← Development updates · NewBrain · Avatar Builder
Reviewed November 4, 2026 · Independent Kira Labs research status
When a test completes, the outcome matters—even if it fails.
No new accepted neural model, human likeness or full biological body. This report reproduces reviewed aggregate test-status records, not independently rerun private cases. It excludes private experiments, training curricula, personal media, rejected image renders, character reference videos, raw states and model files. Mirrored research records are not independent replications.
NewBrain: completed V9 study, negative learning outcome
The controlled native V9 study completed fixed teaching and evaluation. Complete-answer scores were the same before training, immediately afterward and after fresh-process restart: directly taught facts 0/6; rules 0/2; paraphrases 0/6; new two-fact compositions 0/3; correct abstention on unsupported questions 1/3. These denominators belong to this experiment, not to a different benchmark.
All 20/20 post-training answer outputs repeated exactly after restart, but matching an incorrect answer is persistence, not successful learning. Teacher-target likelihood increased in seven of eight reviewed rows; none passed the fixed success threshold. The older protected Original12 result is 2/12, not rerun, and must not be called improved or regressed based on this separate study.
A later controlled V3 evidence-pathway experiment has been launched and remains active, without a completed scored result. Its planned conditions must not be presented as execution success.
Avatar Builder: 13 rejected head views and failed V7 full body
The original Evelyn V11 procedural head completed its numerical solve and generated 13 static review views. Visual review rejected the face because of jaw, neck, ear, lid, nose and lip problems; an independent strict native floating-point audit also failed. The approved original blue eyeballs are preserved, but there is no newly approved head.
A separate actual V7 full-body attempt processed five coarse body regions and then hit the unchanged 2 GiB allocation guard. No usable body or export was produced. This is not evidence of functioning anatomy, a successful motion rig, or integrated digestive, urinary and reproductive systems. No approved separate fictional-character reconstruction was established.
Source-only successors: proposed V12 head source and 17 intended views are not a completed render; proposed V8 body storage/lookup repairs passed bounded review but not a complete native full-body build. No clinical or personal-source imagery is included.
Evidence and publication boundary
Source: repository commit-status history and reviewed saved test-result summaries dated November 4 (evening, U.S. Eastern). These negative results are not proof that all future approaches will fail. Details such as private held-out questions, raw answer text, model snapshots, protected reference media and rejected renders remain nonpublic. The website and Iris describe only approved public-safe aggregates; each new model or anatomy claim requires separately executed, dated evidence.
Ask Iris about the published research → · Development timeline →