.0s-8.0s; the skin appears slightly over-smoothed or airbrushed, lacking the fine pore structure seen in the reference photos.
Required fix
Add keywords like 'unretouched skin', 'natural pores', and 'realistic imperfections' to ensure a more authentic dermatological finish.
What Glove added around the model call
Stage
Durable output
Why it matters
Creative direction
Aethel, model definition, hero coat, three scenes
The human asked for the outcome; the agent supplied the campaign.
Identity
8 generated references; 3 selected anchors
The same visual contract feeds generation and independent review.
Three jobs remain one resumable production rather than unrelated prompts.
Inspection
7 stored verdicts with evidence and revision prompts
A provider success never becomes an automatic creative success.
Delivery
All-or-nothing sequence decision
One passing hero clip cannot conceal two campaign-breaking shots.
The receipts
Measure
Recorded value
Identity references generated
8
Video drafts generated and reviewed
7
Generated video duration
44.2 seconds
Recipe-attributed media cost
$5.9019
Account-reconciled complete task
$9.3401 / $10.00
Unspent hard limit
$0.6599
The account figure includes earlier experiments, model-selection smoke tests, director calls, video generation, and reviewers. Keys are read at runtime and never written into the gallery data or client bundle.
The delivery rule is code, not a suggestion
campaign-policy.tstypescript
// The agent chooses the concept and runs the production.await agent.processRequest("Create one recurring brand model across three scenes.");// Each selected shot must have a latest passing actual-video review.await glove_video_flow_deliver({ run: campaignRun, replacements: revisedShots,});// The tool refuses the entire sequence when any shot is unreviewed,// below threshold, marked revise, or has a major/critical issue.
That is the point of glove-video: not pretending every expensive generation is good, but giving an agent enough memory, tools, evidence, and authority to run the creative process—and enough policy to stop it from shipping what it should know you will reject.