Kira Labs / Research

Source-report reading edition · 20 October 2026

Report 70 · Evelyn’s Three-Angle Head Rendering

Three genuine compatibility-render views passed visibility and camera-difference checks; blink, voice, lip synchronization and release readiness remain unverified.

Original document: 70_Evelyn_Verified_Head_Rendering_2026-10-20.pdf

This is a complete text reading edition extracted page by page from the supplied PDF, not the original binary PDF. Text, tables and spacing may differ from the PDF; screen captures from Report 70 are linked separately. The original source is retained, and the latest status is on the evidence page.

Actual saved Evelyn head captures

These are the same actual compatibility-render screenshot captures currently published in the site's evidence gallery, not generated illustrations.

Source page 1 of 2

KIRA LABS | RESEARCH EVIDENCE


Evelyn Hart: a visible 3D head
A successful owned-browser test rendered the head from three distinct camera angles. The
images below are actual captures from that test.




                                                                                     Three-quarter view




Front view
                                                                                     Side view



Verified result
The face, eyes and hair were visible. All three views passed the fixed foreground and
camera-difference checks. The independent saved-evidence audit passed, and original screenshots
agreed with compositor captures.

This is a human-shaped 3D preview with simplified materials. It is not a photorealistic appearance
result. The lower neck crop and stylized hair/eyebrows still need refinement.

Speech was disabled in this rendering test. Blinking, animated facial motion, synchronized speech,
generated dialogue, memory across restarts and installation remain separate, unverified milestones.




Asset credit: Base Female Character, Black Project, labeled CC BY 4.0 in the local source record. Original attribution and
distribution permissions still require final release review. These images document a private research preview.




October 20, 2026 | Experimental private preview                                                                              1

Source page 2 of 2

KIRA LABS | RESEARCH EVIDENCE


How the head test was performed
1. Freeze the implementation. The test used the reviewed compatibility-diffuse512 preset:
simplified diffuse materials with bounded texture sizes. The existing geometry and camera
crop were preserved.

2. Apply a fresh test gate. Other work tracks were held. Available RAM was checked before
launch against the unchanged 8 GiB startup floor, with a 3 GiB reserve rule. The owned head
process remained subject to its 4 GiB job limit.

3. Capture controls and three views. The test recorded an empty-scene control, then front,
three-quarter and side views at 800 x 800 pixels. It checked that the required object was
visible and that the views actually differed.

4. Check the original evidence. A separate saved-evidence audit verified the original PNG
provenance, pixel acceptance, compositor agreement and owned-process evidence. The
three accepted views had no recorded WebGL error or context loss.

5. Close the test. Browser, server and component cleanup were checked before the test
evidence and finite resource grants were closed. No test replay was used to replace a failed
original.

Measured scope
Foreground occupied approximately 17.9%, 18.6% and 22.2% of the three captured views. Pairwise
view differences were approximately 18.6%, 24.3% and 19.0%. Original-to-compositor disagreement
was 0% for each accepted view.

What this result does not establish
The test does not qualify the original physically based materials, continuous RAM availability, a
complete graphics-memory total, realistic skin mechanics, conversation, voice quality, lip
synchronization, persistent learning or an installer. Successful rendering alone is not a completed
Evelyn release.

NewBrain remains a separate primary research track. Original12 remains 2/12 and Qwen has not been
tested on that comparison. The latest diagnostic setup stopped before a model call; it produced no
new learning result.




Evidence identities: actual head test 1037; independent saved audit 1038. Earlier black captures and failed attempts are
preserved. Public release remains pending private installation and user review.




October 20, 2026 | Experimental private preview                                                                            2

← Back to research evidence and all reports