Sources & context

Know what
you’re looking at.

Published information, working prototypes, and future directions keep their own labels.

Iris uses dated, reviewed public knowledge

Iris preserves the original website guide’s versioned public foundation: Robert’s biography and screen credits, book catalogue, dated development notes, and ShiftBrief guidance. Newer reviewed project summaries replace the older answers for their specific projects. The current website’s project selection is applied after the source files pass integrity checks. The original records and source revisions remain preserved in GitHub history.

Iris preserves the September 12, 2026 core source snapshot, version 2026-09-12.1, for the original biography, credits, books, and project-guide foundation. The website retains a reviewed public extension dated September 21, 2026 for Healthspan Lab and Synthetic Robert / Digital Twin status. A separate integrity-checked public project pack, version 2026-11-04.4, covers thirteen reviewed topics, with immutable older versions preserved for regression tests. Iris now includes the November 2 failed NewBrain learning result (0/6 in each scored category; Original12 remains 2/12), the narrowly tested Avatar Builder workspace launcher, and a bounded public review of Aster's October 25 synthetic retention study integrated November 1. The rejected human head and body are not approved; preserve the separately accepted round eyeballs. Replaying incorrect answers, passing software checks and small synthetic retention gains do not prove realistic anatomy or general conversation. The November 4 optional reader and regional scheduler remain limited source tests. Later NewBrain checkpoint-memory experiments also lack a full native RAM measurement or new learning success. Original head V5 was actually constructed and rendered in 13 views with the approved blue eyeballs preserved, but still failed anatomical review. Latest November 4 source reviews also describe NewBrain V8v2 as unexecuted preparation, and distinguish Avatar Builder's failed V6 attempt from an unexecuted V7 source candidate; separate video and body proposals remain unqualified. The newest November 4 review also documents an unscored NewBrain V9 learning study and a proposed first-answer credit-flow diagnostic. The original V8 head was rendered and rejected; its jaw/neck narrowing was traced to the early control mesh, while corrected V9 head and V6 body sources are not yet qualified. The real full-body V5 build failed its allocation limit. Iris's project replies use dated public summaries, not private runtime access. Iris is the website guide only; Sarah in ShiftBrief and Sarah Travel has not been renamed.

Iris can retain a project topic, recognize common project-name typos and question paraphrases, ask which project you mean, distinguish movie and television questions, compose source-based mini biographies, list books, and add reviewed detail in follow-up answers. Answers run inside the browser with no AI model, backend service or model subscription. These are dated notes, not live research or general-purpose AI. She cannot verify unpublished work, invent credits, supply full manuscripts, or access private records.

Robert’s published screen credits · Book listings · About Robert

Status labels matter

Actual screenshot or output: published development material, not automatically a fresh test. Working prototype: a demonstrated part of the work, not a completed product. Active development: engineering in progress. Concept visualization or future goal: direction, not a current implementation.

The lab timeline follows Kira Labs and the projects currently featured on the site. A video’s publication date is not an independently verified date of implementation. Future goals do not imply release dates.

Explore the Kira Labs timeline · Current projects

Unknown information and direct contact

When the published notes do not answer a question, Iris should say so. She does not make commitments or send messages on Robert’s behalf. Use the contact form to send Robert your own message without an email app.

Ask Iris or contact Robert · Privacy

Iris knowledge / October 24, 2026 / Verified name decision

Person Creator replaces Temporary Creator

Kira Labs decided on October 24, 2026 to call its synthetic-person creation research Person Creator, rather than Temporary Creator or Temporary AI Creator. The new name acknowledges that intended experts, companions and characters may be either temporary or continuing individuals; permanence is a property of the person, not of the tool.

Iris's public project knowledge includes the updated name and recognizes the previous names as aliases. Her information remains sourced to public Kira Labs research—not to private model files or visitor conversations. The rename does not establish automatic individual creation, working identity persistence, complete mind/body/voice integration or a released installer. Read the Person Creator announcement and research limits → · Ask Iris about the rename →

Current reviewed public project knowledge · October 21, 2026

What Iris knows now

Iris now has independently versioned, integrity-checked public summaries of NewBrain, Kira World, Evelyn Hart, Patient Horizon, Avatar Builder, Aster, IdeaForge, Humanoid Researcher and other research projects. These are dated source reviews rather than live measurements. Iris remains a browser-only website guide: she does not operate the research programs, connect to personal accounts, expose private files, or send visitor questions into NewBrain.

Latest NewBrain scientific audit: the independently reviewed saved Original12 training candidate (experiment 1078; audit 1109) declined 2/12 → 0/12 after 256 updates while the no-training control remained 2/12. Teacher-objective loss decreased sharply, but no assessment answers improved and two previously correct answers were lost. The candidate was not promoted, and the accepted reference benchmark remains 2/12. The separate 3/3 after-restart result used explicitly rehearsed questions; it does not establish transfer to different assessment prompts. Read the audited negative transfer result →

Separate NewBrain engineering update: a corrected Windows process-owner launch passed 26/26 predefined synthetic metadata compatibility checks, including clean narrow termination and review. The earlier 819-byte memory-limit mismatch remains preserved as a historical failed attempt. These checks are not learning questions. Read the passed Windows result →

Earlier separate NewBrain learning evidence: a separate, balanced 256-update synthetic three-question study improved 2/3 → 3/3 → 3/3 after a distinct fresh restart, while the untreated control stayed 2/3. All three examples were explicitly taught. Another curriculum experiment fell 2/3 → 0/3, and a no-update diagnostic identified two weak next-token decisions. The main Original12 benchmark remains 2/12, and Qwen has not been tested in the intended comparison. These small exercises do not establish general language, new real-world facts or independent university replication.

Other projects: Evelyn's controlled eyelid closure/reopening reached 69.9% left and 72.5% right against a 65% threshold. Her planned complete-body opening and life-loop zoom to her face are not yet verified in a running application; voice, visitor-memory restart and an installer are also pending. Existing body-rig and 512 sampled weighted-skin-vertex engineering results may be reused for Evelyn and Patient Horizon, but do not establish a complete animated body or functioning organs. Aster's synthetic weight-protection experiment did not show benefit beyond replay; IdeaForge passed 14 scoped injected project-identity checks.

Read the newest Original12 audit → See all new test records → Ask Iris a project question →

Historical project review · October 6, 2026 · newer October 21 status above

What the current project notes establish.

Approved high-level summaries of current source and documented evidence. This review is not a new end-to-end run, safety certification, release promise or live connection to any project.

These visitor-facing summaries publish only approved project capabilities and boundaries. Private repositories, research archives, personal records and local runtime data are not available through Iris. Measured results below belong to their stated tasks; proposed code and unrun fixtures are labeled separately.

NewBrain

NewBrain is a separate local-first research project toward a learned cognitive backend for Kira World. It explores learning and retention while keeping each synthetic person’s identity and stored history separate. A small controlled command learner has measured results; general conversation and connected brain/body abilities remain experimental.

Current status. Reviewed October 3, 2026: the small command learner has measured learning and retention results. The newest source adds a small token-generating dialogue decoder, but all 16 engineering fixtures, training and actual generated replies remain unrun. The separate plasticity component has now passed all 20 local pure-Python engineering checks, with independently reviewed saved evidence. Its delayed-cue, context/rule-change and lesion/sham scientific campaign remains unrun, so improved learning or damage recovery is not yet demonstrated. NewBrain is not the active Kira World backend, and no public installable release or trained-weight download is verified.

Evidence and validation. Five repeat trials on the same exposed, tightly limited command task retained every old and new answer when earlier lessons were rehearsed. Without rehearsal, old-task retention fell to about 42%. A capable hand-written grammar also answered perfectly, so neural superiority is unproved. An earlier association benchmark tied simple indexed memory. The visual learner memorized training images but failed unseen color/shape combinations. The newest dialogue component’s 16 fixtures remain unrun. The separate plasticity component has now passed all 20 local pure-Python engineering checks, with independently reviewed saved evidence. Its delayed-cue, context/rule-change and lesion/sham scientific campaign remains unrun, so improved learning or damage recovery is not yet demonstrated. Passing legacy core CI does not execute the separate research fixtures.

Approach, requirements, privacy and limits

How it works

The research tests whether learned systems can acquire new skills, preserve earlier lessons and handle corrections. The next dialogue checks cover short learned conversations, source-grounded corrections, retrieval, old-skill retention and a genuine restart. Strong simple controls are required before integration. Rehearsing earlier lessons helped the tiny command learner retain them; that is a bounded task result, not general conversational ability.

Technical approach and requirements

The project contains separate research tracks: recorded neural-data comparisons and GLIF neuron simulation, small engineered association learners, a from-scratch command learner, visual-learning trials, and a proposed token-generating dialogue decoder. Short GLIF circuit scaling tests reached 5,120 units; 1,280 units met both provisional timing targets. That is neither a whole-application capacity limit nor sustained no-lag performance. The dialogue decoder has not yet been trained or run.

Privacy and external services

NewBrain is private, local/offline research rather than a visitor service. Iris has only this approved high-level summary and cannot open private research data, media, identities, histories, model weights or runtime state. The website does not send your Iris questions to NewBrain.

Limits

NewBrain is not a Qwen replacement, complete human brain, ordinary conversational model, robust vision system or evidence of consciousness. The tiny 16-command comparison with Qwen establishes no general intelligence or efficiency advantage; its latency and RAM methods were unequal. More RAM alone does not establish learning or generalization. No connected brain/body ability or finished cognitive architecture is claimed.

Availability and cost

No public NewBrain product price, installable release, trained-weight download or reuse license is verified in these notes. Iris itself needs no AI subscription or paid model service; this does not imply that future NewBrain compute or products will be free.

NewBrain’s relationship to Kira World

IdeaForge

IdeaForge is a local-first invention research and project-planning workspace. It turns an idea into goals, subsystems, research questions, reference choices and staged prototype plans while distinguishing available technology, experimental work and fictional or undemonstrated goals.

Current status. Implementation reviewed October 3, 2026: the source supports multiple persistent projects, typed/local-model chat, optional Windows voice, project variants and selected reference images, online research and a watch queue, equipment records, project memory, and text/image-assisted troubleshooting. The latest storage update gives new projects separate numeric identities so duplicate display names can remain separate. This is a source review, not a fresh end-to-end run or public packaged-release claim.

Evidence and validation. Reviewed October 3, 2026 against current implementation and documented boundaries. Focused storage checks cover separate project identities and duplicate display names. They do not validate crash recovery, the native GUI, real model-answer quality or speech-recognition accuracy. Limited mesh/build-volume checks and a simulation plan do not prove an invention works or is safe. No new end-to-end run or engineering certification is claimed by this summary.

Approach, requirements, privacy and limits

How it works

Start with an idea and define its goals, known engineering facts and unknowns. Organize subsystems, compare research and reference images, and build a staged prototype plan. The research queue and recurring watch operate while the app is open. The automatic geometry path currently makes parametric enclosures only when required dimensions are supplied; other shapes need custom CAD. An optional PyBullet runner takes an existing URDF, rather than automatically validating an entire invention.

Technical approach and requirements

The documented setup is Windows with Python 3.11 or newer, a prepared Python environment and local Ollama with qwen3.5:9b for chat. qwen3-vl:8b is optional for image analysis. OpenSCAD can export enclosure STL when separately installed; PyBullet is a separate optional tool. Voice needs an explicitly built .NET/System.Speech helper, an installed English recognizer, a suitable Windows voice and a working microphone.

Privacy and external services

Project files, equipment records and history are stored locally, and the default model endpoint is local. Research is online: outgoing searches can include the project name, subsystem questions and saved engineering facts. IdeaForge should not be described as fully offline or as guaranteeing that invention details never leave the computer. Voice records no raw microphone audio, but recognized questions enter ordinary chat/project records. Iris cannot access those records.

Limits

Automatic CAD is limited to the documented parametric enclosure path with known dimensions. Other geometry remains custom CAD or TBD. OpenSCAD output, a mesh check or an optional existing-URDF simulation does not certify an invention’s safety, manufacturability or real-world function. Speculative goals and unsupported technology must stay labeled.

Availability and cost

These notes establish no public IdeaForge price, packaged installer or reuse-license grant. Its documented local models and optional CAD/simulation tools require their own setup and computing resources. Iris itself has no AI-service subscription or per-question model charge.

Humanoid Researcher

Humanoid Researcher is a local-first research library and candidate-design workspace for humanoid robotics and body subsystems. It organizes source material and proposed designs while keeping verified facts, candidates and unresolved details separate.

Current status. Implementation reviewed October 3, 2026: the source collects papers, robotics news and public robot repositories with attribution, saves research locally, supports local research/design chat and optional screenshot interpretation, and produces candidate body revisions, conceptual parts lists and assembly/procurement packets. Candidate proposals do not replace the approved body. NewBrain integration remains disabled and unqualified; no fabrication-ready release is established.

Evidence and validation. Implementation and stated boundaries were reviewed October 3, 2026; this is not a fresh end-to-end run or robotics certification. Compatibility scoring is a heuristic over recorded interfaces and ratings. It does not establish mechanical, electrical or thermal safety. Blueprint views remain diagrammatic placeholders unless verified geometry exists. CAD-derived views, physics/electrical/thermal validation and a fabrication-ready release gate remain roadmap work.

Approach, requirements, privacy and limits

How it works

Collect and categorize research, retain source metadata, and compare possible body subsystems against recorded requirements and equipment. Candidate tools can propose revisions, BOM and fastener schedules, assembly documentation, supplier searches and diagrammatic blueprint sheets. Keep VERIFIED, CANDIDATE and TBD distinct. Supplier links and lower prices do not prove compatibility or authorize substitution.

Technical approach and requirements

The app is Windows-oriented Python 3.11 or newer. Chat defaults to local Ollama qwen3.5:9b; screenshot interpretation needs optional qwen3-vl:8b. Collection uses online sources. A daily Windows scan is separately opt-in. Optional Windows voice starts off and requires the speech helper, installed recognizer and voice, plus a microphone. Embodiment modes are planning architectures, not deployed hardware.

Privacy and external services

Research, media, library and equipment records remain local in the documented setup and are not website content. Source collection and supplier searches use online services. Paper and image reuse still depends on original rights. The daily scan is opt-in and voice is initially off. Iris sees only this public summary and cannot retrieve local equipment records or private project memories.

Limits

This is a research and candidate-design tool, not an autonomous safe robot builder. Heuristic compatibility scores, conceptual BOMs and diagrammatic blueprints are not engineering certification or fabrication-ready geometry. No demonstrated mind uploading, consciousness transfer, deployed robot or active qualified NewBrain integration is claimed.

Availability and cost

No public Humanoid Researcher price, packaged consumer release or reuse-license grant is established by these notes. Local models, speech components and engineering tools need separate setup. Iris’s free browser-only answers do not certify or fund hardware construction.

Kira World

Kira World grew from the Kira Project, begun in summer 2025 with an attempt to create an artificial companion who could do more than answer questions: someone who could provide companionship to people who live alone and to people dealing with challenges such as anxiety, PTSD, autism or social isolation. The project evolved from a humanoid-robot concept into local-first research on continuing synthetic identities, reviewed memory, conversation, voice, virtual places and interchangeable embodiment. Working local conversation and continuity prototypes plus a basic virtual-world shell exist today; more complete bodies, shared worlds and NewBrain integration are still being developed and tested.

Early Kira + Friend prototype. The earliest recovered record is from August 2025. A desktop dual-chat app placed Kira beside an unnamed Friend and let the two exchange mostly scripted or rule-driven responses, including automatic back-and-forth. This was a working chatbot prototype, not two independent persistent AIs. It introduced the social design idea that Kira should have another synthetic companion; the unnamed Friend later became the conceptual predecessor of Lisa. Modern Lisa is a later, separately modeled identity with her own memories and continuity.

Current status. Reviewed October 5, 2026: Kira World still provides the existing one-person-at-a-time text/voice workflow, selectable profiles and 3D World Shell, while Avatar Builder work is being advanced separately. Recent source corrections now pass explicit project/dependency paths instead of guessing folders and consume the Avatar load witness before later body/pose checks. These changes improve the test route, but a native Blender body run with accepted movement, save/reopen, deformation, skin, breathing and physiology feedback is still unfinished. NewBrain remains a separate research project rather than the active Kira World backend.

Evidence and validation. The documented portable offline lane and bridge tests remain bounded interface evidence rather than proof of natural conversation or physical embodiment. Recent Avatar Builder work corrected source/path handling and the load-witness chain so future native runs can be attributed more reliably. That is engineering progress, not a passed moving-body result. No accepted native Blender movement/save-reopen/body-physiology result has been published in the current reviewed notes.

Approach, requirements, privacy and limits

How it works

The current design separates the continuing person from replaceable model and body infrastructure. Person-specific identity, profiles, reviewed memories, relationships, conversation and world state remain distinct; changing the underlying language model is not meant to create a different person. The main interface selects one person at a time. The basic World Shell can show named orbs for profiles without accepted bodies. A draft World Builder records requests; it is not a finished autonomous builder. The current conversation baseline remains a local language model. NewBrain is a separate possible future backend and does not control the live system in this work.

Technical approach and requirements

The local conversational baseline is Ollama Qwen3.5:9b. Voice is a separately installed and authorized dependency. Clean source setup needs Git LFS and external dependencies; it is not a ready-to-install consumer package. Avatar Builder belongs to Kira World: its newest source trial would build and save an inactive generic carrier, reopen it separately, then check actual mesh movement and geometric feedback. Those trial steps are not yet executed.

Privacy and external services

The source snapshot excludes private runtime state, raw conversations, model weights and unreviewed media. Individual identities and histories must stay isolated. Local-first describes the architecture, not a blanket guarantee that every optional dependency or authorized action is offline. Iris has only approved public notes and cannot access Kira World’s live memories, voice input or private state.

Limits

No finished anatomy/physiology package, accepted native moving body, completed shared life, working Video Studio production bridge or physical Hanson integration is established. NewBrain is not the active replacement for Qwen in Kira World. Source-reviewed body routes, compiler checks and preserved load witnesses do not by themselves establish an embodied synthetic person.

Availability and cost

The Kira World source/review repository is public, but no packaged GitHub release or verified consumer price is documented here. Running it needs suitable hardware, Git LFS, model/runtime dependencies and authorized voice setup. Public source does not itself grant blanket reuse rights. Iris’s browser-only guide needs no paid AI service.

Kira World project page · Public source reviewed October 5

BlueBook

BlueBook is a local-first workspace for investigating UAP and anomalous-phenomena records while preserving attribution and uncertainty. It organizes cases, sources and reported descriptions without treating an unresolved report as proof of an extraordinary explanation.

Current status. Implementation reviewed October 3, 2026: the source includes imported case/source records, a 2D map and 3D globe, reported-entity cards, related cases, program/document records, chat, and a reconstruction gallery with provenance. Actual tabs are Map, Globe, Details, Library and Gallery. Discovery adds provisional records for review while running. The official-record library is a seeded, in-progress collection, not an exhaustive verified database or packaged-release claim.

Evidence and validation. Implementation and boundaries were reviewed October 3, 2026, not freshly validated end to end. A government document establishes that a record exists, not that every interpretation is true. Witness and news claims remain attributed; unresolved does not mean extraterrestrial, and descriptor similarity does not prove identity. AI reconstructions are visualizations based on reported descriptions, not photographs or evidence. Hypotheses have no assigned probability or winner.

Approach, requirements, privacy and limits

How it works

Read attributed case records, inspect source material and reported descriptions, and compare related cases. Competing-hypothesis templates list assumptions, missing evidence and falsifiable checks. They assign no probability or winner and cannot promote a hypothesis into a finding. Automatic discovery periodically checks configured feeds and archive pages while the app runs; discovered records remain provisional until reviewed.

Technical approach and requirements

BlueBook uses Python 3.11 or newer and a local server. Local Ollama qwen3.5:9b is optional because deterministic routing remains available. Reconstruction prompts work without an image backend; rendering requires a separately configured local AUTOMATIC1111-compatible API. Hands-free uses the server computer’s microphone and installed Windows speech components, starts off, stops when the page is hidden/closed, and allows only one app or tab to use the shared speech engine.

Privacy and external services

User media and generated images stay in local folders in the documented setup. Online archive/news requests, external map tiles and CDN scripts mean the application is not fully offline. Optional hands-free requires same-computer local access and permission to use that computer’s microphone. Iris cannot read BlueBook’s local case notes, media or chat history.

Limits

BlueBook does not prove extraterrestrial life or turn a reported entity into a verified identity. AI reconstructions are not evidence. The record library is incomplete, and full timeline or relationship-graph UI remains architectural planning rather than shipped tabs. Missing evidence, competing explanations and source reliability must remain visible.

Availability and cost

No public BlueBook price, packaged release or reuse-license grant is verified in these notes. Optional local models and image generation need separately configured tools and computing resources. Iris itself uses no paid AI service or model subscription.

Equipment information

A dated picture of the workstation.

The equipment and upgrade priorities on the Support page include published September 8, 2026 notes; later October 18, 2026 research-partnership and equipment estimates are planning scenarios, not live shopping prices or proof of purchases. They are not live readings from Robert’s computer or a promise that a purchase has been made.

More memory, graphics capacity, and testing equipment are development goals. Read the current description and contact Robert before offering hardware.

View the equipment and priorities →