← Development updates · NewBrain · Avatar Builder
Kira Labs · Research and source review · November 2, 2026
What actually changed—and what to test next
No new approved language model or human body. This report distinguishes repository engineering, separately executed negative measurements, and external research proposals. Studies mirrored across NewBrain, KiraWorld and Patient Horizon are not independent replications. Private body renders, identity recipes and model states are not included.
NewBrain: durable state, but meaning acquisition still fails
A November 1 opt-in native Adam candidate recorded exact optimizer moments/age, guarded zero-update adoption and a separate-process reload. This is persistence engineering, not improved answers. An optional score-normalized attention candidate then passed scoped source/math checks, including invented finite-difference fixtures. A subsequent separately measured continuing attention-policy study returned 0/6 direct acquisition, 0/6 paraphrases and 0/6 unfamiliar use, with 2/6 unsupported-question results, both immediately and after delayed restart.
On November 2, the isolated CONTENT_STATE_MEAN_DIRECT readout proposal averaged delivered content-position states rather than relying only on the terminal recurrent state. Its portable source review covered eleven groups and 37 negative cases; the ordinary test suite reported 79 passes and 34 optional skips from 113 tests. The proposed readout had not run a new independent learning qualification. Preserving the wrong answer across restart is not learning it correctly. The earlier protected Original12 result remains 2/12, not retested.
Next discriminating test
- Freeze a new authorized lesson set, a genuinely unseen held-out application set, and unsupported-question controls before any training.
- Use the same examples and evaluation for legacy direct readout versus the optional mean readout; separately compare existing SGD and opt-in Adam without conflating changes.
- Evaluate complete free-generated answers, paraphrases, unfamiliar use, regressions and delayed independent-process restore—not teacher-forced loss alone.
- Keep experiment custody, exact checkpoint fingerprints, optimizer state and all failures. Report unchanged or negative results without declaring Qwen replacement.
Outside research worth testing, not adopting blindly
GradMem (ICML 2026) optimizes small prefix memories with fixed main-model weights; compare a memory-only baseline to costly neural updates. FOREVER (ACL 2026) schedules replay by optimizer update magnitude rather than merely counting steps; test for old-control regression. Replay Can Provably Increase Forgetting (2026) cautions that careless replay can worsen retention; include a no-replay control. These papers test different models and datasets; none is evidence that NewBrain has acquired general language ability.
Avatar Builder: a safer workspace, not a complete person
The November 1 experimental pipeline prepared 68 caller-supplied exterior shape controls and mathematical fields for torso, limbs, palms, fingers and feet, with an opt-in vectorized sampler and guarded meshing adapter. A separate bounded 2D-reference module proposes proportions from authorized local images with explicit occlusion and uncertainty flags. Its local host-dispatch source reaches the intended native pipeline, but its software fixtures mock or intercept real meshing and inference. No accepted body results from these tests.
The November 2 workspace launcher update uses a build-identity check before reopening a local Avatar Builder server. It passed twelve scoped standard-library tests and an actual Windows HTTP identity/state/page check. The native GUI, body editing, new full mesh, material, deformation, anatomical accuracy and motion acceptance remain unqualified. Earlier visually rejected original heads and bodies remain rejected. Original-only production and private-media safeguards remain unchanged.
Next actual Blender experiment
- Run one original, neutral profile through the real pinned local native mesher; preserve genuine input, output, source and failure receipts. Do not substitute imported character geometry.
- Open and reopen one editable Blender scene and GLB, inspect front, profile, rear, three-quarter and clay views; measure collisions, surface continuity, limb proportions and neck seams.
- Only after visible anatomy passes, add rig/skin weights and posed shoulder/elbow/hip/knee tests; keep the original eyeball milestone separate until truly fitted.
- Register named internal structures separately from exterior geometry, with anatomical landmarks and appropriate source/license provenance; do not infer organs or physiology from 2D silhouettes.
Outside reconstruction and anatomy references
SAM 3D Body (CVPR 2026) estimates body, hand and foot pose using a parametric body representation and optional 2D guidance; benchmark its silhouette/landmark consistency without copying its meshes. SOMA (March 2026 preprint) demonstrates adapters across incompatible human-body topologies and rigs; study a neutral common landmark vocabulary.
CADS (Nature Communications, October 28, 2026) describes 167 anatomical CT labels across 22,022 volumes. It is a segmentation reference, not a fully animated person, a patient-specific digital twin or biological simulation. CADS distribution offers model variants with different licenses; each training-data source has separate reuse terms. For Avatar Builder's original-only policy, start with independently authored anatomical geometry and use licensed datasets strictly for permitted validation and educational comparisons.
Evidence and release boundary
This is a public summary of reviewed technical source commits and published external research, not a production deployment or fresh execution by the website. No private person records, photos, model state, unreleased voice, original uncensored body media or patient records are published here. The site and Iris should direct visitors to the dated record, while remaining clear that neither a Qwen replacement nor a realistic full-anatomy body has been demonstrated.