Kira Labs / Research evidence / October 18, 2026
A single trained-on NewBrain answer was correct after a fresh-process reload. A separate 12-case experiment regressed. Kira World also recorded bounded avatar-geometry and synthetic sensory-component results. These are internally reviewed experiments—not a working artificial brain, verified nerves or clinical system.
What was measured
These scores use distinct question sets and must never be added together or substituted for one another. Each result belongs to its own declared evaluation.
After 64 updates and 128 presentations of repeated examples, NewBrain correctly output the single-token target ball after saving and reloading in a fresh process. The three-token target cup word one instead ended immediately without an answer.
This is one trained-on example, not held-out learning, broad language ability or delayed retention.
A separate frozen 12-query experiment used 64 updates and 384 repeated presentations. The two baseline passes were missing-information responses—not learned historical facts. Following training and reload, neither passed. The original 2/12 state was preserved.
The cause remains unknown. No hidden answer repair or easier benchmark was applied.
A separate saved-data review independently checked 30 sampled parameter coordinates at each of three fixed step sizes: 90 central differences drawn from 181 original objective calls, across ten parameter arrays containing 11,914 model parameters. All selected comparisons met the declared tolerances.
This verifies a local sample at one Public2 checkpoint and batch, not every gradient, an explanation of the Original12 regression, improved learning, or external academic replication. Earlier failed audits remain documented.
Avatar and sensory engineering
Geometric movement and invented-voltage controls are useful infrastructure; they are not biological skin, nerves or a closed-loop learning body.
A recorded forearm-command study moved all 512 sampled weighted skin vertices at the nonzero pose settings and restored the exact saved coordinates and pose channels.
No proof of passive elasticity, sliding skin, touch sensing or visual realism.
One fixed synthetic skin-event component passed eight predefined criteria, including zero/nonzero inputs, duplicate suppression, two-avatar isolation, session/epoch rejection and state/reset behavior. The nonzero invented input produced 0.00025 V versus 0 V for the zero control.
Units and inputs are invented engineering values, not calibrated biological signals.
Future plans define anatomy, typed sensory events, a local simulation adapter, Pulse/Three.js physiology research and a bounded ROS 2 intention bridge. No live skin-to-learner response, breathing-lung model, physiological validation or deployed robot is demonstrated.
What the research could support
Temporary AI Creator: KiraWorld contains draft candidate profiles, source plans, personal interests and review workflows for specialists including Emily Carter (programming), Jessica Hale (robotics), Ryan Hale (quantum mechanics), Laura Mitchell (legal review) and Sarah Bennett (public relations). A historical H. H. Holmes draft is anchored to a late-1893 knowledge boundary. One preserved deterministic historical-persona test uses repeated constrained replies, so it is not evidence of free-form historical conversation.
Education: a proposed Civil War guide would distinguish researched events from fictional backstory, allow the learner to request a break, and later resume from a verified lesson state. Accuracy, reliable resume, privacy across students and educational value remain to be tested.
Synthetic Robert: a separate local identity/voice configuration and authorized memory records exist. The owner reports conversations and a familiar Chatterbox voice, but fresh live speech/conversation qualification and a realistic twin body remain unverified. This is not a biological person, medical patient simulator or a proven NewBrain integration.
Patient Horizon: a synthetic-only scenario planner and versioned medical-research workflow are the current foundation. Real patient ingestion, organ simulation, clinical evidence synthesis and treatment predictions are future development requiring separate validation.
Collaboration request / research phase
Kira Labs is seeking faculty expertise in machine learning, education, human-computer interaction, biomechanics and robotics; one or two research partners for carefully scoped replication; secure laboratory/remote access; and approved shared computing. The proposed memory, GPU and equipment tiers are planning targets, not measured minimums. More hardware alone will not resolve the Original12 regression.
The read-only objective-versus-generation diagnostic has now passed its own four execution controls and independent internal saved-data audit. Short and long row losses were 1.3116488 and 2.4736416; combined loss 2.0863106 matched the token-weighted row values within 1e-12. Generation remained ball versus immediate EOS for the three-token target. Next, test token-position losses, EOS competition and a fresh acquisition/stopping protocol; replicate body findings separately. The fixed Qwen comparison, delayed retention and held-out generalization remain untested.
Evidence cutoff: October 18, 2026. Based on the internally preserved NewBrain Original12/Public2 records, dated numerical-control reports, KiraWorld geometry and synthetic-event records, and relevant published source summaries. Review by separate project roles is not outside-university replication. No personal/private runtime data, patient records or unpublished conversation transcript is included here. This dated page is not live monitoring of Codex's desktop.
