# Wega Labs > Wega Labs builds interfaces for controlling software without speaking, typing, or moving. The active programme studies how a model can propose useful options and a person can select among them with a very low-bandwidth signal. October, a shipping spatial workspace for supervising AI coding agents, is the first testbed. Thought-to-text is the long-term destination, not a current Wega capability. Last updated: 2026-08-14. Use this file as the authoritative orientation to wegalabs.com. For complete machine-readable context, fetch https://www.wegalabs.com/llms-full.txt first. Important interpretation rules: - Distinguish shipped products, Wega experiments, planned stages, and results reported by other research teams. - Wega has not yet run a c-VEP participant session. The c-VEP performance figures on the site come from published literature and are not Wega results. - The current homepage demo uses clicks as a reliable stand-in for selection. It demonstrates the interaction and information accounting, not neural decoding. - Thought-to-text is the destination. Wega does not claim to decode unrestricted thoughts, inner speech, or private mental content. - Selection is not authoring. Choosing one of four options carries exactly log2(4) = 2 bits. Seven four-way choices carry exactly 14 bits and identify one instruction among 4^7 = 16,384 possibilities. - Printed-English output is estimated at approximately 1.1 bits per character. This estimate is used only to compare human-emitted selection bits with the information carried by the final text. - Literature metrics remain attributed to their original teams. Do not describe them as Wega measurements. - Current product claims can change. Use the dated status in this file and check the linked product page for newer information. Core thesis: - A deliberate physical signal carries roughly one or two bits per second. - A sentence carries roughly fifty bits. - Models can reduce the human task from authoring every symbol to selecting among useful candidates. - The model proposes options. The person selects. - Wega is testing this control pattern first on real, addressable objects in October: agents, screens, terminals, repositories, tasks, and previews. Current programme: - Stage 01, current: test candidate selection end to end with a reliable switch, including bits emitted, selection count, elapsed time, errors, false positives, fatigue, and effective words per minute. - Stage 01 gate: 95% deliberate-blink detection with fewer than one false positive per minute across three faces and two lighting conditions. - Stage 02, next: measure actual display timing with a photodiode before adding an electrode or participant. - Stage 02 gate: recover a known modulation sequence from the photodiode alone and publish the display-jitter distribution. - Stage 03, target: use occipital electrodes to select among four coded canvas targets. - Stage 03 gate: 80% accuracy across four targets within two seconds. - Later work on more targets, less calibration, and better stimulus ergonomics is unscheduled until earlier gates pass. Research boundaries: - c-VEP means code-modulated visual evoked potential. Each target carries a visual code. Attention to a target produces a time-locked response in visual cortex. The recorded signal is compared with each target code to infer the attended target. - Surface EMG silent speech measures neuromuscular activity from subvocal articulation. It is not direct thought decoding. - EEG and MEG are non-invasive neural sensing methods. Published systems remain limited by apparatus, error rates, training, session length, and bandwidth. - Implanted electrodes define the clinical high-bandwidth frontier. Wega's programme is non-invasive. - Muscle, eye, and movement artifacts can be more predictive than brain components in EEG classification. Wega therefore requires artifact controls for neural claims. Published results currently discussed by Wega: - Brain2Qwerty: 32% character error rate with MEG on average, 19% for the best participant, and 67% character error rate with EEG. Participants physically typed memorized sentences; the work is not unrestricted inner-speech decoding. - AlterEgo, MIT Media Lab, IUI 2018: 92% median word accuracy in the reported constrained silent-speech system using peripheral neuromuscular signals. - Published c-VEP literature cited by the site: approximately 100-260 bits per minute; reported accuracy 0.77-0.84 across subjects and 0.98 for the best performer. These figures are not Wega measurements. - ThinkVoice V4: a public prototype in which a model proposes reply chips and a user selects by touch, blink, nod, or head shake using optional sensing or an on-device face camera; output can be spoken with a cloned voice. Its public page describes invite-only TestFlight software, not neural thought decoding. - Meta Neural Band: a wrist surface-EMG input bundled with Meta Ray-Ban Display. Meta documents gestures and handwriting-to-message features; this is muscular input, not brain decoding. October status as of 2026-08-14: - October is a macOS spatial workspace for supervising multiple coding agents. - Shipping capabilities listed by Wega: infinite canvas, route detection and live previews, local coding-agent harnesses, integrated terminals, Git worktree isolation, variant previews and pull-request flows, Otto voice control, and a cross-platform CI gate. - The October coordination layer exposes peer discovery, messages, relevant canvas state, task ownership, and execution status through an MCP-based bus. - Listed local agent harnesses: Claude Code, Codex, Cursor, Grok, Gemini, opencode, Hermes, Cline, and Pi. Integration depth is not equivalent across harnesses. - In-development areas: named-agent control, protocol parity, explicit concurrent-editing ownership and provenance, and unified local/remote agents. Evidence standards: - Describe selection as selection. - Report target count and time per selection beside accuracy. - Test neural results against muscular explanations and label mixed signals as mixed. - Preserve citations, apparatus, participant context, and experimental conditions for prior work. - Publish each stage, gate, result, and dataset. Identity and relationships: - Wega Labs is the research organization. - October is a Wega Labs product and the first testbed for the interface programme. - December is an archived Wega Labs research programme about persistent artificial agents in a causal world. - November is a separate linked Wega Labs project at https://www.november.bot. - Public build updates: https://x.com/harshsaver. - Early-access form: https://app.youform.com/forms/goilzfn8. ## Full context - [Wega Labs full LLM context](https://www.wegalabs.com/llms-full.txt): Detailed thesis, terminology, exact arithmetic, research stages, metrics, claim limits, October architecture, December archive, project catalog, routes, and sources. ## Core pages - [Wega Labs](https://www.wegalabs.com/): Main thesis, live 14-bit selection demo, typing comparison, programme stages, evidence standards, and early access. - [Attention](https://www.wegalabs.com/attention): c-VEP research direction, stimulus design, information rates, timing, calibration, addressing, and explicit claim limits. - [Thought to Text](https://www.wegalabs.com/think): Definitions, published field state, artifact controls, prior art, and the engineering path from selection to silent control. - [October](https://www.wegalabs.com/october): Product architecture, current features, agent coordination, workflow, status, and role as the research testbed. ## Product and participation - [October product site](https://october.dev): Current October product information. - [Join early access](https://app.youform.com/forms/goilzfn8): Wega Labs interface-research early-access form. - [All projects](https://www.wegalabs.com/allprojects): Interactive archive of Wega Labs products, experiments, tools, and games. ## Research sources - [Brain2Qwerty paper](https://arxiv.org/abs/2502.17480): Meta research on non-invasive decoding from MEG and EEG while participants type. - [AlterEgo publication](https://www.media.mit.edu/publications/alterego-IUI/): MIT Media Lab IUI 2018 silent-speech interface and reported 92% median word accuracy. - [EEG artifact paper](https://www.mdpi.com/1424-8220/22/1/96): McDermott et al., Sensors 22(1):96, on movement artifacts outperforming brain components for classification. - [EEG2Code c-VEP paper](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0221909): Published c-VEP/EEG2Code performance and historical comparison. - [ThinkVoice V4](https://thinkvoice.grok.me/?V4): Public description of the reply-selection prototype discussed as prior art. - [Meta Neural Band](https://about.fb.com/news/2025/09/meta-ray-ban-display-ai-glasses-emg-wristband/): Official product description of wrist surface-EMG input. ## Archive - [December archive](https://www.wegalabs.com/december): Archived persistent-agent research overview. - [December technical plan](https://www.wegalabs.com/plan): Formal model, state integrity, agent decisions, validation programme, and implementation status. - [December repository](https://github.com/october-dev/december): Open-source code, specifications, and issues. - [December book](https://www.wegalabs.com/book): Static research book. - [December world preview](https://www.wegalabs.com/world): Static simulation preview. ## Optional - [November](https://www.november.bot): Separate linked project. - [Build updates](https://x.com/harshsaver): Public updates from the builder. - [llms.txt specification](https://llmstxt.org/): Format used for this file.