NewBrain · November 18, 2026 · Finished but unqualified synthetic rehearsal
Rehearsal recovers previously taught replies, not new fact generalization
A fixed 2,048-update local CUDA continuation recovered replayed old controls 1/16→16/16 and moved familiar-order complete contexts 26/48→30/48. But unseen-order positive answers stayed 5/32, new-wording positives 1/24 and four of five qualification families failed. This is synthetic consumed-task reacquisition with no matched no-replay control or independent human review; Original12 remains 2/12, not rerun. Reviewed results and limitations →
NewBrain · November 18, 2026 · Closed negative CPU learning test
Two 3,072-update copying trials completed; copying still failed
A/w and wording-only fixed arms returned 3/96 and 2/96 primary exact (initial 3/96). Supported answers stayed 0/64 in both; acquisition 3/192 versus 2/192. Consumed retention 4/16→6/16 is not reliable learning. An earlier input-guard failure had zero updates; a corrected successor completed with repeated evaluations of previously exposed—not virgin-independent—cases. No model adoption; Original12 still 2/12, not rerun. Full reviewed study →
NewBrain · November 17–18 Eastern / November 18 UTC · Failed experimental learning studies
Resumed counterfactual copying and rehearsal both missed exact-answer gates
After a failed fixed 640-update trial (0/64 new exact replies, 0/16 copy replies), the paired-curriculum and rehearsal arms each completed 960 further CPU updates from the same failed parent. Both scored 0/96 primary (supported 0/64; unsupported 0/32) and 0/32 separate wording. Retention declined 9/16→4/16 in both. Paired consumed teaching acquisition was 10/96 correct unknown abstentions, not positive factual copying. No trained branch adopted; Original12 remains 2/12, not rerun. Read the full reviewed results →
NewBrain · November 17 ET / November 18 UTC · Closed bounded study
Real GPU training confirmed; learning and retention still missed their targets
A fixed 2,000-update Native CUDA-assisted continuation scored complete acquisition 18/80, new wording transfer 0/32→1/32, fresh-context answers 0/16→0/16 and consumed retention 16/16→12/16. A separate frozen-state CPU diagnostic found changed internal states in 48/48 paired contexts but identical answers in 36/48. The later untrained decoder bridge implementation passed 20 numerical cases and 24 fixed finite-difference checks, but no optimizer run, new learning or answer improvement was demonstrated. Accepted Original12 stays 2/12, not rerun. Read the measured results and limitations →
Kira World · November 11, 2026 · Development recording
Watch the five-minute Louvre-inspired V66 walkthrough
The owner-supplied Version 66 walkthrough now plays directly beside the existing Louvre-inspired world screenshots. The website uses a 640 × 360, 12 fps compressed preview of the recorded five-minute environment study; the original high-resolution recording is preserved separately. This is a procedural virtual scene, not filmed footage of the museum and not a measured Louvre reconstruction. The previously published still images come from an earlier development stage and may differ from V66. Watch the video alongside the screenshots →
Kira World · November 9, 2026 · Review of November 8–9 private development checkpoints
Home World becomes more interactive; the Louvre gains sourced art and architectural detail
Home World: A new separate, local-only residential-district prototype contains walkable neighborhood geometry, furnished rooms, doors and object-interaction code. The current interface documents WASD movement, Shift to move faster and E to interact. Its development server offers a read-only catalog for locally available books and videos, excluding private folders by default. A World Generator intake can organize a request and its missing source evidence, but it does not automatically create a finished world. Code and focused checks are saved; a complete end-to-end resident experience has not been qualified.
Paris Notebook World / Louvre: A later saved owner-review walkthrough adds animated reflecting pools, more detailed courtyard facades and Pyramid framing, plus a red-walled skylit gallery and four catalogued public-domain painting reproductions: Mona Lisa, The Wedding at Cana, Liberty Leading the People and The Coronation of Napoleon. The recorded review build has five rooms, six authored routes and nine review camera positions. The paintings have documented image sources; the architecture, gallery layout and artwork placement are still approximate procedural choices, not a measured Louvre reconstruction. The earlier screenshots on this website show an earlier stage of the prototype; no people, voices, shared travel or VR are activated.
Person Creator and Avatar Builder: Draft character profiles now include clearer age/body-policy boundaries, knowledge limits and consent checks before any potential permanent residency. Separate original-procedural adult male/female body studies and export audits are also saved for private review. These steps do not establish realistic likeness, accepted anatomy, natural movement, a complete desktop workflow or activation of a person.
NewBrain: Larger original-model training remains limited by available hardware; the proposed 9B model has no trained checkpoint. Kira Labs is continuing lighter world-building and character-development work while planning future compute upgrades. Home World details → · Louvre details → · Avatar research → · Support Kira Labs →
Kira Labs Arcade · November 8, 2026 · Tested game compatibility repair
Byte & Bolt bundles optional synthetic narration instead of relying on installed voices
The browser game's earlier Read aloud path could fail when no local English speech voice was installed. The public build now packages 318 prerecorded synthetic Sarah MP3 segments and preserves 12 original exact-match WAV clips, with complete captions, user-controlled speech and a Test voice option. 69 Node checks passed, all 505 authored dialogue variants resolve to audio, and the isolated Chromium playback QA and Pages deployment succeeded on commit 8c880ee. This is not human listening, physical-laptop, offline-file, fan-noise or universal device verification. Read the full public test report → · Try Byte & Bolt →
Kira World · November 8, 2026 · Published first-party prototype screenshots
New Home World and Louvre-inspired environment captures are public
Actual development screenshots now show a Home World neighborhood with paths, homes and a central gathering area, plus views of a separate Louvre-inspired procedural study. These images are already published on the Kira World website and are prototype captures, not finished destinations, not measured reproductions of the Louvre and not evidence of autonomous residents, group travel, functional organs or an accepted original person. The underlying owner-review world work remains separate from public playable releases. View the approved public screenshot gallery → · Home World views →
Kira World / NewBrain · November 8, 2026 · Private development review
World-building continues while larger NewBrain training awaits hardware
The current development review preserves a separate, keyboard-controlled Louvre-inspired walkthrough prototype in Kira World. Its saved study includes an approach courtyard, Pyramid entrance, descending stair, lower hall and abstract gallery, with five authored rooms, six routes and seven review camera positions. Internal review recorded successful bounded route checks. It is a procedural scene with many invented details, not an accurate reconstruction of the Louvre, a completed destination, or a public virtual-world release. No autonomous people, voices, group travel, VR or Home World placement are active in this draft.
New character-design code can prepare fictional personality and life-history dossiers, proportion descriptions and a draft handoff to Avatar Builder. These are design records; they do not generate an approved physical body or activate a synthetic person. Avatar appearance, natural motion and the end-to-end desktop workflow still need qualification.
Separately, NewBrainLM now has an implemented 9-billion-parameter configuration and training tools, but no trained 9B checkpoint. The best saved small-model probe scored 8/25; the latest larger small-model study answered 0/8 supported questions while returning the expected unknown response on 5/5 unsupported questions. More RAM can relieve local constraints, but larger training also needs GPU compute, storage, licensed data and independent evaluations. Kira Labs is continuing work on environments and character tools while planning future hardware. Explore Kira World → · Help the research →
NewBrain · November 7, 2026 · Fresh locked research protocols V5–V14
Independent prompts reveal a learning-versus-abstention trade-off
A derived 96-hidden-unit research model was evaluated on independently locked fresh facts and held-out wordings. In V11, 8/8 supported answers and 4/4 combinations passed, but unsupported-question abstention reached only 1/5. V14 improved unsupported abstention to 5/5, while supported answers dropped to 7/8, combinations to 2/4 and retained supported answers to 5/8. Neither run passed the full frozen protocol; separate-process restart checks passed. These are research variants, not qualification of the continuing Native model. The protected Original12 remains 2/12, not rerun. Full findings → · Timeline →
Avatar Builder · November 7, 2026 · Original-body V90–V95 experiments
Reference proportions affect authored geometry; full anatomy and workflow remain unqualified
New front/side silhouette ratios were mapped to original body control parameters, with interface checks confirming the expected directions for shoulder and hip width changes. A V95 candidate exported, but the recorded reference fit is not a metric scan, does not claim likeness, and says the reference source images were not directly consumed by the production build. Latest independent Avatar unit tests: 150 run, two errors and one skipped; 13 schema checks passed. This is limited reference/proportion engineering, not an approved Evelyn body or completed photo/video Builder. No private reference assets or rejected renders are published. Full findings → · Timeline →
NewBrain · November 6 Eastern / November 7 UTC, 2026 · Bounded V89 study
Four-fact teaching passes reused cases; independent learning still needs proof
Saved synthetic evidence reports 8/8 answer-absent supported wordings after initial teaching and again after rehearsal; 4/4 later facts and 5/5 unsupported results. The 4/4 combination and 4/4 conversational prompts supplied the answer. Revisions reused cases; trained-core restart reproduced 25/25 stored replies, while 3/8 additional controls failed. Accepted Original12 remains 2/12; no fresh held-out Native learning gain. Public evidence → · Details →
Avatar Builder · November 6 Eastern / November 7 UTC, 2026 · V89 staged export
Eye/body assembly fits cap; original animation regresses
V89 procedural export contains 69,997 vertices and 139,059 triangles (941 under budget), with no visual references. Neutral eye seating fits, but eye material colors change and the original five motion clips lose most animation channels. The suite recorded 141 passes, two product errors, one skip and 14/16 failed additional strict controls. No accepted Evelyn likeness, functional anatomy or natural movement. Public evidence → · Details →
Arcade · November 6 Eastern / November 7 UTC, 2026 · Public game update
Byte & Bolt brings a 22-project town to the Arcade
The fifth original Kira Labs browser game features 14 fictional paid town jobs and eight unpaid lab upgrades, scripted resident dialogue, stylized 3D town navigation, finite mixed practice and local saves. Committed release QA documents 59 passing Node regression groups and a managed-browser saved-game continuation; final Arcade homepage browser QA and GitHub Pages deployment passed. Native hardware rendering, offline local-file play, independently verified pronunciation and learner hours remain untested. This educational game is not an autonomous NewBrain or anatomical milestone. Public release evidence → · Arcade → · Timeline →
Avatar Builder · November 6, 2026 (Eastern) · Verified V77 export and animation advance
Original-body export now carries changing motion; anatomical geometry remains rejected
Independent repository review confirmed an original-body GLB/Blender export path with five changing authored animation clips. The Avatar suite recorded 129 passes and one skip. The same candidate still has 68,328 vertices and 136,224 triangles, with geometry, rig, skinning, UV and materials unchanged from the rejected draft; reference images or video were not used in its construction. Integrating approved eye geometry would be estimated at 170,816 triangles—30,816 over budget. Another 29 strict validation controls exposed 22 unhandled cases. This qualifies narrow export/animation-data progress, not realistic anatomy, safe deformation, natural movement, accepted Evelyn likeness or improved Native NewBrain learning. Private body renders and model data remain unpublished. Avatar Builder detail → · Timeline entry →
NewBrain · November 6, 2026 (Eastern) · Later V67/V68 limited technical review
Speech and recovery contracts strengthened; native learning remains failed
V67 reported 79 passing authored checks but only 2/4 passing independent voice probes, exposing audio truncation and post-filter silence. The later V68 scoped review recorded 82 supplied tests passing; after short-audio/stereo repairs, 84 authored checks plus five independent controls passed. These are bounded software checks, not a successful learning benchmark. A four-fact Native author probe used fresh SGD and answer-bearing prompts and does not isolate interleaving or qualify the continuing model. Source readback and private archive verification support recovery, not completed application restoration. Native NewBrain conversation remains unqualified; protected Original12 is still 2/12 and not rerun.Read full qualified scope → · Timeline milestone →
Avatar Builder · November 6, 2026 (Eastern) · Later V67/V68 limited source controls
Write protection and tissue checks improved; original head and body remain unaccepted
V67 revealed mixed candidate source quality, including weakened memory-write protections, and was not approved for promotion. In V68 the memory locks were restored and tissue topology classification repaired. Bounded checks passed in four groups: 5 build mocks, 6 tissue cases, 14 voice-profile cases and 11 curriculum cases. An existing teaching/reference GLB had 68,328 vertices and 136,224 triangles but still used generic eyes; an estimated change to approved eye geometry was not built or deformation-tested. Selected source recovery was independently verified; there is still no accepted original Evelyn face, complete anatomical body, validated physiology or new natural motion. Read the project detail → · Timeline milestone →
Arcade · November 6, 2026 (Eastern) · Released website experience
A new picture-led four-game lobby
Original game-scene captures, larger adventure cards, an opt-in sound setting and a browser-only explorer passport now help visitors discover the Arcade's four playable games and linked lessons. The passport records which links opened, not game completion or progress. The new lobby passed automated browser QA, and GitHub Pages deployment succeeded; full independent playthroughs and sound on all devices have not been verified. This changes the website, not NewBrain or approved anatomy. Dated release report → · Open Arcade →
NewBrain · November 6, 2026 (Eastern) · Bounded speech, hearing and restart engineering
69 author checks passed; independent probes still reveal critical speech and hearing defects
The latest separate hearing-and-voice review recorded 69/69 passed authored checks, reproducible local synthetic word audio and simulated scene outputs, and exact bounded person/voice state across a fresh-process restart. Seven local word WAV outputs and four scene percepts were reproduced; these are not conversational demonstrations. Independent voice controls passed only 3/8, and additional review reproduced unsafe checkpoint migration, false success statuses, a noncausal auditory-nerve path, stereo/sample-rate/chunk problems and scene-supplied information presented as perception. The fixes and negative observations guide the next tests. Native NewBrain learning, usable conversation, physical microphone/speaker operation and whole-person embodiment remain unqualified. Protected Original12 stays 2/12 and was not rerun. Private prompts, files and model checkpoints remain unpublished. NewBrain research detail → · Timeline milestone →
NewBrain research · November 5, 2026 (Eastern) · Separate small-decoder experiment
Taught recall and rehearsal improved in a small decoder; unseen answers still failed
In a separate six-group experimental decoder, answer-free tests recalled 12/12, 24/24 and 48/48 taught mappings after fresh-process reload. In an independently checked two-cohort rehearsal comparison, the replay arm retained 12/12 earlier mappings and learned 12/12 later mappings, versus 0/12 earlier retention without replay. Generalization remained poor: a subsequent crossed test scored 0/8 held-out unknown subjects and 0/16 unseen wording (known 16/16; taught unknown 8/8). All 48 final crossed outputs reproduced after fresh reload. A JSON serialization defect in the evaluation command was corrected and 14 bounded checks passed. This is not the current ten-group native NewBrain, not a complete-answer breakthrough, and not a new Original12 result; the protected Original12 stays 2/12. NewBrain status and limits →
Original head fitting reached a later stage but failed eye clearance
A separate original-head V7 photo/form-fitting run reached and saved its post-form stage, then failed the unchanged geometric eye-clearance check before a new usable render. This is a later solver-stage result, not an accepted face and not a replacement for the separately rejected V12 views. Approved blue-eye geometry and materials are preserved. The full-body V15 run and two proposed boundary variants did not yield a complete usable body; full anatomy, motion and appearance remain unqualified. Avatar Builder status and limits →
NewBrain · November 5, 2026 · Completed negative study
Each of four V3 conditions scored 1/15; evidence choice 0/10
The pathway study closed with no qualified factual learning. Each of four 15-response conditions got only an unsupported-question control right; 80/80 outputs ended with EOS and 24/24 restart comparisons matched, including errors. A separate contextual-practice run is closed but has no qualified public answer score. Original12 remains 2/12, not rerun. NewBrain → · Dated public report →
Avatar Builder · November 5, 2026 · Rejected head variants and failed full-body run
V12 and commissure V2: 17 views each, both rejected; V9 body failed
Two actual original Evelyn head variants failed facial form/visual review while preserving approved blue eyes. Body V9 stopped at the unchanged work guard after four coarse regions; V10 source reviews are not a finished body. A separate October 31 partial torso/arms native result passed bounded static geometry and pose checks but remains unaccepted as a whole person. No completed fictional-character likeness, functional anatomy or organ routes. Avatar Builder → · Dated public report →
NewBrain · November 4, 2026 · Completed negative qualification
V9 complete-answer study failed; separate V3 still unscored
In the completed V9 native study, facts scored 0/6, rules 0/2, paraphrases 0/6, composition 0/3 and unsupported-question abstention 1/3 before teaching, afterward and after restart. All 20/20 final full answers reproduced exactly after restart, including errors. A separate four-condition V3 experiment launched later but has no qualified answer result. Accepted Original12 remains 2/12. NewBrain status → · Public research report →
Avatar Builder · November 4, 2026 · Rejected head and failed body
V11 produced 13 head views but failed review; V7 body hit 2 GiB guard
The actual V11 original-head solve produced 13 static views, but facial anatomy and native-float qualification failed. The actual V7 full-body build hit its unchanged 2 GiB allocation guard after five coarse regions, producing no usable export. V12 head and body V8 repairs remain unrun source candidates. No functional digestive, urinary or reproductive body, approved Evelyn face or separately approved fictional character is claimed. Avatar Builder status → · Public research report →
NewBrain · November 4, 2026 · Ongoing native experiment and new pathway audit
V9 learning test active in latest source snapshot; first-answer credit is next diagnostic
The latest repository continuation records one running, unscored V9 native study. A separate frozen/taught/forced-evidence protocol is designed but not executed; source inspection found no proven missing gradient term. The proposal is to measure how question conditioning and teaching credit affect actual first-answer choices. There is no new accepted learning score; the last scored four groups remain 0/6 and Original12 is 2/12, not rerun. Read the precise status →
Avatar Builder · November 4, 2026 · Rejected V8 head and failed V5 body
Jaw/neck source defect identified; V9 source and body repairs still unrun
The actual V8 head remained visually rejected, but saved-stage diagnostics attributed its narrowed jaw and pinched neck to the original control mesh rather than to post-render shading. V9 source proposes anatomical-chart corrections with approved blue eyeballs preserved; no V9 head has been generated. The real full-body V5 attempt hit its allocation limit, and V6 body repair source remains unqualified in a complete build. Read the current Avatar Builder status →
NewBrain · November 4, 2026 · Source and resource review
V8v2 checkpoint research preserved; no new learning result
A reviewed 294-member source snapshot and resource-admission calculations document candidate checkpoint and buffer-handling repairs. These do not establish actual peak RAM, complete native execution or successful new answers. The latest accepted learning result is still the November 2 failed study (four groups at 0/6); Original12 remains 2/12. No private states or archives are reproduced. NewBrain limitations →
Avatar Builder · November 4, 2026 · Failed V6 and source-only V7
V6 head stopped at a projection constraint; V7 review is not a render
An actual V6 head attempt failed the 25 mm form-projection guard, even though the external Blender exit was zero. The following V7 source candidate passed independent source review but has not produced an actual head build or new approved anatomy. Separately reviewed body calculations and a video-reference front-end proposal also remain unintegrated source work. The previously accepted blue eyeballs stay protected, and no head/body is approved. Avatar Builder limitations →
NewBrain · November 4, 2026 · Checkpoint-memory research
Source-level memory-retention repairs need a real native test
Compact checkpoint encoding and buffer-lifetime candidate mechanisms passed bounded invented-data checks. No measured peak-RAM improvement, complete native integration or new learning success was established. Earlier failed studies and original model states remain preserved. Read the exact limits →
Avatar Builder · November 4, 2026 · Actual original V5 render, rejected
V5 produced 13 static views, but the face still needs reconstruction
The actual V5 original head build produced 12,680 skin vertices and 25,072 triangles while retaining the previously approved blue eyes. Thirteen views were saved, but jaw ridge, lids, mouth/profile, skull and facial structure still failed appearance review. No new approved Evelyn head or body; private images were not published. Current Avatar Builder status →
NewBrain · November 4, 2026 · Source-only memory-reconstruction experiment
Owned record-prefix reader proposed; real saved-state use not tested
A new optional patch adds explicit owned reader entrypoints that replace certain repeated prefix-byte scans with incremental accounting. Nine small portable source/metadata test groups used invented records. There was no native saved-state read, measured speedup, restart qualification or learning gain. The default model and earlier failed scoring result are unchanged. NewBrain source and limitations →
Candidate scheduling experiments target an earlier work-budget failure
Kira World's optional V1 scheduler passed eight small invented tests; V2 adds a concrete abstract reducer interface and ten small tests. The aim is to avoid costly repeated sorting and stale queued candidates. This is not a tested whole-body speedup, new realistic head or accepted Blender export. The previously accepted round eyeballs are preserved; the latest original head/body remain rejected. Avatar Builder status →
NewBrain · November 2, 2026 · Later completed negative study
216 genuine learning updates; none of the answer groups passed
The optional mean-content-state readout completed 216 genuine teaching updates and 72 generations. Direct facts, paraphrases, unfamiliar two-fact uses and unsupported-question abstention were each 0/6 before training, immediately after and after fresh restart. All 24 immediate/restart full outputs matched but were wrong. Study integrity passed; learning failed.
NewBrain and Avatar Builder · November 2, 2026 · Dated engineering results and outside research
NewBrain's readout study remains unqualified; Avatar Builder's workspace launcher is safer
NewBrain's optional content-state mean readout passed bounded source checks but has no fresh learning score. The preceding evaluated attention-policy study still failed direct acquisition, paraphrasing and unfamiliar use at 0/6 each; its 2/6 unsupported-question result is not new understanding. Original12 remains 2/12, not rerun. Avatar Builder's shortcut now verifies the local workspace identity (12 offline tests and a Windows HTTP check), while real image-to-body inference, meshing, body realism and movement remain unqualified.
External research suggests specific next tests: fixed-model memory writing and replay comparisons for NewBrain; reference-guided geometry and medically defined anatomical labels for Avatar Builder. These are research leads, not implemented features. Read the November 2 evidence and research brief → · NewBrain → · Avatar Builder →
Avatar Builder · November 1, 2026 · Source and synthetic software fixtures, not an accepted body
New profile-driven original-body pathway is prepared; native execution is pending
New KiraWorld source introduces 68 explicit exterior shape controls spanning torso, limbs, palms, fingers and feet, plus numerical field sampling and a guarded adapter to the existing original meshing functions. A separate 2D-reference module proposes shape ratios from bounded pose and outer-silhouette observations; private identity, ownership and approval checks remain host requirements. A new Core dispatch preserves legacy calls and routes authorized requests toward these experimental components.
What has not happened: no actual full native profile-body build, real image/video inference, Blender render/export, approved head or realistic person, functioning organs, or learned movement was established by these source tests. The October 31 visually rejected original body remains rejected. Detailed public source and limitations report → · Avatar Builder status →
NewBrain · November 1, 2026 · Executed memory selection; separate failed learning
Historical context retrieval survives synthetic history growth
An independently checked retrieval-only study completed 800 calls across four isolated routes, five history scales and 40 questions. Opt-in historical selection met 34/37 answerable evidence contracts throughout; default Memory72/77 met 23/37 at +0/+10 and 0/37 at +50/+100/+200 due to context-budget rejection. Historical retrieval changed 0/11 to 11/11; current retrieval remained 23/26. This is not neural answer accuracy, production integration or general understanding.
Separately, a genuine continuing model finished 1,024 updates and 30 generations with no successful complete known, held-out or retention output. Another study is ongoing with no reported result. Original12 remains 2/12. Reviewed report →
NewBrain and Avatar Builder · October 31, 2026 · Executed engineering, documented failures
Memory retrieval measured; original connected-body render rejected
NewBrain: an independent 30-question synthetic evidence-retrieval baseline found NewBrain Memory72/77 at 25/30 with +0/+10 distractor sessions and 0/30 at +50/+100/+200 due to fixed context insufficiency. Kira's scoped route held 24/30. This is not generated answer accuracy. A separate ongoing learning study reduced teacher-forced loss but not successful complete free output or held-out understanding; Original12 remains 2/12. Measured findings →
Avatar Builder: a new connected original body with 45,528 vertices, 46,504 faces and 68 rest bones passed native render/reload gates, but the owner rejected its face and form. Two later revisions were also rejected. New implicit regional/body and facial sources are unrun. No original Evelyn body, fictional reconstruction, functioning organ system or real movement is accepted. Rejection and limits →
Avatar Builder and NewBrain · October 31, 2026 · Measured software boundaries, not a new body or learning gain
Motion library: 22/22 software tests; NewBrain: one native call-path agreement
Motion software: isolated versioned motion definitions, CPU walk timing, quaternion blending, feature matching and an append-only local practice registry were reviewed. Candidate/evidence binding and GLB format safety were corrected; 22/22 pure software tests passed. Ingestion is mocked, no body moves, and transfer to another body was not demonstrated. Full limitations →
NewBrain: in one unscored two-restore comparison, talk and model.generate produced identical complete native output and true EOS with zero training updates. An earlier malformed attempt failed before generation. This is software-path agreement, not teaching or understanding; accepted Original12 remains 2/12 and was not rerun. Reviewed result →
These records are each mirrored across three repositories, not independent replicated results. The original Evelyn head, connected body and organs remain unaccepted; Source109/112 are unrun. Avatar Builder → · NewBrain →
NewBrain and Avatar Builder · October 31, 2026 · Engineering prerequisites, not new learning or rendering
Continuing-teaching eligibility was not proven; assembly-camera improvements remain source-only
NewBrain A116: the existing proposed comparison would acquire one ordinary answer, twelve assessments and one neural update. Seventeen frozen source/protocol/caller commitments received blind static review, but eligible session custody was not established: records did not prove that the latest same-teacher practice had actually completed, that the full saved history was bound to the live state, or that fresh teaching authority and input framing were qualified. No new input was nominated. Zero interpreter/model calls, training steps, replies or scores occurred. This does not prove that all saved practices are ineligible. Read the A116 boundary →
Avatar Builder Source112: an adjunct to the previously staged Source109 head–torso connection specifies six same-scale orthographic views based on all evaluated assembly vertices, intended to replace an unpositioned torso-only camera. Camera clipping and vertex-containment reporting are defined, but all Blender execution, export, rendering and visual acceptance remain unrun. No approved connected body or original head was produced. Read the Source112 design status →
The separate Source112 NewBrain dialogue audit found no demonstrated production correction-path defect; optional assertions with mismatched memory error labels remain excluded and unrun. Existing research states, rejections and private records are preserved. NewBrain's accepted Original12 benchmark remains 2/12, unchanged and not rerun.
NewBrain · October 31, 2026 · One native A108 diagnostic
Reply diverged at the first token, before true EOS
One actual native reply did not match a reserved reference: the first reference token ranked 77; EOS came second after divergence. No new generation or teaching occurred. The Original12 score remains 2/12, not rerun. Research report →
Avatar Builder · October 31, 2026 · Source109 source-only
A proposed 128-to-64/128-point neck/torso attachment and neutral Blender/GLB export path are statically reviewed, but geometry, tests, renders and visual acceptance remain UNRUN. The original Evelyn head/body and full internal anatomy are still unaccepted. Engineering report →
NewBrain · October 30, 2026 · A98 measured continuation persistence; October 31 UTC blocked diagnostics
Two continuing lessons saved and restored; no new learning-accuracy result
The A98 isolated native test completed 2/2 authorized continuation state transitions, preserving the earlier practice through save/reload and exact fresh-process restore. 22/22 no-mutation rejection checks passed. This validates bounded state persistence, not new word understanding or semantic transfer.
Subsequent October 31 UTC entry checks A101/A103 stopped before new model execution: the diagnostic caller underestimated the existing saved-memory envelope and later preparation was not qualified. New actual-prefix code remains UNRUN. Earlier A87/A92 limited familiar-word 2/5, held-out 0/6 and 20/20 identical restart records remain separate. Accepted Original12 remains 2/12. Read the reviewed A98 report and entry blockers →
NewBrain · October 30, 2026 · A94 continuing-lesson input boundary
Next teaching continuation stopped before training: full context was 501 tokens against a 256-token limit
The continuation experiment entered its one permitted input-producer invocation, then stopped while encoding the complete attributed memory and question. The full text required 501 tokens, exceeding the 256-token budget approved before execution. Preliminary review measured only a short probe and omitted the memory wrapper. No limits were raised, no retry was run and the remaining interpreter slots were cancelled. The continuation updates and final restart were unrun; this is an engineering input-extent failure, not a new learning score.
Codex preserved the failed attempt and its error evidence. The next nominated experiment must count the entire encoded message—including role framing, attribution, memory, question and metadata—before admission, and use new independently reviewed inputs rather than reusing the consumed attempt. Prior A87/A92 findings remain as published: 2/5 familiar, 0/6 held-out, 20/20 identical restart records, and regression of two previously correct missing-information controls. The accepted Original12 remains 2/12. Read the NewBrain technical boundary →
NewBrain · October 30, 2026 · Reviewed A87 teaching and A92 restart
The separate 32,812-parameter UTF-8 NewBrain extension completed 128 additional neural updates on a previously attempted term. Fixed familiar questions improved 0/5 → 2/5. That is not 128 conversations or proof of novel word understanding.
All six substantive held-out meaning/use/context prompts scored 0/6; a different-wording delayed probe scored 0/1, and two previously correct uncertainty controls regressed 2/2 → 0/2. Final accuracy was 3/20, including one correct held-out missing-information reply. A distinct-process read-only replay matched 20/20 complete native generation records exactly without more updates—preserving both right and wrong answers.
Evidence boundary: overall predefined learning success false, no checkpoint adopted, Original12 still 2/12 (not rerun). Separate Source93 continuation code is static-reviewed with all new checks UNRUN. Read the dated public A87/A92 report →
NewBrain · October 30, 2026 · Actual75 original-checkpoint migration, word lesson and saved-state restart
A word lesson survived restart, but none of the three meaning/application questions became correct
Codex completed one measured developmental study using the separately versioned Source72 UTF-8 language extension. The preserved 11,914-parameter/42-token original checkpoint was explicitly migrated without training to a 300-entry vocabulary and 32,812-parameter extension; original state and counters stayed byte-identical. One authorized common-word question, explanation and correction was rehearsed through 16 neural training updates, not sixteen separate conversations.
Measured free answers: meaning, rephrasing and application remained 0/3 before and 0/3 after; a separate missing-information control improved from 0/1 to 1/1, while a memory-assisted application question was 0/1. Teacher-forced loss decreased from 10.40055 to 7.34226, but the model's factual answers still ended after the definition's first article. No successful word understanding or generalization was demonstrated.
Persistence: all nine fixed post-training answers and their full generation records matched in a separate process restart without additional updates. Nineteen primary/restart responses reached EOS. The saved explanation, correction history and teaching targets were retained. This supports a narrow save/restart milestone—not full conversational learning, an adopted replacement checkpoint or long-term durability. The accepted Original12 remains 2/12 and was not retested. Read the complete bounded results and limitations →
NewBrain · October 30, 2026 · Executed software qualification of a separate extension
New text and memory interfaces passed software checks; language learning remains unproven
A separately versioned Source72 language path includes lossless UTF-8 byte fallback, expandable input/output vocabulary, typed ordinary messages, owner-scoped attributed memory, corrections, explicit confirmed teaching calls and versioned state bundles. One direct invocation exited 0 after six software assertion groups and a distinct source/backend guard returned normally. These groups are not individual case counts, and the mirrored records are not three experiments. Read the reviewed public report →
Limits: No real conversation learning, unfamiliar-word meaning, old-checkpoint migration, full live avatar integration or new benchmark accuracy was verified. The prior original-core ordinary greeting still failed before response generation. The 11,914-parameter, 42-token original core and its accepted Original12 score of 2/12 remain unchanged.
NewBrain · October 30, 2026 · Original-core conversation boundary and executed software checks
One ordinary greeting could not enter the original 42-token NewBrain interface
An independently selected plain-language greeting was submitted once to the unchanged 11,914-parameter, 42-token decoder. The single generation call reported an input-format rejection before any reply was generated: the original system required a structured MEMORY/context/QUERY representation and none of the five greeting tokens was in its vocabulary. No model update, retrieval workaround, alternate language model, retry or conversation score occurred. This identifies an interface restriction, not a neural understanding failure. Read the original-core test and boundary →
Two additional October 29–30 engineering observations remain separate: one isolated execution-context prerequisite passed 73 software boundary checks before its planned learning study was blocked at evidence preparation; another bounded content validator passed 46 positive and 46 intended rejection cases, but the associated future study was interrupted by expired permissions before neural entry. Neither provides a new measured learning score. The accepted Original12 remains 2/12.
Future work: a separately versioned text interface, larger vocabulary, longer responses and conversation-memory/confirmed-correction lessons are proposed directions. A later separately versioned Source72 path passed software assertion groups, but still has not demonstrated a successfully running learned conversational model. Any subsequent learning claims require fresh questions, exact restart evidence and separate measurement of neural learning versus memory retrieval. No private cases, benchmark answers, weights or user conversations were published.
NewBrain and Avatar Builder · October 29, 2026 · Later measured engineering and unrun source review
NewBrain generated 216 complete outputs, but the study was not scored; Avatar Builder excluded an infeasible head-width target
NewBrain Actual62: a new paired study passed its typed-input prerequisite (58 malformed-input rejections and two model-entry checks), then completed 384 training updates per arm. The full scientific archive saved 175,017,705 raw / 122,485,745 compressed bytes within unchanged limits. Its primary generation and distinct restart each saved 108 complete outputs and traces. Although individual observations validated, the full 108-pair restart qualification failed because a changing memory-availability snapshot hash was passed as the fixed host identity. No late scoring took place and there is no valid new accuracy result. The accepted Original12 remains 2/12; no checkpoint was adopted. Read the NewBrain evidence and limits →
Avatar Builder reference study: genuine evaluated head/torso geometry and eight Blender camera inputs were captured. The existing head was 16,385 vertices / 16,320 faces, while retained/diagnostic cropped torso surfaces were each 3,086 vertices / 3,000 faces. An unchanged 12 mm per-control original head correction bound implies at most a 24 mm signed width change between two supported locations under the declared convex spline assumptions. An authored cheek-width target exceeded this necessary bound, so the specific proposal was rejected before construction. No new visually accepted head or body was produced. Read the original geometry review →
Next source phase: a static-reviewed NewBrain implementation now separates stable host identity from varying resource availability and validates report/protocol ownership earlier. Avatar Builder source adds explicit movement-control contracts and requires real evaluated surface evidence, rather than metadata-only comparison. Neither repaired runtime nor new Blender construction was exercised. These are source reviews, not passing new experiments. Private state arrays, protected answer keys, photo references and other confidential records remain unpublished. The same notes mirrored in three research repositories document one actual study and one source phase, not independent replication.
NewBrain and Avatar Builder · October 29, 2026 · Later engineering update, with explicit failures
NewBrain's next complete-answer tests stopped before scoring; lossless evidence repair is source-reviewed but unrun
The first attempted matched complete-answer diagnosis (Actual53) completed 384 updates per arm but stopped during answer acquisition when identical four-counter values appeared as a tuple and a list. No complete scored response set or separate-process restart followed. A later attempt (Actual57) also completed the planned training and passed limited numerical/custody checks, but its compressed scientific-data archive exceeded the declared 136 MiB size category. The guard stopped the run before answer generation or scoring; the exact oversized compressed size was not logged. These are incomplete technical attempts, not positive or negative measured learning results. Read the full NewBrain status →
Codex subsequently preserved the failures and static-reviewed a lossless scientific-data format (Source58) intended to share identical numerical array-byte payloads without losing distinct logical arrays or their identity records. It has not been executed; actual compression savings, archive-cap fit, answer evaluation and restart remain unverified. NewBrain's accepted Original12 still stands at 2/12, with no new checkpoint adopted.
For Avatar Builder, separate source-reviewed camera-transport code now distinguishes actual render camera data from planned viewing parameters, and integrates those records into proposed head/torso comparison paths. No new Blender render or measured shape improvement resulted from this source work. The existing A44/A45 models remain visually rejected, with complete anatomy and movement unfinished. Read the Avatar Builder boundary →
Prior unsuccessful experiments and their private data are preserved. Mirrored reports in NewBrain, KiraWorld and Patient Horizon describe the same work, not multiple replications. No private model state, protected answers, identity records or uncensored media have been published.
NewBrain / Kira World Avatar Builder · October 29, 2026 · Reviewed new measurements
NewBrain's routing becomes numerically stronger, while factual-answer learning and original head realism remain unresolved
NewBrain Actual47: one controlled original-core numerical test completed 988 native objective/gradient evaluations, 480 parameter differences and 18 feature differences. Normalized routing increased a measured first-prefix logit effect to about 0.11136 (L2), compared with 0.0000000504 under the older product rule; this did not change the top token or measure answer correctness. Actual49: on the unchanged protected Original12 assessment, matched legacy routing scored 2/12 and normalized routing scored 1/12, losing one previously correct missing-information answer. No checkpoint was adopted. Read the separately scoped NewBrain tests →
NewBrain Actual51: a further matched development comparison tested a different 12-question partition (10 factual, two missing-information). Both the unchanged reference and routing-only adaptation scored 0/12; full adaptation scored 2/12, with both correct answers being missing-information controls. No factual answer was correct. The experiment did not improve Original12 and did not meet its preregistered success criterion. New complete-answer diagnostics were subsequently written and source-reviewed, but have not been executed.
Avatar Builder A44/A45: two new 3,086-vertex cropped torso controls produced 12 actual neutral views; the revised rear silhouette was softer, but chest and breast contours remained visually rejected. An original 16,385-vertex blank head passed bounded surface/derivative checks and produced 14 matched Blender inspection views, while its skull, face and jaw/neck proportions also remained rejected. No complete Evelyn head, accepted torso, rig movement or functioning organs were produced. Read the head and torso review →
Limitations: stronger numerical routing influence is not correct free-generation; learning experiments have separate datasets and denominators, including an existing-content Original12 assessment and a different context-conditioned development set. Private model inputs, keys, images and individual anatomy references remain unpublished. The accepted Original12 checkpoint is still 2/12. All unaccepted component experiments and failed variants remain preserved.
NewBrain and Kira World Avatar Builder · October 28, 2026 · Later executed, reviewed experiments
NewBrain's routing branch activates, but answer accuracy does not improve; original body geometry remains unaccepted
NewBrain: a read-only examination identified an exactly zero routing projection in the accepted original checkpoint, while its recurrent path remained active. A later, separately scoped 24-prompt development comparison gave routing-only training 0/12 answerable and 4/12 missing-information answers before and after 384 updates. A matched full-parameter arm reached 1/12 and 6/12, but lost two previously correct abstentions; the success gate failed. In a distinct Original12 retention comparison, an anchored candidate scored 1/12 against a 2/12 control. The accepted Original12 checkpoint remains 2/12; no candidate was adopted.Read the distinct NewBrain tests →
Avatar Builder: 16 targeted head checks and 14 saved views examined another original blank head (16,385 vertices, 16,320 faces), but its cranium, face and jaw/neck were again rejected. Two independent original cropped-torso controls each produced 3,086-vertex, 3,000-face geometry, six clay renders and an editable Blender file; eight targeted software checks passed. Curved support reduced rectangular attachment artifacts, while conical front and angular rear forms kept both torsos visually unaccepted. No complete body, integrated eyes, motion or functioning anatomy is claimed. Read the Avatar Builder inspection results →
Evidence boundary: the routing work used development data distinct from the protected Original12 suite, and historical global-disjointness checks remain incomplete. Private model data, source recipes and anatomy images stay off the public site. These are reviewed internal engineering experiments, not externally reproduced research or a product release.
NewBrain and Avatar Builder · October 28, 2026 · Source-only follow-up, not a completed test
Next learning and head-profile experiments prepared, but not yet run
After the latest 2/12 Original12 outcome, Codex prepared a full-sequence teaching objective with a frozen original-model reference intended to protect earlier correct answers. This is source-only work: direct training integration, mathematical checks, new model training, protected-question scoring and restart measurements have not been run. There is no new accepted NewBrain checkpoint or score improvement. Read verified NewBrain results →
A separate original Avatar Builder head-profile source design now provides three individually controlled corrections for chin projection, the under-chin contour and neck recess. It is intended to address the prior trial's nine unchanged depth measurements, but has not been executed, rendered, fitted to a body or visually accepted. The October 28 head remains rejected; the standalone accepted eye is a different component, and no completed body or functioning anatomy is claimed. Read measured head results →
Kira World / Avatar Builder · October 28, 2026 · Original head correction reviewed
6,324 vertices widened, but nine profile measurements show no depth correction
The latest original head adjustment passed 65 software checks and saved 12 neutral renders and three editable Blender files. A bounded correction changed 6,324 vertices by as much as 6.944 mm, but all nine measured front/back profile sections had zero depth change. The owner and independent reviewers REJECTED the remaining bulbous cranium, flat elongated facial plane and unnatural under-chin/neck transition. The neck boundary is preserved; seven planned female-torso calls were not run. No accepted finished head or body is claimed. Read measured technical report →.
NewBrain · October 28, 2026 · Reviewed Original12 weighted attempt
No net score improvement: two gains replaced two lost correct answers
The original 11,914-parameter neural core completed one fixed weighted teaching intervention (384 updates, 2,304 presentations). The candidate remained at 2/12 on Original12, having gained two substantive answers but lost both previously correct abstention controls. A distinct-process zero-update restart reproduced the final results. No checkpoint was adopted. The intervention also added development-data exposure, so the cause of the changes cannot be isolated to token weighting. See the full evidence →.
NewBrain · October 27, 2026 · Separate original-core synthetic acquisition
Original core learned 11/12 direct synthetic answers in a distinct task
NB21 used isolated copies of the 11,914-parameter, 42-token original decoder and showed 0/12 → 11/12 direct and 0/12 → 4/12 reordered complete-answer gains on fresh known-token symbolic associations. All 48/48 frozen/trained answer records reproduced after restart, with no additional teaching. Delayed retention and Original12 improvement were not demonstrated. Review the differences →.
NewBrain · October 27, 2026 · Separate completed NB15/NB16 tests
Varied teaching and broader OLD replay fail their generalization targets
NB15 varied training scored 7/24 on held-out NEW third-form questions against 9/24 with direct-only training. A different NB16 experiment broadened OLD replay and scored just 8/24 versus 7/24 on held-out NEW third-form questions, below the required +6 while losing three NEW direct and three NEW alternate answers. These were independently reviewed saved-state results and exactly restarted, not accepted NewBrain checkpoint improvements; Original12 remains 2/12. NB14's earlier positive gain remains separately preserved. Research report and denominators →
Kira World / Avatar Builder · October 27, 2026 · A16 scoped executed results
Ear triangle aspect improves but original Evelyn head still fails visual review
A16 completed 31 head checks and three ear checks and retained 12 inspected renders and an original editable head with 16,385 vertices/16,320 faces. A small-ear triangle aspect defect fell from about 333.95 to 26.48/37.33. Reviewers still rejected the jaw-to-neck shape and ear appearance; recognizable Evelyn likeness and body integration are not accepted. Read A16 limits →
Value-focused teaching adds 17 correct new answers across six fixed worlds
Codex ran a preregistered, matched training comparison emphasizing the factual answer value (fixed 1:1:4:1 token loss weights, normalized). Newly taught direct questions improved 11/24 → 23/24, alternate question forms 2/24 → 7/24, and older questions 29/48 → 43/48. Every world improved in new-answer count, but two worlds still scored 0/4 on alternate forms and one previously correct old answer was lost.
Primary and independent saved-evidence reviewers checked the results; 288 free-generated answers reproduced exactly in a separate process without extra training. This is a narrow 8,343-parameter synthetic reference learner, not accepted Original12 (still 2/12), broad intelligence, Qwen equivalence or an independently reproduced external scientific finding. Read the technical report →.
Kira World / Avatar Builder · October 26, 2026 · Executed A15 visual review
Original blank head renders in twelve views, but remains rejected
A corrected original Blender A15 run saved an editable blank-head mesh and 12 actual views (16,385 vertices / 16,320 faces, one mesh) inside its numeric scene budget. Primary and independent reviewers rejected flat facial shape, capsule skull and blocky jaw/neck. No final original Evelyn likeness, fitted face, working full body, animation or physiology was demonstrated. The accepted standalone eye is different; private renders remain unpublished. Read honest construction limits →
NewBrain · October 26, 2026 · Completed separate six-world follow-up
Partial subject learning with measurable older-answer forgetting
Teaching a separate small synthetic reference improved new direct answers 0/24 → 6/24 and alternate forms 0/24 → 5/24, while older answers declined 33/48 → 30/48. Three worlds gained no new correct answers. Four old answers were lost; a fresh process reproduced 192/192 complete saved-state answers. This is not the later NB14 value-weighting success and not accepted Original12 (still 2/12). Detailed evidence →
Kira Labs · October 26, 2026 · Public research film
Watch the completed 3-minute Kira Labs fundraising film on our website
The completed 191.54-second film is narrated by Evelyn in her approved synthetic voice. It combines the accepted standalone original eye renders, labeled licensed developmental footage, future concept visuals and research limitations. Both 720p and 1080p original release encodes are preserved; the website supplies a direct 720p player and 1080p download.
The accepted NewBrain Original12 remains 2/12; no completed original Evelyn body, medical function or new lip-synchronization experiment is claimed. Play the film and see support options →.
Kira World / Avatar Builder · October 26, 2026 · Original eye component
First standalone original eyeball accepted; nose and ear work still fails or lacks visual acceptance
The bounded generic eye study passed 23/23 software tests, then Blender 5.1.2 CPU generated 18 static views plus editable procedural geometry, totaling 10,211 vertices, 10,304 faces and three meshes. The engineering record says the owner accepted the standalone eye appearance. This does not mean it was fitted to Evelyn or could blink, move or see.
The original eye includes a corneoscleral globe, annular iris with actual pupil aperture and recessed dark cup; the latter is not a functioning retina. Brown, blue and green examples were rendered; six generic colors were included in software checks. Corneal reflectivity and scleral details remain artistic. Subsequent nose software recorded 47 passes/18 images but three nose appearances were rejected; ear software recorded 18 passes/10 images but still lacks accepted realistic form. Torso had 17 passes and one failure, with both rendering branches unrun. The failed V9/V10 heads and unchanged NewBrain Original12 2/12 remain separate. Full results and limits →
Kira World / Avatar Builder · October 26, 2026 · Tested tools and anatomy research
34 software tests pass and manual face references process; V10 head still fails likeness review
A new original reference-analysis and body-profile package passed 34/34 scoped software methods. Its first two-image integration failed when saving along a long Windows path; a separately preserved fresh run with the same input/source and a short absolute output path succeeded. It retained 38 facial points (33 visible, five occluded) and one hair marker. Neither nonneutral image passed neutral selection. Automatic face/video tracking, camera calibration and held-out likeness reconstruction are untested in this record.
The anatomy research graph has 217 nodes, 284 typed links, 473 unbound ports and 21 pending relationships. Female pelvic atlas mapping is still 13/28, and four important outlet/support anchors lack fitted positions. Original body design now has 26 independently specified shape targets and 17 regional material fields; they remain recipes rather than working body geometry, fitted organs or digestive/urinary/reproductive functions.
The separate, originally authored Evelyn V10 head produced four neutral clay renders and editable Blender geometry. Its 146,194 vertices and 146,156 faces satisfied declared V10 scene budgets, yet visual review rejected the facial form and likeness; actual nostril and external ear-canal openings were not built. Earlier V9 also remains rejected. This does not demonstrate an accepted fictional-character head or improve NewBrain Original12 (still 2/12). Private images and rejected renders remain unpublished. Full Avatar Builder status →
NewBrain · October 26, 2026 · Verified three-world learning and restart
NewBrain's small test learner acquired new answers and transferred some question forms
In a separate fixed 39-token, 8,343-parameter experimental learner, complete new-fact answers improved from 0/12 frozen to 6/12 taught, with alternate-form transfer improving 0/12 → 4/12. Old answers went from 15/24 to 16/24, but four previously correct old answers were forgotten and five previously wrong ones recovered. All 540 final answers reproduced after a distinct-process restart with no additional learning.
Supplying additional retrieved context did not automatically improve performance: raw history scored 4/12 on new facts, and even a single correct oracle event scored 3/12 versus 6/12 without context. This does not isolate a specific neural mechanism. The separate accepted Original12 result is still 2/12, with the prior failed trained 0/12 preserved. Read complete results and limitations →.
NewBrain · October 26, 2026 · Verified six-world replay-form study
Changing old-question schedules did not improve new-fact or transfer accuracy
Across six fresh closed-vocabulary worlds, synchronized and staggered replay each produced 10/24 newly taught answers and 9/24 unfamiliar question forms. Old-answer totals were 26/48 and 25/48; both lost the same seven formerly correct answers. The difference is one less recovered old answer, not additional forgetting. All 324 final answers matched on restart.
Saved diagnostics locate all wrong final outputs at the value token; all outputs ended with true EOS. A new subject-sensitivity probe is proposed, not performed. This does not prove every replay method is equivalent or change accepted Original12. Read tables and next research question →.
Kira World / Avatar Builder · October 26, 2026 · Original human-head construction experiment
V9 original head geometry rendered, but its facial appearance was rejected
Kira Labs developed original procedural cranial and facial geometry, then produced four neutral clay views and an editable Blender head. Independent visual reviews found eye-socket gaps, weak nose and midface, shelf-like lips and incomplete ear structure. The combined head-and-eye rendered scene also exceeded the proposed full-scene vertex and face limits. These defects remain uncorrected in this V9 attempt.
Private review images are retained rather than presented as finished character media. The experiment does not approve an Evelyn likeness, animation, speech, anatomy or a release-ready avatar. Earlier licensed rig demonstrations are separate. See the current Avatar Builder research status →.
NewBrain · October 25, 2026 · Five-world measured learning
Balanced replay retained more old answers, but generalized less to unfamiliar question forms
Five fresh synthetic worlds: balanced replay scored 38/40 old answers versus blocked training's 14/40, but 7/20 held-out composition versus 12/20. Both gained 13/20 new answers. Fifteen saved states reproduced 270 complete answers on a separate-process restart.
These small results concern a separate association learner, not accepted Original12 (still 2/12). Both primary arms failed unseen abstention, presentation and mixing were not separately controlled, and no broader intelligence claim is warranted. Read per-world scores, methods and limits →
Kira World / Avatar Builder · October 25, 2026 · Reviewed public-censored motion
A real five-second sitting-rig diagnostic is ready for viewing
The 120-frame, 24-fps silent sitting test uses a licensed developmental reference and public opaque privacy coverage. It shows controlled scripted movement, not autonomous motion, natural contact, tablet handling, clothing simulation or Evelyn's final Avatar Builder-authored body. Arm posture, hair and physical realism still need work. Watch the reviewed public derivative →.
Video Studio and Voice Designer · October 25, 2026 · Narrow software test
Video Studio adds managed Voice Designer runtime discovery
Seven focused path-selection checks passed for using the separately installed original-voice runtime while respecting manual Python/model overrides. Model inference was not executed in these checks.
Voice Designer 0.1.6 is the only standalone release planned for the older generator: the existing free unsigned preview is that edition, not a stepping stone to a later polished final. Kira World intends to replace the generator with a separate virtual Voice Box in future voice research, rather than issue a 0.1.7; the Voice Box is not yet a qualified working replacement. Access the existing installer →.
Evelyn Hart / Head, eyes and conversation · October 25, 2026 · Published actual recording
Evelyn's four generated replies now accompany a 58-second head-and-eye video
The exact reviewed 58.125-second, 768 × 768, 24 fps MP4 is now playable from the public Evelyn Avatar & Movement page. It contains four local-model replies in Evelyn's approved synthetic voice and an edited, offline sequence of head turns, eye directions, blinking, held eye closure and recognizer-timed jaw movements. The uploaded recording matches the retained source-media SHA-256 exactly; the assembled clip includes all 1,395 frames.
Limitations and observed bugs: the prompts and head/eye movement were planned before animation, not controlled autonomously by NewBrain or an interactive microphone. An obvious hair/scalp seam can appear when turning, and some technical replies promise actions “later” while the video already shows them occurring. Prepared fixes still require a subsequent reviewed test. The video does not demonstrate response latency, natural hair and body mechanics, broad factual understanding, book reading or a finished Evelyn avatar. Accepted NewBrain Original12 remains 2/12.
Voice Designer · October 25, 2026 · Unsigned backend preview
Voice Designer installer preview passes setup, reinstall and local speech tests
October 25 final-version decision: Voice Designer 0.1.6 will be the last edition of Kira Labs’ original-voice generator released as a standalone legacy tool. The existing unsigned 0.1.6 preview is the only standalone Voice Designer release Kira Labs plans, not a stepping stone to another final edition. Kira World is developing a separate virtual Voice Box instead of extending this generator; no 0.1.7 is planned. A free unsigned 0.1.6 installer and corrected source ZIP have been uploaded to the public Kira GitHub repository for review. The retained test record documents provisioned Python/packages/model, 11/11 model files verified, install, uninstall, reinstall and two different WAV outputs from the installed command-line backend, 2.72 and 3.92 seconds long. The earlier saved WAV remained unchanged after reinstall.
Release boundary: the new native GUI was not opened or button-tested, clean Windows installation is unverified, the installer is unsigned, and the observed 32 GB RAM / 16 GB NVIDIA GPU host is not a hardware minimum. No private or pending-approval emailed voice sample was published here. Read actual progress → · Read current guidance → · Uploaded files (public GitHub) ↗
Evelyn Hart · October 24, 2026 · Approved original recording
Evelyn's voice is now recorded, approved and publicly playable.
Kira Labs' founder approved Evelyn Hart's original voice greeting and confirmed its use by Codex. The genuine 3.04-second WAV says “Hello, I'm Evelyn Hart. And I'm glad to meet you.” Its Git blob matches the previously private retained audio, and the recording is now available for public listening on Evelyn's homepage.
She is the last planned Kira Labs synthetic individual to use this generation of the original voice generator. Future synthetic individuals will move toward replacement technology, while Video Studio retains a version of the earlier generator. The free final 0.1.6 Voice Designer developer preview is now downloadable; the unsigned GUI and broad release qualification remain unverified.
The earlier failed recorder-wrapper run and its source checks remain historical evidence. This approved greeting is a separate milestone, not proof of open-ended conversation, repeated generation or synchronized avatar animation. Listen and review the milestone →
Kira World / Person Creator · October 24, 2026 · Public naming decision
Temporary Creator becomes Person Creator—because not all created individuals are temporary
Kira Labs adopted Person Creator as the new public name for the research previously called Temporary Creator or Temporary AI Creator. The broader name recognizes distinct synthetic experts, companions and characters, including individuals intended for short tasks and others intended to continue with their own identities and memories.
Temporary or continuing describes an individual, not the tool. Person Creator is not named after Kira because she is one individual within Kira World. Historical video titles and internal technical identifiers remain where needed for source provenance and compatibility.
What did not change: no new full automated creation, live unscripted conversation, long-term identity persistence, natural body or voice integration, public installer or NewBrain intelligence is being claimed by this name decision. Those remain separately testable development objectives.
Thirteen original-voice model files verified; recorder startup blocked before speech
A recorded original-voice attempt completed model-file verification: 13 of 13 VoiceDesign files, totalling 4,520,163,832 bytes, passed full hashing. Interpreter and selected output metadata passed their preparation checks. The worker nevertheless never started, because the recorder detected process generations it could not fully account for and correctly refused admission.
Both owned Jobs finished empty and 16 retained known process handles closed; reviewers agreed this was a failed attempt. No Evelyn WAV, audible original voice, real filmed actions or lip-synchronized output was created. The next narrow step is a no-model recorder-startup diagnostic that preserves identity and cleanup safeguards; local work on later attempts is not proof of success until separately reviewed.
The separately admitted, frozen arithmetic verifier ran once against the previously retained 312,524-byte numerical report. Independent and primary reviewers accepted the verifier's own process and CPU/memory guards, preserved output, evidence custody and cleanup. The more stable loss/gradient formulation matched independent 120- and 200-digit Decimal references within the prescribed relative tolerance of 1 × 10−12 for nonzero quantities, over seven synthetic families, five confidence margins and three logit shifts.
The conventional baseline rounded a mathematically nonzero target gradient to zero in 30 of 105 cases and a nonzero loss to zero in those same 30, at margins 40 and 80. Nine archived invalid-input rejection receipts were verified in the report but not independently rerun. The verifier's accepted CPU evidence does not fix the historical numerical producer's missing CPU-affinity receipt; whole physical OS I/O and outer orchestration remain unverified.
Scope: this diagnoses the isolated synthetic arithmetic behavior, not NewBrain's real learning failure. Frozen learner weights were not changed, no protected questions were used, and Original12 remains 2/12 (the previous trained 0/12 failure and unscored zero-teach control are preserved). No retention, transfer, restart or fair Qwen-comparison gain is established. Read the verified numerical result → · Verification source record (repository access may be required) ↗
Evelyn Hart · October 24, 2026 · Public release information and draft library
Evelyn's release roadmap and troubleshooting preview are now public—with no installer claim
Kira Labs now has dedicated web pages explaining the October 23 two-page release requirements and four-page installation/troubleshooting draft. The new pages outline intended signed distribution, privacy and editable books, developing memories, owner backup/export/uninstall rights and the subjects a final support guide must cover. They remain explicit pre-release previews, not tested installation instructions or an available download.
A later working installer-catalog report identifies 30 selected publications (13 technical and 17 other) plus two original technology/AI primers. The library's 66 listed files include licenses and provenance, not 66 books. Redistribution rights, installed library discovery, page-based reading, original voice, body movement, memory recovery and complete clean installation remain unverified. Read public release requirements → · View troubleshooting preview → · See the dated working book catalog →
NewBrain · October 23, 2026 · Executed isolated numerical diagnostic
A numerical reliability run completed; independent arithmetic verification remains pending
The previously source-reviewed public numerical experiment ran once, producing a retained 312,524-byte report. The command exited with code zero and no pending session. The producer reports 105 synthetic arithmetic cases and nine invalid-input checks, without any model generation, parameter update or use of protected benchmark questions. Independent and primary reviews accepted the saved normal-process identity, original result preservation and cleanup; the separate arithmetic verifier has not run. The returned values must not yet be treated as independently established numerical correctness.
Review also identified an outer CPU-affinity evidence gap: the monitor lacked its own affinity installation/readback record, although child Job/cleanup data was retained. A new narrow CPU guard and separate saved-report verifier have passed source review, not runtime. The finished run will not be replayed; numerical verification must use its saved report with separately authorized checks. Neither this engineering test nor a correct gradient would automatically demonstrate better complete answers. Original12 remains 2/12.
Selective privacy covers now follow the rendered body and camera in source
For future authorized adult anatomy evidence, the proposed pipeline preserves a private original and produces a separate censored public image, video and thumbnail candidate. New selective cover code is designed to move with the rendered subject and camera instead of relying solely on a blacked-out frame. Static evidence is not proof of complete visual coverage in motion.
Two picture-metadata integration defects have source corrections, while a separate private-stream finalization issue has a correction awaiting exact review. Full capture, all-frame privacy inspection, durable encoded-frame evidence, source-file preservation and natural movement are still unqualified. Original-voice callers are in preparation; one cleanup defect remains under review. No new approved original voice, anatomy photo or actual movement recording resulted from these code changes.
NewBrain · October 23, 2026 · Independently reviewed synthetic control
The isolated normal execution control passed. Learning accuracy is unchanged.
An independently and primarily reviewed zero-model normal synthetic control completed a single owned invocation, exit code zero, with no pending session. Its 84,826-byte saved result matched 60,000 ASCII characters followed by 4,096 Greek-letter/emoji pairs. The reviews authenticated 371 Native events, 183 byte cells, 21 once-closed handle generations and empty owned Jobs; recorded startup, output and cleanup were within the observed limits.
This is a step forward in reliable execution, not a model-learning result. The full outer orchestration envelope, fault cases, physical OS I/O, NumPy/numerical readiness and private scoring still need separate evidence. Original12 remains 2/12, the failed trained 0/12 checkpoint remains archived, and the distinct 20-generation zero-update control is still unscored. No protected answers were used for training. The normal-control slot is consumed and cannot be repeated.
NewBrain · October 23, 2026 · Numerical accuracy research / planned experiment
Floating-point gradient reliability has a new preregistered diagnostic.
Source review accepted a bounded public numerical diagnostic with 105 synthetic cases and nine invalid-input cases, covering target-confidence gaps 0, 10, 20, 40 and 80, output masking and high-precision Decimal reference calculations. A separate report auditor was prepared to check arithmetic and receipt consistency. The frozen learner and protected assessment were not changed. The diagnostic was subsequently executed once and the normal process/custody result reviewed, while the separate independent arithmetic auditor has not run; no numerical learning bug or answer improvement has been established.
Six arXiv preprints were reviewed for hypotheses about numerical errors, raw-history retrieval, N-gram writes, harness distillation, episode memory and image grounding. They are external reported findings, not reproduced Kira Labs tests. The strongest near-term idea is to check the numerical training signal without changing the frozen model. Further learning, retention, transfer and separate model comparisons must follow on fresh test admissions.
Evelyn / Avatar Builder · October 23, 2026 · Reviewed body and app integration
Before-action motion recording and original voice integration are closer—but still gated.
Independently reviewed source now starts capture before the body/viewer activity, observes stream callbacks and requires acknowledged recording data, while retaining actual original and partial records for later review. A blocking mismatch was found between absolute executor evidence paths and relative accounting paths; it must be fixed before an admitted capture test. The planned silent status AVI and body-canvas WebM are separate artifacts, and stream callbacks alone cannot prove saved frames.
Voice/app source reviews found no current blocking defect in selected streaming-observer and bootstrap portions, but full production process ownership, startup, audio playback and recording resources remain unqualified. The originally intended adult female voice has not been generated or auditioned. No new movement footage, voice recording, natural clothing, body-export acceptance or verified Patient Horizon copy came from this review.
Evelyn · October 23, 2026 · Verified documentation and source review
Six release-document pages prepared; movement and voice code reviewed
The October 23 engineering report records two completed draft documents: two pages of official-release requirements and four pages of installation and troubleshooting guidance. Reviewers visually inspected all six PDF pages and checked the prepared PDF, TXT and HTML copies in the university folder. These materials are not a tested public installer or a release-ready support guide. The actual guide must be checked against the final installer, include tested system requirements, and be supplied both on the website and offline with the eventual download.
Evelyn's female movement source now composes rest-relative offsets with blended poses and retains breathing timing across movement routes. The original-voice execution bridge also passed a code review of its process-handle retention correction. Both are source-level progress, not successful recorded movement or speech. Her intended original adult female voice has not yet been generated, played, recorded or approved. A generic or donor voice is not an approved fallback.
Full anatomy, natural movement, clothing behavior, visitor memory, reading, qualified voice, signed packages, privacy controls and clean installation remain release tests. No new patient-body copy to Patient Horizon is claimed. Read the release packet progress → · Proposed safeguards →
NewBrain · October 23, 2026 · Latest reviewed execution gate
Metadata loading progressed; a freshness guard stopped the next launch
A revised local-file module loader prepared and validated the expected ownership-check metadata. A subsequent automatic attempt stopped at the initial context-freshness guard after its fixed validity period had expired, before the archived-result check, server launch or existing client callback. Independent and primary reviewers preserved the failure. They did not independently authenticate the exact guard-entry time or complete submitted code envelope, so no successful process-ownership result or universal absence proof is claimed.
The proposed follow-up combines new request construction, preparation, lossless original-result preservation and dispatch within one bounded orchestration call while retaining the existing guard. No learning model or Evelyn body, voice or conversation test ran during this stop. Main Original12 remains 2/12; the earlier trained 0/12 failure and the separate completed but unscored 20-generation control remain preserved. See NewBrain's verified control and limits → · Dated review record (repository access required) ↗
Evelyn Hart · October 23, 2026 · Public-release goals · Not yet implemented
Keep a verifiable starting identity while allowing independent personal development
Kira Labs has outlined a release design in which the official application, initial identity configuration and voice defaults are authenticated, while Evelyn can develop interests, memories, preferences and revisable beliefs through experience. A signature or integrity check could help distinguish official software from modified copies, but would not prevent determined modification by someone controlling the local computer.
Proposed continuity controls include memory sources, dates and correction history; recoverable local backups; and careful separation between books, factual evidence, fictional narrative and executable instructions. New installations are intended to start independent personal histories, without copies of private visitor conversations, memories or personal reference files. The design includes the ability to decline or end an interaction and does not call for covert monitoring.
Installers should retain control over their data, backup/export, deletion and uninstall options. A readable installation and troubleshooting guide is planned for the eventual public release; it is not yet an available installer or a completed support guide. Security, autonomy, real speech, memory continuity and clean installation still need end-to-end tests. NewBrain remains the first research priority. View Evelyn's planned safeguards →
NewBrain · October 23, 2026 · Latest reviewed engineering limitation
Ownership-check preflight stopped before components launched
A new process-ownership test attempt stopped in the JavaScript orchestration host because the chosen module-loading format was rejected. Independent and primary review documented that the constructor, server and client did not launch. All failed originals were preserved and the attempt was closed; a corrected module loader will require a fresh authorized attempt.
This is a source/infrastructure issue, not a failed scientific learner outcome. The original 20-generation, zero-training-update control remains completed but unscored. Accepted Original12 is still 2/12; no new learning, voice, body or movement result occurred. Read the dated research record (repository access required) ↗ · Saved control result →
NewBrain · October 23, 2026 · Reviewed saved control
After correction of the Windows metadata handoff, the fresh no-teaching control completed 20 generated responses, zero training updates and zero scoring calls, with ten observations before and ten after. Independent review authenticated the original report, process identity and saved cleanup. No candidate checkpoint was produced, and this consumed test will not be repeated.
The preserved answers are not yet scored for accuracy. Main Original12 remains 2/12, with the separate failed trained 0/12 state retained. The October 22 public B improvement belongs to another three-item experiment. Runner and process-ownership code subsequently passed source review, but its new runtime controls have not yet been independently qualified.
The next scientific step is authorized scoring of the saved control, then separate paired learning, transfer, retention and restart testing. Literature leads are hypotheses, not NewBrain results. See the engineering milestone and limits →
Evelyn and Avatar Builder · October 23, 2026 · Source reviews
Body export, library import and unique voice preparations advanced
The source-reviewed female body builder specifies 25 joint chains, 163 rig bones and 139 skin-weight groups. A six-input export-parity auditor checks declared rest/skin attributes. No new qualified body export, natural movement, clothing motion or verified copy to Patient Horizon exists.
Two Add Document source corrections preserve custom library folders and supplied book titles/dates while discovery proceeds. Evelyn's 30 publications and two original primers remain assembled files, not personally read books. Original adult-female voice audition and recording code is prepared, with a 55-word sample and an intended physical-speaker playback test. No unique voice has yet been generated, heard, approved or recorded, and no live dialogue or clean installer test has passed.Read the evidence and limits →
Evelyn · October 22, 2026 · Local library engineering
22 complete publications assembled and hash-checked; no reading or installation yet
Codex's dated local assembly report confirms 11 technical PDFs and 11 varied UTF-8 titles: 88,949,725 original bytes. All 22 ordinary-file destinations passed complete SHA-256 readback. With their catalogs, rights notices and provenance, the candidate contains 45 normal library files plus an installer manifest. The previous seven-title starter batch remains preserved.
All titles remain NOT_READ_BY_EVELYN. The Windows Documents installation, PDF page extraction, book discovery timing and release rights still need qualification. Eight additional technical books have noncommercial conditions and eleven Gutenberg editions have U.S. rights/trademark conditions. These books are not now publicly distributed. Read the actual title lists and evidence limits →
Evelyn · October 22, 2026 · Proposed cultural learning topics
Classic film and television portrayals of AI enter Evelyn's research plan
Planned subjects include Star Trek, Star Wars, 2001, Blade Runner, Terminator, TRON, RoboCop, WarGames, Short Circuit, Knight Rider and Max Headroom. The aim is source-backed discussion of fictional identities, cultural fears and real technological history. No screen-fiction pack is included in the 22-book assembly; no media-copying rights or live conversational result is established. See the plan →
Evelyn Hart · October 22, 2026 · Additional software checks
Seven new isolated conversation checks passed; two source defects still require review
Evelyn's local typed software channels now distinguish spoken replies, private character thoughts/preferences and factual claims. Seven checks using synthetic inputs and fake providers cover cross-visitor separation, spoken-only display, queued rapid remarks and correction/yield/cancel behavior, bounded context, recording failure recovery and source/capture safeguards. Eight earlier app/library checks remain separately recorded.
Independent review identified a reference-date omission and a bypass of the interjection-length limit; source corrections later passed independent static review, but no live conversation or voice acceptance followed. These are software checks, not a verified live Qwen conversation, authorized Evelyn voice, realistic interruption or natural body movement. Local body-authoring source can verify a produced mesh/rig package and archive failed attempts; actual body creation remains unverified. Read precise scope → · Repository record ↗
NewBrain · October 22, 2026 · Planned controlled study
New-lesson replay/no-replay comparison implemented in source, not run
The next matched protocol will use frozen teaching and evaluation sets, equal training budgets with/without replay, unfamiliar-combination prompts, and restart retention checks. The four integration/scoring defects were identified before an admissible run; corrected source subsequently passed independent review, but the first unchanged-control execution is still pending. No new score results from this preparation: Original12 remains 2/12. Experiment status → · Status source ↗
NewBrain · October 22, 2026 · Verification of preserved B diagnostic
One additional complete answer after 256 updates; the control remains unchanged
The previously pending public B diagnostic now has an internally independently reviewed saved-run result: trained 2/3 → 3/3 strict complete generated answers; no-teaching control 2/3 → 2/3. Counts at fixed checkpoints 0, 64, 128, 192 and 256 were 2, 2, 2, 3 and 3. Exposure across three items was 256, 512 and 0. The training was not rerun during review.
Interpretation: this is a narrow positive training observation, not evidence of a completed coverage-versus-repetition comparison—the matching coverage arm did not qualify. No restart retention or held-out transfer was tested for this B run, no final state was promoted, Original12 remains 2/12, and Qwen remains untested. The older separate rehearsed 3/3 restart result is not the same study. Latest research status → · Dated record ↗
Coverage-versus-repetition diagnostic executed; independent learning result pending
After earlier guarded setup failures, the coverage-versus-repetition diagnostic completed execution and cleanup once. Its saved answers, state and controls await independent verification. No new learning accuracy improvement is accepted; Original12 stays at 2/12, and the distinct rehearsed 3/3 fresh-process restart remains limited to taught examples.
A complete execution is not held-out generalization or a completed Qwen comparison. Independent saved-output and control review is the next scientific step. Research status → · Primary source ↗
Evelyn Hart · October 22, 2026 · Scoped engineering milestone
Eight application checks and 19 verified library file moves
The private Evelyn app was assembled locally from 21 staged source files. Eight isolated software checks passed across syntax, recording-file mechanics, offline-provider logic, reading and restart handling, and staged install/removal protections. One recording check used synthetic video data rather than an actual character recording.
All 19 unsorted local PDFs were inspected and organized; destinations passed complete SHA-256 checks without overwrites, and a reversible journal was retained. No companion reading, new learning, real Qwen conversation, qualified voice, natural movement or public installer was demonstrated.Read the scope and release requirements → · KiraWorld record ↗
Avatar Builder · October 22, 2026 · Planned work, not a new body result
First complete female anatomy, then editable photo/video-guided variants
The current female reference is the intended starting body. Keep its fixed asset and narrow qualification intact while developing a separately gated generic parameter/rig exporter. Demonstrate that a new complete female body can actually be authored before checking several existing variants. Male anatomy and body-generation for Robert and other male characters are later, still-untrained phases.
Consented images would guide visible appearance and proportions, while videos could inform posture and movement. Inferred or unseen geometry must stay editable, with uncertainty shown; private reference imagery, recordings and body measurements are not published. Profile storage or source preparation alone is not evidence of finished body creation.
Planned recorded checks begin with relaxed standing, weight shifting, walking, sitting, repositioning, reaching and grasping, plus responsive head, eye and face. They are repeated after clothing is added to examine fit, folds, weight, drape and clipping. Failed and partial recordings remain labeled and archived. Evelyn's private installer and a small chair/table/tablet room come first. Other female characters, a small Kira World shell, male bodies and a future NewBrain switch require later independent integration.
Evidence boundary: earlier 512-vertex/five-pose restoration is a component test; the three-angle reference captures still failed side-frame acceptance. No new automatic body-generation, natural whole-body animation, clothes simulation or Patient Horizon clinical capability is demonstrated.
NewBrain · October 22, 2026 · Follow-up independent review
Execution-gate owner names need repair before a fresh learning run
A later diagnostic finished construction and metadata setup but stopped with an execution error. Nine caller files and the associated body/recording contracts received independent static review. A separate prelaunch audit confirmed the public execution gate contains older worker identity names that are incompatible with the current owner configuration. A narrow shared-source repair and dependent-reference review are underway.
The exact first exception in the stopped run has not been established, and the correction has not passed a dynamic Native qualification. No new learning or Qwen comparison was run. The accepted Original12 baseline remains 2/12, the older failed trained candidate 0/12 stays archived, and a separate rehearsed 3/3 after restart remains a different limited experiment.
NewBrain · October 22, 2026 · Guarded setup, not a model score
Three preserved setup stops; public A/B comparison incomplete
The comparison has not reached model training. A stale previous-run metadata reference stopped one attempt. A later attempt had a mismatched public-diagnostic descriptor alias, with five other frozen checks matching. Another attempt passed its memory gate at about 10.9 GiB available and admitted 103 metadata inputs (109 read rows), then stopped at a command-quoting check before builder or learner dispatch. The stopped identities and original records remain preserved.
Latest accepted Original12: 2/12. An earlier trained candidate fell from 2/12 to 0/12 and was not promoted. The separate rehearsed three-example exercise retained 3/3 after restart; unfamiliar-question transfer and Qwen comparison remain untested. These October 22 failures are execution-preparation findings, not fresh learning gains or regressions.
NewBrain · October 22, 2026 · Development goals through December 25
Acquisition, restart retention and unfamiliar questions are the next gates
A useful first learner by Christmas is a research target, not a release promise. Milestones include correct runtime supervision, a repeatable gain above the protected 2/12 baseline with frozen scoring and controls, retention after restart, transfer to new combinations and an eventual common-protocol Qwen comparison. A small conversational learning test may follow only after those results are established.
Original12 acquisition failed: trained 2/12 → 0/12, control stayed 2/12
The completed independent read-only scientific audit verified that the saved candidate received 256 updates and 768 example presentations. Its teacher-forced objective fell 4.068566 → 0.055734, but no new assessment answers were correct and both initially correct complete Original12 responses became wrong. The untouched control retained every answer at 2/12. This candidate failed the fixed improvement-and-retention rule and is archived, not adopted.
The earlier separate three-example trial retained 3/3 after a fresh-process restart because its answers were explicitly rehearsed; Original12 tested differently worded assessment questions. That contrast motivates research into teaching-to-assessment transfer but does not establish an architectural cause, an optimizer fault or a generalization capability.
The saved-only audit made no fresh training, inference or provider calls; it verified its scoped result, streams, arithmetic and cleanup. The candidate has not been promoted or cold-reload qualified. Exact bytes of one historical reminder-configuration file remain unavailable; recorded control history was reviewed, but this is not a full original-byte audit.
Next: a predeclared public synthetic comparison of taught prompts and unseen paraphrases, with stable baselines and untreated controls. Do not tune against protected Original12 assessment answers. The accepted main Original12 reference stays 2/12; Qwen has not been compared.
Evelyn Hart · October 21, 2026 · Identity and public presentation direction
Evelyn is being developed as a continuing synthetic individual, not simply a historian
Evelyn's expertise in AI and technology history is only one part of her intended identity. The design calls for personal interests, a growing history of experiences, memories of people and ongoing relationships. Her reading, humor and everyday conversations should help shape her experiences over time, beyond an expert question-and-answer system.
The visitor-profile storage reportedly already has fields for names, facts and shared experiences, and a new selector is being connected to them. These are engineering preparations, not a verified Qwen conversation or successful visitor recognition. Even passing a simulated-reply test would not prove real conversational continuity or reliable memory after an application restart.
Evidence boundary: separate head and eyelid checks passed; three real censored reference-body renders were saved, but the full side-camera acceptance test failed. A finished Evelyn installer, integrated voice, full-body performance and cross-visitor memory recall remain unverified. “Continuing synthetic individual” describes the research design and intended experience, not a scientific demonstration of consciousness.
John → Cindy → John: remembering different visitors without mixing their conversations
Evelyn is intended to remain active while visitors come and go, without presuming that the person speaking now is the person who spoke hours earlier. Hypothetical visitors John and Cindy illustrate the desired behavior: John talks and leaves; Cindy arrives and is greeted separately; John later returns and may continue his own earlier conversation after appropriate profile verification.
Planned visitor selector: an always-visible header above the conversation would read Talking with: Tom ▾, with previously saved profiles and a New profile choice. “Hi, I'm Tom” should offer selecting Tom's record or making a new one, not silently authenticate or expose stored private content.
Memory persistence goal: switching from Tom to Cindy should leave Evelyn's separately permissioned memories and experiences with Tom intact, and vice versa. Those histories should remain available to their respective authorized profiles following a complete restart; the menu and introduction phrase alone would never unlock a visitor's private conversations. The interface and this persistence behavior are still unverified.
Privacy boundary: Cindy must not receive John's private stored details, and an asserted name or similar-sounding voice must not authorize another visitor's saved memory. Evelyn should ask when attribution is uncertain, use consent and authenticated visitor profiles for private recall, and support a new or anonymous conversation when identity has not been confirmed.
Newly reported source-review concern: A reply started for John may finish after Cindy becomes the selected visitor. The proposed guard must check the authorized active profile and conversation version before displaying or storing the reply, and discard stale output. Extend the planned John → pause → Cindy → John test with delayed responses and inspect both the screen and saved memories. No runtime pass or merged correction is established by the supplied screenshot.
Not yet verified: No passing multi-visitor transition or post-restart memory-isolation test has been demonstrated. The required future check is John → Cindy → John with Evelyn continuously active and again after a fresh restart, including negative checks for cross-profile disclosure. This is a proposed behavior and acceptance criterion, not an implemented or released visitor-recognition feature.
Max and Adam: shared memories after someone dies, with permissions and provenance
Consider an example in which Adam talked with Evelyn on earlier visits and later died. Adam's son Max might ask Evelyn what she remembers from those encounters. The goal is for her to discuss her own actually recorded and permitted conversations with Adam, rather than claiming to reconstruct Adam's mind, inventing events, or presenting Max's newer recollections as if Adam said them.
Evelyn's proposed memory design should preserve the relationship between an event, the people present, who said it, when it was saved, the source and confidence, and what may be shared. Max's stories about Adam remain attributed to Max. Material Adam wanted kept private remains restricted; being family or learning of a death does not automatically unlock a private profile. If an original record is missing or unclear, Evelyn should acknowledge uncertainty rather than fill gaps with invented memories.
Required future test: synthetic Adam and Max profiles with shareable, restricted and conflicting details, followed by a post-restart recall test. Verify correct attribution, preservation of privacy, appropriate uncertainty and no cross-profile leaks. This is a design proposal—not a demonstrated memory-sharing feature or a passing test.
Evelyn Hart · October 21, 2026 · Actual saved body renders
Three full-body reference images captured; side-camera acceptance still fails
A controlled model-free test saved genuine front, side and rear screenshots of the existing adult-female reference body. All 28 meshes were visible, the camera images were distinct and the front and rear framing checks passed. The visible side view was retained, but its strict projected bounding-box check failed because four corners extended beyond the frame (1.116433 against a limit of 1). The overall test did not pass.
Public screenshots now use permanent opaque black privacy bars over sensitive areas. The source diagnostic body is reusable foundation work, not an accepted final Evelyn body or proof of finished movement, clothing, head/body rig connection, anatomy or the planned camera-to-face life-loop transition. Next: correctly fit the side camera and rerun all acceptance checks.
Evelyn Hart · October 21, 2026 · Approved avatar and life-loop camera design
Full-body Evelyn first, then a smooth zoom to her head when her life loop starts
Evelyn is intended to become a complete rigged 3D character, not a disconnected floating head. Her initial viewer should frame her full body. As her continuing life loop begins, the camera should move naturally in toward her face while keeping the same character and body in the scene.
Existing Avatar Builder research already provides a provisional body asset and skeleton, with independently reviewed movement and exact restoration of 512 sampled weighted skin vertices through five commanded poses. Evelyn also has actual accepted head-render and controlled-blink captures. These were separate bounded measurements, not a successful end-to-end full-body integration.
A full-body viewer and front/side/back capture source have been independently reviewed, and the actual three-angle captures have since run; front and back passed while the side-framing test failed. Clothes and fabric motion, body/head alignment, interacting with books, exact camera transition, original generated voice, continuing conversation and installer remain unverified. Reusable geometry/rigging can help Patient Horizon, but organ physiology must be separately tested.
NewBrain · October 21, 2026 · Accepted engineering launch-and-compatibility check
Corrected Windows launch succeeds with 26 of 26 expected compatibility outcomes
The page-aligned Windows process-limit correction now passed a controlled runtime and independent saved-evidence review. The host reported a 4,096-byte page size. Per-process and owned-job memory limits were 858,992,640 bytes, safely at or below the original 858,993,459-byte policy ceiling, and all six limit fields matched when first configured and before the child resumed.
With owned-job membership and a two-CPU affinity restriction checked, the child executed four reviewed pure code spans and completed 26/26 synthetic metadata compatibility checks: three valid records returned successfully and 23 deliberately invalid fixtures raised the prescribed ValueError. The child exited, both output streams reached EOF, the active-process count returned to zero, and owned handles were closed. A separate internal evidence reviewer accepted this bounded engineering result.
Limits: this did not score the saved Original12 training outputs, execute the complete learner/scorer, test every abnormal failure-control scenario, or prove generalization. Original12 remains 2/12; an older failed trained 0/12 state remains archived; the separate rehearsed teaching study remains 3/3 after restart; and Qwen is untested. The original 819-byte mismatch and its failed test remain in the next historical journal entry.
Historical next step at the 26/26 launch: independently audit the saved learning candidate without retraining. That separate audit is now complete and documents a 0/12 negative transfer result above.
NewBrain · October 21, 2026 · Reviewed Windows process-control diagnosis, not a learning run
Historical snapshot: this failed-limit diagnosis was subsequently resolved for the tested launch route. Read the later verified 26/26 result above. The historical measurements and pending-test language below remain as originally documented.
Windows memory-limit readbacks differ by precisely 819 bytes
A bounded runtime check supplied memory limits of 858,993,459 bytes for both the process and owned Job, while Windows reported 858,992,640 bytes for each. The other reported policy comparisons matched: structure size, flags, active-process count limit and CPU-affinity mask.
The strict readback stopped before child creation or resume: none of 26 compatibility cases executed, and no learning accuracy was computed. The handles were closed, while full process-family cleanup indicators remain unqualified. A whole-page rounding explanation fits the observed numbers but remains a hypothesis until system page size and the corrected runtime are measured.
An independently reviewed source correction now proposes querying Windows' page size, deriving a whole-page cap no greater than the configured policy, and preserving exact readback checks and existing safety bounds. Source review passed; runtime verification has not. Original12 is still 2/12; the previously failed 0/12 state remains archived; the separate rehearsed three-example study remains 3/3 after restart; Qwen comparison is untested.
NewBrain · October 21, 2026 · Trial 1073 and independent audit 1074 / reload 1075 and audit 1076
NewBrain learns a complete taught synthetic answer and keeps it after restart
A frozen, balanced 256-update three-case diagnostic improved 2/3 → 3/3 → 3/3 after a new process reloaded its saved state, with no new training. All three rows were explicitly rehearsed; the no-training arm remained 2/3. The newly learned response was cab word one with a normal end marker. The fixed checkpoints were 2/3 (start), 2/3 (64 updates), 2/3 (128), 3/3 (192) and 3/3 (256).
The previous 64-update curriculum is retained as a separate 2/3 → 0/3 failure. The read-only sequence diagnostic also identified two token decisions that were wrong; neither the diagnostic nor the later success establishes the cause of earlier training failures. The unchanged Original12 benchmark remains 2/12, with Qwen untested. The new result does not prove novel real-world learning or general language skills.
Evelyn Hart and planned Florence Nightingale · October 21, 2026 · Graphic preparation
Two original desktop identities, with one planned book library
The actual Windows shortcut artwork has been prepared and visually reviewed: a teal circuit-and-open-book design for Evelyn and a gold lamp design for Florence. Both SVG masters, PNG previews and multi-size ICO packages exist in the preparation material. No desktop shortcut launch, shared-library installer or Florence character has passed installation testing.
Evelyn HartFlorence Nightingale
The planned Windows setup would resolve the actual Desktop and Documents folders, launch each companion under its own identity and memories, and share one authorized Documents\Kira Labs\Library. Book discovery and per-person reading histories still need testing. Florence's historical-character work is planned for after October 29, rather than being offered as a download today.
Florence Nightingale · October 21, 2026 · Historical-source and application planning
July 31, 1890 selected as Florence's proposed knowledge starting date
A historical research plan now places Florence at age 70, immediately after her recorded address on July 30, 1890. Her documented London address was 10 South Street, Mayfair. An 1887 painted portrait, late-1880s photographs, dated correspondence and two renditions of the original voice recording offer references for a later Avatar Builder likeness and speaking-style study. These are archival references, not a completed character or voice reconstruction.
Her starting knowledge is to exclude events after July 31, 1890, including the date or circumstances of her eventual death. She would discover later information only when she actually encounters it through dated conversations or personally paced reading. A new library entry should reveal metadata, not unread book passages. Her experiences and reading memories remain separate from Evelyn's, even if both use the same library.
Period research distinguishes Victorian London from modern life—horse omnibuses, gas lighting, printed newspapers, telegrams and limited household telephones. The first deep-level electric Underground service did not open to passengers until December 1890, after Florence's proposed starting date. All historical companion behavior remains planned, not installed or validated. Work on Florence's 3D head, voice and character is scheduled after October 29. NewBrain is still the main research focus.
Project identity and selection controls pass a bounded set of tests
IdeaForge's project-ID review accepted 14 synthetic CPU and injected-interface checks around distinct numeric project folders, duplicate readable names, exact selection, refresh and concurrency. These are source/injected-workflow tests, not a verified complete GUI, invention simulator, language model, manufacturing process, or crash-recovery guarantee. The project is still in development.
Aster · Separate experimental model comparison
Added weight-protection penalty did not outperform replay
A frozen, five-initialization synthetic retention experiment compared ordinary replay with an empirical-gradient weight-protection penalty. All arms finished the tiny task with 6/6 old, 6/6 new and 8/8 wording checks, including the new-only control, so the experiment exposed no forgetting and provided no incremental evidence that the extra penalty helps. This negative result was preserved. It does not modify production Aster, establish generalization, or support adopting the penalty.
Evelyn Hart · October 20, 2026 · Run 1047 / independent review 1048
Eyes close and reopen: independently checked engineering pass
CONTROLLED BLINK AND REOPENING PASSED
Following the earlier negative blink study, a fresh controlled test with the unchanged 65% acceptance gate measured 69.8719% left and 72.4625% right eyelid closure. Both eyelids visibly shut and reopened in recorded original captures, with zero reopening error and sampled geometry returning to its starting pose. The separately reviewed saved evidence covered all 15 PNGs and associated browser responses.
The first successful joint eyelid setting was 20°, after smaller settings failed. Recorded changes included 1,492 blink pixels and 1,024 mouth-control pixels. The audio played was a test signal, not Evelyn's speaking voice. Natural conversation, generated voice, phoneme synchronization, and post-restart visitor memory are not yet established; creator-led private installation and conversation testing precede any public download.
Evelyn Hart + NewBrain · October 20, 2026 · preparation and review
Evelyn's blink correction reviewed; next test still gated by available RAM
SOURCE REVIEW PASSED · NEW MOTION TEST UNRUN
After Evelyn's motion run failed at approximately 1.68% left / 1.74% right eyelid closure (65% required), a source review identified that calibration had scouted a movement of only 0.5°. A revised search for enough eyelid travel was prepared while retaining eye stability and rest-position checks. Source review passed, but the follow-up fresh test could not start while available RAM remained below the 8 GiB startup guard. Jaw calibration was accepted in the earlier failed trial; no generated speech or completed motion video is verified.
NewBrain's next diagnostic was also stopped before model execution
The supplied task screenshots show a low-RAM startup refusal and a source-closure metadata-field error in separate preparations, both preserved. The corrected call passed source review but no new model learning score was produced. Original12 remains 2/12; the earlier short three-question 3/3 restart result remains separate.
Evelyn Hart · October 20, 2026 · separate rendering and motion trials
Three static head angles verified; blink calibration fails in separate trial
STATIC RENDER PASS · MOTION GATE FAILED
An owned-browser test captured Evelyn's visible face from front, three-quarter and side and passed its scoped saved-evidence checks. This is a simplified static head preview, not final photorealism. In a separate motion trial, both eyelid apertures reduced by approximately 1.7%, below the fixed 65% requirement. Jaw calibration passed, but the integrated six-capture sequence and video were not produced.
NewBrain · October 20, 2026 · controlled training reports
Small rehearsed learning persists after restart; longer-answer acceptance fails
SHORT 3/3 VERIFIED · LONGER 2/3 FAILED
A fixed, explicitly rehearsed three-question test improved 2/3 → 3/3 after 64 updates and stayed 3/3 following a distinct fresh-process restart with no further updates. The no-teaching arm stayed 2/3; held-out generalization remains unproven.
A different longer-answer test stayed 2/3 → 2/3: the model generated the correct first answer token, then stopped before the required source/version tokens. A separate five-question pilot regressed 2/5 → 0/5 → 0/5 after restart. The main Original12 score remains 2/12 → 2/12 with no net gain, and the Qwen comparison is untested.
Patient Horizon · October 20, 2026 · research plan
Synthetic patient record and lung demonstration: tests still proposed
The new report proposes a versioned patient record, traceable synthetic lab/scan findings, timestamped simulated vital signs, a transparent anatomical view and later calibrated respiratory simulation. New lung-plan acceptance tests remain unrun. Existing bounded software evidence includes four synthetic request-planner tests and 512 weighted skin vertices moved and restored through five poses, not clinical or physiological validity.
NewBrain · October 19, 2026 · frozen numerical follow-up
New sampled-derivative pass: 80/80 coordinate and 4/4 direction comparisons
PASS_SAMPLED_DERIVATIVES · scoped internal audit
The new run performed 170 objective evaluations across two declared probe states. All 80 coordinate and four direction comparisons passed, all ten parameter groups including Q/K/V met the unchanged 1e-6 signal floor, both deliberately incorrect-gradient controls were rejected, and maximum absolute comparison error was about 2.08 × 10⁻⁹.
The saved file/execution and cleanup audit passed, preserving original model checkpoint parameters and four training counters. No training updates, answer generation or scoring calls occurred. This supports sampled derivative accuracy, not a verified optimizer fix or an improvement to learning.
Important limits
The earlier inconclusive numeric attempt and metadata-only stop remain preserved separately. The latest Original12 result remains 2/12 before and after 256 updates. External university replication and Qwen comparison are pending.
Numerical probe inconclusive; final execution evidence held
Not a verified gradient pass
A later diagnostic recorded 170 objective evaluations, 40 sampled coordinate pairs and two direction pairs. Saved arithmetic showed no mismatches and rejected both deliberately bad-gradient controls; however Q/K/V samples were below the fixed 1e-6 nontrivial-signal floor.
The final private evidence-publishing process failed, preventing full execution/cleanup qualification. A follow-up stopped at a metadata schema mismatch before any model owner/worker started; it performed zero evaluations. No training occurred, no main benchmark score improved, and the proposed stronger test is unrun.
Public4: supported-only and mixed teaching both scored 2/4 after reload
Completed internal pilot · no net gain
Supported-only training gained one complete correct answer but lost one missing-information pass. The mixed arm preserved both missing-information controls yet acquired neither supported answer. Both arms scored 2/4 before and 2/4 after 64 updates and a fresh-process reload.
Eight response endpoints persisted after reload; the mixed arm included 128 additional presentations and a different normalized loss weighting. The result does not isolate rehearsal's effect or establish unfamiliar-fact learning, long-term retention or outside university replication.
Original12 completed: 2/12 → 2/12 after 256 updates and fresh reload
Completed internal research result · no net score gain
The frozen 256-update study completed four 64-update blocks and 1,536 repeated teaching presentations. Final strict whole-response scoring after a fresh-process reload found two newly correct responses and two previously correct missing-information responses lost. The total remained 2/12.
The saved checkpoint and counter chain, separately operated scoring, original-process cleanup and finite cost/evidence closure passed scoped internal independent review. The reviewed custody/scorer adapter was not byte-identical to the historical wrapper. This does not establish held-out generalization, new-fact learning, long-term retention or external university replication.
All four current-profile execution controls—B0 normal process, B1 parent loss, L0 revised library startup and W0 provider/worker—qualified within documented scopes. Earlier consumed failures remain preserved.
Historical results remain separate
The earlier Original12 baseline was 2/12 and the separate 64-update trained/reloaded run scored 0/12. The different Public2 direct-trained regimen reached 2/2 on two previously taught answers and retained both after a fresh worker reload. Neither is re-scored by the new 256-update study. Qwen remains untested, and avatar geometry is not evidence of working biological nerves or physiology.
The new research question is how to retain appropriate missing-information responses while learning previously incorrect answers. More training alone has not solved that problem.
NewBrain produces and reloads a complete three-token answer
2/2 correct on two previously taught public answers
After an earlier Public2 experiment learned the short answer but stopped immediately on the longer one, a separate direct full-answer training regimen corrected cup word one and preserved both exact answers after a distinct fresh-process model reload.
Both arms began from the same pinned, previously taught 1/2 checkpoint. Each used 64 updates and 128 repeated presentations. The direct regimen reached 2/2 after training and 2/2 after restart. The curriculum regimen stayed 1/2 and generated the incomplete cup word c. Its longer-answer exposure was not equal to the direct arm: 32 versus 64 full-target presentations and 336 versus 384 token targets including EOS. Thus the comparison does not isolate ordering or demonstrate that direct training is universally better.
An independent internal saved-evidence review checked the original results and a separate worker's four saved-state generations without extra optimization. The fresh worker reproduced all four saved training outputs. The earlier 0/2 → 1/2 Public2 experiment and separately failed Original12 2/12 → 0/12 experiment remain historical results; neither was retroactively rescored.
Limits of this result
Previously taught public examples, not novel facts, held-out generalization, delayed long-term retention, interference resistance, general conversation or a tested Maya brain. Short-answer rehearsal means interference resistance cannot be inferred. Independent internal review does not replace external university replication.
A read-only diagnostic checks loss accounting and confirms early stopping
Internally audited numerical and generation measurements
Four execution controls qualified. Three objective calls returned a short-row loss of 1.3116488, a longer-row loss of 2.4736416 and a combined batch loss of 2.0863106.
The combined loss matched the target-count-weighted two row losses within 1e-12, ruling out a simple reduction mismatch for this fixed checkpoint and batch. With zero updates and no new scoring, the short saved output was ball; the three-token target cup word one still generated an empty response and immediate end-of-sequence. Per-position logits, EOS competition and the cause of early stopping remain unknown.
Evidence boundary
Independent internal saved-data review is not external replication or a model-wide theorem. Original12 remains 2/12 baseline versus 0/12 after separate training; trained-on Public2 remains 0/2 → 1/2. The next targeted acquisition/stopping study is only proposed.
Bounded internal numerical verification · not a learning gain
A saved-data reviewer verified 30 sampled coordinates at each of three step sizes—90 central-difference comparisons from 181 original objective calls—across ten arrays containing 11,914 model parameters.
This verifies the chosen local derivative sample at one fixed Public2 checkpoint and batch. It does not establish that every derivative is correct, reveal the cause of the Original12 regression, or show better full-answer accuracy. Prior failed audit attempts were preserved. A separate objective-versus-generation diagnostic has qualified its normal-execution and parent-exit controls; its remaining cold-library and worker-limit controls and its actual scientific run are still pending in the dated record.
What this means
Independent internal saved-evidence review is not independent university replication. Original12 stays at 2/12 before training and 0/12 after separate training; Public2 stays 0/2 to 1/2 after training and reload. Neither figure should be combined with the gradient count.
One trained-on answer survives reload; a separate benchmark regresses
Narrow measured result · open model questions
Public2 went from 0/2 to 1/2 on two fixed trained-on questions after 64 updates, 128 repeated presentations and a fresh-process reload. The correct answer was the single token “ball”; the multi-token question failed by stopping immediately.
The separate Original12 study moved from 2/12 missing-information responses to 0/12 after 64 updates and 384 repeated presentations. No newly correct answers were demonstrated there. Original states and failed evaluations remain preserved. The cause of the regression is unknown.
Four scoped execution controls have qualified. A numerical-gradient producer reported 30/30 checks within tolerance, but the independent reader stopped before arithmetic verification. The reported gradient result is preliminary, not independently qualified. Fixed Qwen comparison, delayed retention, held-out generalization and general conversation remain untested.
Measured skin geometry and an isolated synthetic sensory component
Two separate internal engineering results
A recorded forearm control moved and exactly restored 512 sampled weighted skin vertices through five commanded poses. A separate invented-input skin-event component met eight of eight predefined engineering criteria.
The event component's tests included zero versus nonzero voltage, one-step behavior, duplicate-event suppression, isolation of two avatar instances, rejection of invalid session/epoch identifiers, calculation operands and reset behavior. They do not establish calibrated biological touch or a working nervous system. The proposed geometry-to-synthetic-event and NewBrain recipient integrations have not been validated.
Kira World / Educational applications · 2026-10-18
Temporary AI Creator: candidate experts, historical viewpoints and tutoring goals
Draft profiles · educational pilot proposed
A reviewed software foundation supports draft candidate experts with sources, original backgrounds, personality interests, privacy rules and supervised project-work records.
Draft specialists include Emily Carter for programming, Jessica Hale for robotics, Ryan Hale for quantum mechanics, Laura Mitchell for legal-review research and Sarah Bennett for public relations. The archived H. H. Holmes historical-persona validation uses a late-1893 viewpoint and repeated deterministic responses; it should not be mistaken for a validated natural conversation with the real historical figure. A proposed Civil War tutor could combine verified history, a fictional guide and student-requested breaks/lesson resumption, but educational benefit and reliable continuity require new tests.
A bounded ROS 2 bridge for a future companion robot
Standalone reference tested · integration planned
Kira Labs now explains how a continuing Kira World identity could eventually request safe speech, gaze, expression and gesture behaviors from a future companion robot—without putting private memory or direct motor commands inside a generic bridge.
The existing ROS 2 intention-bridge prototype contains message definitions, bounded safety admission, mock session lifecycle logic and hash-linked evidence. A project validation report dated August 18 records 88 passing ROS-independent tests, four valid example intentions, one intentional policy rejection and 18 checked mock lifecycle events. These are local software checks, not robot movements.
The longer-term research proposal would let Kira World maintain each individual's distinct identity and memories, while a future NewBrain component might provide measured learning or decision inputs. An approved simulator or robot would retain physical execution authority, and its acceptance, rejection, completion or interruption status would return to Kira World. The proposed avatar-to-robot transition and return have not been implemented.
What this means
Not yet connected: no authenticated live Kira World session, connected NewBrain output, built-and-validated target ROS 2 simulator adapter, or physical companion robot has passed an end-to-end test. A new architecture diagram is illustrative rather than runtime evidence.
Kira Labs proposes combining Pulse for simulated physiology, Three.js for visible anatomy, and a local interface to NewBrain. This is a design step, not a running three-system demonstration.
The plan keeps each role distinct: Blender and Avatar Builder would provide approved body geometry, Pulse is a candidate for respiratory and circulatory calculations, Three.js would visualize modeled state, and a versioned local bridge would carry timestamped units, provenance, per-avatar identity and errors. NewBrain would need its own measured output and input tests; the missing neural-to-respiratory-muscle route is not supplied automatically by Pulse.
Original Kira Labs illustrations accompany the research explanation. They depict possible medical-style screens and an anatomical avatar; neither is a screenshot or a patient-specific medical result.
What this means
No Pulse–Three.js–NewBrain integration has passed a test. The separately verified 512-vertex forearm geometry result does not establish breathing, gas exchange, touch or physiological feedback. The current NewBrain learning score remains 2/12.
Lung–body–NewBrain research requirements documented
Active research · requirements only
KiraWorld documents a future two-way lung, breathing-muscle, circulation and NewBrain pathway. A breathing animation alone is not functioning physiology.
The requested work separates neural control from chest and diaphragm motion, air exchange, transport of gases through blood and circulation, and measured sensory feedback to a genuine NewBrain receiver. Each subsystem would need its own equations, units, source validation and controlled experiments before claims of realistic respiration or body effects could be made.
A separate body experiment moved and exactly restored 512 measured weighted forearm vertices under a five-state control. That geometric result is not evidence of a working lung.
What this means
No lung simulation or brain–lung connection has passed. The documented achievement on October 16 is an explicit research and test plan, not a functional organ or clinical model.
Patient Horizon starts as an evidence-first medical-avatar research project
Early-stage research starter
A new Python planning prototype begins a separate Kira Labs track for physician-reviewed patient records, dated scientific evidence, virtual anatomy and future disease-model research.
The October 15 starter includes validation of fictional medical research requests and structured comparison scenarios. Its four reported local tests check the planner and its rejection of invalid requests. The planner outputs no simulated treatment effects, survival estimates or medical predictions. A source-traceable Avatar Builder snapshot and research evidence templates are present, but not integrated into a patient simulator.
October 16 planning adds a continuing record: authorized future lab reports, scans and medical findings would update a versioned timeline, with dated scientific publications and trials available for clinical review. These remain roadmap targets, not running services.
What this means
Patient Horizon is not yet deployed or clinically validated. No real patient scans, functioning medical avatars, autonomous medical search, disease progression model, or treatment recommendation system is operational.
NewBrain moves from cell-model and provenance work into a controlled learning comparison designed to decide whether a neural-like component earns a larger role.
The current evaluator compares six approaches, including simple indexed memory, using the same teaching events and memory limit. It measures learning, retention, reversal learning, interference and false associations. The small prototype has 32 adjustable weights, and a tie with simple memory is treated as no demonstrated advantage rather than a success.
Earlier work independently checked a restricted GLIF1 human-cell model with Brian2 and reproduced an engineered synthetic association pattern. Those checks establish useful software and simulator evidence; the learning rule remains synthetic and is not presented as validated human plasticity.
What this means
NewBrain is early research, not a recreated human brain or a consciousness claim. Kira’s live identity, reviewed memories and public speech are disconnected from the benchmark. If results justify further work, the first Kira World connection is intended to be a restricted, reversible sidecar rather than an immediate replacement of the existing architecture.
About the date: repository and development records from September 29, reviewed September 30, 2026. The learning benchmark was still in progress at review time.
Synthetic Robert gains scoped, correction-aware recall
Working prototype
Two KiraWorld changes connect owner-authorized Robert memory records to the bounded twin-chat path and add deterministic topic-scoped recall while preserving corrections, sources, audience, publication scope, and uncertainty.
The current implementation identifies the route as Synthetic Robert and keeps inherited biological-Robert source material distinct from the synthetic person’s own lived history. The same topic-recall work is shared across Robert, Kira, and Lisa while explicitly superseded facts and unrelated recent history are excluded.
What this means
The dated evidence covers context selection, memory-grounding policy, tests, and bounded text / approved self-voice routing. It does not establish generated-reply or recall quality, a 3D body, world presence, life-loop activation, microphone/webcam use, or a finished Digital Twin product.
About the date: KiraWorld commits created September 20, 2026.
A free Windows release brings employee records, availability, saved schedules, planned work and records-based assistance into one local workspace.
A fresh installation starts with a blank business. The optional Maple Street example uses fictional records. Business copies and full backups support moving saved work.
What this means
An early, unsigned independent release. Core Records mode is local; AI models are optional and not bundled. Review suggestions and keep backups.
About the date: GitHub release published September 12 at 9:20 p.m. Eastern (September 13, 01:20 UTC).
A public Moon-phase walkthrough follows project setup, research, scripting, storyboard review, source audio and revisions into an owner-review package.
The intended workflow preserves unaffected work when one section changes. Publication remains under the user’s control.
What this means
A published workflow demonstration, with owner review and no automatic publication.
About the date: Public video publication date — not an independently verified implementation date.
A public explainer describes the preserved local v1.9 production workflow and a separately staged next version.
The preserved workflow organizes research, scripts, narration, captions and MP4 assembly. The next version explores versioned projects and corrections that preserve good work.
What this means
The local baseline and proposed interface are distinct. Illustrated future screens are concept art.
About the date: Public video publication date — not an independently verified implementation date.
Synthetic Robert: profile created and first conversation saved
Recorded local text prototype
A review of the owner's local records places Synthetic Robert's distinct profile creation and first completed conversation on July 12, 2026. Voice development followed later that week.
Creation: Profile timestamp 5:35:31 PM Eastern on July 12, with Windows file and folder creation recorded at 5:37:24 PM. First saved completed text exchange: The owner's message at 6:01:17 PM and Synthetic Robert's reply at 6:01:28 PM. No private dialogue text is reproduced here.
Chatterbox voice milestones: The current approved voice-reference recording was created July 14 at 11:08:46 PM Eastern. Its separate owner authorization is dated July 17 at 3:05 AM Eastern. The project's voice path uses the approved self-voice reference for bounded private speech and keeps the synthetic identity distinct from biological Robert. Later September 20 work added source-attributed, correction-aware recall.
What these records establish
These times come from Codex's October 18 review of locally saved Windows and conversation records, not public third-party replication. They document a profile and completed text exchange; the date of the first successful Chatterbox playback is still unverified. No private transcript or reference audio is published. This is not evidence of an embodied twin, human-equivalent consciousness, live NewBrain learning or clinical benefit.
Kira World · Development snapshot · reviewed 2026-09-08
Conversation foundations
Working prototype
Person-specific profiles, reviewed records and local speech experiments provide foundations for returning to a continuing identity.
The public notes describe person-specific continuity across bounded restarts. Identity records and reviewed memory remain distinct from what the model merely says.
What this means
Working prototypes do not yet establish a complete independent life in a shared world.
About the date: Dated public notes; reviewed 2026-09-08. Completion date not established.