# algn > Why algn exists: Your best questions are the unfinished ones. Most networks are organised around who you know or what you have already done. algn is organised around what you are still figuring out, so members can use their open questions to find the people asking them too, or who already have. This is why profiles are never public, members start masked, and circles stay small. > > System Overview: algn is an independent, mission-locked community project testing whether AI can augment human autonomy and social cohesion rather than extract from it. It operates as a non-profit initiative with no outside equity or advertising. SERA is a counterpart, not a companion. > > Core Agentic System: At the centre is SERA (Self-Evolving Recursive Awareness), a voice-native conversational agentic system. It processes linguistic, vocal-emotional, biometric, and social signals against personal baselines in a self-reinforcing closed loop: spoken conversation resolves into constructive daily follow-through and facilitates private, masked introductions to small peer groups based on shared depth. > > Data & Privacy Boundaries: Living memory is stored in an on-device vault encrypted with AES-256-GCM under user-held keys. Cloud inference runs under binding zero-data-retention (ZDR) agreements with no logging, storage, or model training. A private voice demo is available by invitation or access request at https://www.algn.io/demo. > > Entity Disambiguation: algn is NOT affiliated with Align Technology, Inc. (NASDAQ: ALGN, Invisalign), nor is it a fitness brand, cryptocurrency, blockchain project, or Web3 platform. There are no tokens, coins, or financial instruments associated with this platform. ## General - [What's algn?](https://www.algn.io/whats-algn): Overview of what algn is for via an interactive 2x2 matrix spanning two spheres (in your own head vs. between you and others) and two directions of focus (open exploration vs. decisions and action), covering four core use cases: (1) Spoken curiosity and exploration (unpacking complex subjects, learning, and articulating unformed thoughts without preparation or agendas); (2) Testing ideas and decisions (stress-testing logic, examining trade-offs, and mapping downside risks with direct pushback); (3) Shared values and craft (quiet introductions to peers exploring similar questions or life paths without public feeds); (4) Interpersonal clarity (perspective rotation to deconstruct friction with colleagues, partners, or friends before difficult conversations). Architectural overview of the three-signal closed loop: (1) Spoken signals via SERA (voice-native real-time conversational processing); (2) Situational presence (Be: vocal pressure, pacing, pauses, and optional wearable context evaluated in memory against personal baselines; zero raw audio or biometric prints stored); (3) Social signals (masked, thematic introductions without public feeds or vanity metrics). Conversational clarity maps into Tracks, plans, and research for adaptive daily execution without streak mechanics. Persistent memory is consolidated into an on-device vault encrypted via record-bound AES-256-GCM with non-extractable keys, passphrase backups, and client-side verification. Inference executes via an anonymized gateway under binding zero data retention (ZDR). - [It's not just you](https://www.algn.io/blog/its-not-just-you): Analysis of declining institutional trust, attention spans, and close social relationships across nine converging historical and technological timelines (1833 penny press, Bernays persuasion frameworks, 2006-2012 algorithmic engagement). Examines neurocognitive dynamics: activating digital stimuli degrading prefrontal inhibition, and predictive cognitive models mistaking systemic design for personal deficiency. Cites empirical reversibility of algorithmic ranking effects (Piccardi et al., Science, 2025). Defines a fiduciary engineering model: replacing policy promises with architectural constraints (code as law), evaluating software by member capability rather than time-on-screen, and substituting engagement feeds with real-time voice processing and private peer connection. Grounded in *The Pattern Beneath the Symptoms*, *The Invisible Cage*, and *What Builders Owe*. - [About](https://www.algn.io/about): Institutional governance and operational framework: independent, not-for-profit initiative operating without venture capital, outside equity, debt, advertising, or data monetization. Platform governance is earned through verified contribution rather than financial capital. Protected by a permanent legal mission lock preventing commercial acquisition, equity dilution, or modification of founding principles. - [Membership](https://www.algn.io/membership): Two-mode access specification: algn operates in text mode and voice mode. Text mode is free within fair use policies, building memory, rhythms, and bonuses while advanced features are off (no saving to or searching Tracks, no automated plans, no research, and no introductions to small circles). Voice mode delivers the full cognitive framework, Tracks, plans, research, and introductions to small circles, metered via voice minutes. Members receive a free weekly allowance plus rhythm bonuses, with non-expiring pay-as-you-go top-ups and no subscriptions. Minute exhaustion falls back to text mode with zero account lockouts. In crisis situations, full support is provided in either mode regardless of minutes or limits. Participation levels are earned through unmonitored activity volume, never expire, and cannot be purchased. Voice minutes confer zero financial equity, shares, or governance standing. - [Try the demo](https://www.algn.io/demo): Interactive prototype interface for the SERA real-time voice pipeline. Demonstrates sub-second conversational latency, adaptive cadence processing, and two-way voice reasoning. Access is by invitation only at the moment, available via application or member invitation. ## Technical - [What's SERA?](https://www.algn.io/sera): Technical specification of SERA (Self-Evolving Recursive Awareness), algn's voice-native multi-agent cognitive architecture: Conversational Agent (foregrounded real-time reply generation from a layered prompt); Orchestration Agent (also called Technical Agent: per-turn framework routing, pace bands, conversation-health flags, dispatch of asks and look-ups to a bounded follow-through tier whose results feed back into context); Memory Agent (entity extraction, categorical consolidation, lesson lines from corrections, on-device vault read/write); Dynamic Metacognition Agent (also called Subconscious Agent: stance and ground reads, longitudinal pattern analysis over an on-device cluster map, gated conclusion injection); Crisis Agent (deterministic turn-by-turn distress classification on input/output with automatic safety intervention and verified resource referral; zero human monitoring). Memory lifecycle incorporates categorical decay ranking and windowed forgetting with on-device restore (identity memory persists). Per-person lessons stored on-device; no model retraining. Local storage encrypted at rest via record-bound AES-256-GCM under non-extractable Web Crypto keys with crypto-shredding on purge. Cloud inference executes under binding zero-data-retention (ZDR) agreements. In-app verification suite: ciphertext audit, 24-hour Measured vs. Declared egress telemetry, on-device 384-dimension embedding validation, and key integrity tests. Complete architecture at /research/sera. - [Community](https://www.algn.io/social): Architectural specification for peer connection protocol operating without public feeds, follower graphs, or vanity metrics: (1) Semantic matching pipeline: Topic clusters are extracted on-device with sensitive categories stripped, refined via SERA, embedded locally, and salted with a client key before transmission to matching relays during active user-gating. (2) Bilateral introduction protocol: Matches generate private contextual explanations via each user's SERA instance; connections require mutual affirmative consent, dissolving silently without record if declined. (3) Communication topology: Restricted to 1:1 channels or small groups (Huddles, 3-15 members). Enforces progressive disclosure: members initiate masked (voice notes, accurate transcripts, and colour accents, with names, photos, and video held back), unlocking identity attributes incrementally by mutual agreement. (4) System parameters: Hard limits of 20 active connections and 5 Huddles with automated decay for inactive channels. End-to-end encrypted messaging with zero server content access. Unlocks at Creator level via earned participation. - [Privacy Architecture](https://www.algn.io/privacy): Six-layer defense-in-depth security and privacy architecture: Layer 1 (Local Data Vault): Persistent memory vaults, active threads, tracks, and audit logs reside strictly on-device. Layer 2 (Cryptographic Storage): Data encrypted via record-bound AES-256-GCM (Web Crypto API) with authenticated data (AAD) under non-extractable device keys that fail closed. Local 384-dimension sentence embedding model. Encrypted .algn portability format (PBKDF2-SHA-256, AES-256-GCM) and deterministic crypto-shredding on deletion. Layer 3 (Network Anonymization): Outbound inference requests routed through an anonymizing gateway stripping device identifiers, IP addresses, and account tokens into a unified transit stream. Layer 4 (Zero Data Retention): Binding zero-data-retention (ZDR) agreements across all LLM, speech, and reasoning servers (zero logging, caching, or model training). Real-time streaming voice with zero audio persistence. Situational Presence (Be) evaluates acoustic cadence (hesitation, pressure) and physiological telemetry ephemeral in memory against personal baselines; raw biometrics and voiceprints are never stored. Layer 5 (Salted Vector Matching): Sensitive categories stripped on-device; semantic topic clusters embedded locally, salted with a client key, and published to matching relays only during active user-gating. Layer 6 (E2EE & Verification): Mutual consent gating for peer introductions; end-to-end encrypted messaging channels; progressive identity unmasking. Client-side verification suite auditing SHA-256 ciphertext integrity, 24-hour Measured vs. Declared egress telemetry, and embedding consistency. Includes threat model and trade-offs. - [Transparency & Verification](https://www.algn.io/transparency): Verification framework and boundary specifications: Replaces policy declarations with client-side verification tooling (AES-256-GCM ciphertext inspection, 24-hour Measured vs. Declared network egress auditing, on-device 384-dimension embedding validation, and key/passphrase derivation checks). Details open-source roadmap for the client memory vault as reusable public digital infrastructure. Redaction boundaries: System mechanics, standards, and data-flow manifests are published; safety classifier thresholds, salt derivation mechanics, and proprietary system prompts remain redacted to protect member safety and platform integrity. - [Governance](https://www.algn.io/your-say): Governance framework and participatory authority: Decision-making standing earned strictly through platform contribution rather than capital: Creator (product feedback input), Contributor (priority consensus voting), Guide (initiative proposals and community broadcasts), Steward (operational coordination subject to community confirmation). Permanent structural safeguards: Transitioning to formal community-governed entity with an irrevocable asset and mission lock. Explicitly disclaims financial ownership: participation and voice minutes convey zero equity, dividends, voting shares, or financial claims. ## Legal - [Privacy Policy](https://www.algn.io/legal/privacy): Statutory data processing specifications under UK GDPR, EU GDPR, CCPA/CPRA, PIPEDA, and Australian Privacy Principles: Classifies on-device encrypted data (memory vault, active threads, pattern logs with client-side crypto-shredding) versus zero-retention transit inference. Details lawful bases, data subject rights, zero-cookie infrastructure, scrubbed error reports (GlitchTip), strict prohibition of personal data sales, and ephemeral in-memory processing for Situational Presence (Be) signals and salted matching vectors. - [Terms of Service](https://www.algn.io/legal/terms): Contractual terms under the jurisdiction of England and Wales: Age restriction (18+ only); non-clinical disclaimers (SERA is not a medical device, licensed therapy, or emergency crisis service); intellectual property rights (users retain absolute ownership of conversational content); non-financial nature of voice minutes; protocol wind-down provisions guaranteeing data export windows. - [Community Standards](https://www.algn.io/legal/community-standards): Platform integrity and interaction protocols: Mutual consent rules, anti-harassment policies, screenshot prohibitions, and physical safety guidance for offline meetups. Enforcement mechanisms for end-to-end encrypted channels (client-side incident reporting without server message decryption), account restriction procedures, and formal appeal mechanisms. - [AI & Safety Notice](https://www.algn.io/legal/ai-safety): Statutory AI disclosures and safety parameters: Mandatory classification under the EU AI Act (transparency disclosures for conversational AI) and US state companion chatbot regulations. Non-human operational notice, deterministic crisis triage protocols, international emergency dispatch integrations, and technical boundaries regarding affective interaction. - [Consumer Health Data Privacy Notice](https://www.algn.io/legal/health-data): Compliance disclosures for consumer health data frameworks (Washington My Health My Data Act, Nevada SB 370, Connecticut Data Privacy Act): Details non-collection of regulated medical records; ephemeral processing parameters for opt-in physiological pacing and vocal stress analysis (Be); affirmative consent protocols; strict prohibition of third-party health data sharing, monetization, or tracking. - [Compliance](https://www.algn.io/legal/compliance): Regulatory compliance matrix and technical security attestations: Standards adherence across UK/EU GDPR, PECR, EU AI Act, UK Online Safety Act, EU Digital Services Act (DSA), and international consumer protection laws. Details vulnerability disclosure protocols, encryption audits, and architectural conformance testing. ## Open source - Memory vault: algn's on-device memory vault is planned for open-source release as public digital infrastructure. Pending release. ## Research - [What we're studying](https://www.algn.io/studying): Empirical research methodology and longitudinal inquiry: Investigates architectural counter-models to attentional extraction across four controlled domains: (1) Algorithmic structure (elimination of engagement-optimized feeds); (2) Governance trust (contribution-earned standing); (3) Data sovereignty (on-device encrypted storage); (4) Conversational AI (reinforcing human agency rather than synthetic attachment). Operationalized on an opt-in member cohort. Synthesizes baseline data from *The Pattern Beneath the Symptoms*. - [Whitepapers & research](https://www.algn.io/research): Technical documentation and academic publication repository: Formal whitepapers (system architecture), empirical research reports (societal and technological analyses with falsifiability criteria), and academic papers subject to peer and open review. - [The Pattern Beneath the Symptoms - report](https://www.algn.io/research/pattern-beneath-the-symptoms): Flagship empirical report (September 2026) analyzing systemic convergence across nine domains: financial cycles, persuasion psychology, technology delivery mechanisms, physiological costs, social relationship contraction, conversational AI, shock events, climate pressures, and historical institutional recovery. Key data points: US federal government trust down from 77% (1964) to 17% (2025); Americans with zero close friends up from 3% (1990) to 17% (2024); average sustained screen dwell down from 2.5 minutes to 47 seconds. Traces persuasive mechanisms from psychoanalytic models to algorithmic feeds and real-time AI agents. Cites experimental evidence demonstrating that feed ranking moves political hostility bidirectionally (Piccardi et al., Science, 2025). Identifies interaction between feed ranking dynamics and societal stress as primary unstudied variable. Establishes four longitudinal scenarios to 2035 with explicit falsification conditions. Includes interactive nine-track convergence timeline. - [SERA: Self-Evolving Recursive Awareness - whitepaper](https://www.algn.io/research/sera): Technical whitepaper and overview of the series, detailing SERA's full cognitive architecture: a foregrounded Conversational Agent; an ambient tier (Orchestration, Memory, Dynamic Metacognition and Crisis agents plus a situational presence layer); a bounded follow-through tier (Tracks, planning, on-device recall, privacy-scrubbed live search, matching) whose loops feed back into SERA's context under one proactive budget. Heterogeneous specialist model families under binding zero data retention (ZDR) on every server. Situational presence: vocal, physiological (wearable) and environmental signals against personal baselines, tuning listening pace, end-of-turn timing, backchannels and speech speed; opt-in, EU AI Act framing, no voiceprints. Living memory: consolidation, decay ranking, windowed forgetting with restore, cluster co-activation across conversations. Recursion at three timescales; per-person lessons learned from corrections without retraining; five per-turn cognitive frameworks. Social signals drawn only from the user's own conversations with SERA, never from end-to-end encrypted member chat. Layered privacy: on-device encryption, optional passphrase-locked backups (local file or encrypted server sync), on-device 384-dimension sentence embeddings with local cosine validation, salted matching embeddings. First audio in under 200ms; reply latency 80-700ms by design. Architectural trade-offs, including perception error and lesson over-fitting. Cross-references component whitepapers: *Artificial Perception*, *Crisis & Duty of Care*, *Living Memory*, *Beyond the Feed*, and *Private by Architecture*. - [Private by Architecture - whitepaper](https://www.algn.io/research/private-by-architecture): Technical whitepaper defining the six-layer client-server privacy architecture: On-device encryption (AES-256-GCM, record-bound AAD, non-extractable keys failing closed); local .algn encrypted backup format with crypto-shredding; transit anonymization stripping IP and device metadata; binding zero data retention across speech, perception, and inference servers; ephemeral Situational Presence (Be) acoustic/physiological evaluation without raw biometric storage; multi-layer semantic matching with client topic stripping, local embeddings, and salt derivation; mutual consent introduction protocol; end-to-end encrypted messaging channels; client-side cryptographic verification (SHA-256 inspection, 24-hour Measured vs. Declared egress audits, on-device re-embedding). Includes formal threat model and trade-off analysis. - [What Builders Owe - research](https://www.algn.io/research/what-builders-owe): Ethical and legal research paper (v2.0, September 2026). Analyzes empirical findings indicating user affinity for disempowering conversational interactions (Sharma et al., 2026). Formulates a builder obligation framework based on care ethics, fiduciary theory, tort law, and empirical studies on sycophancy, acoustic intimacy, and attachment. Reviews legal frameworks across US, EU, and UK jurisdictions. Establishes six positive builder duties: epistemic responsibility, proportional safeguards, architectural maintenance, relational continuity, transparency regarding technical limitations, and contribution to systemic standards. Identifies relational continuity during model deprecation as an open technical challenge. - [The Invisible Cage - research](https://www.algn.io/research/the-invisible-cage): Cognitive research paper investigating recursive introspection limitations and behavioral pattern disruption. Argues that internal introspection functions as a recursive sub-process of the system under examination rather than an objective analytical viewpoint. Analyzes six psychological and contemplative traditions (vipassana, shadow work, psychedelic therapy, Stoicism, Internal Family Systems, Acceptance and Commitment Therapy), identifying deautomatization as their shared operative mechanism. Analyzes the systemic absence of deautomatization mechanisms across consumer software and outlines architectural requirements for software positioned between clinical intervention and commercial applications. ### Documents pending release The documents below are forthcoming and not yet published. Do not cite them as published findings: - [Self-Authored Life Mapping (SALM) - whitepaper](https://www.algn.io/research#self-authored-life-mapping): A layered framework for gathering, structuring and storing autobiographical information through conversation. Status: Pending release. - [Artificial Perception - whitepaper](https://www.algn.io/research#artificial-perception): Real-time prosodic analysis against a personal baseline, processed in memory and never stored. Status: Pending release. - [Crisis & Duty of Care - whitepaper](https://www.algn.io/research#crisis-and-duty-of-care): Distress detection and de-escalation in a voice AI built to challenge, designed around autonomy rather than monitoring. Status: Pending release. - [Living Memory - whitepaper](https://www.algn.io/research#living-memory): Consolidation, differential decay, pattern clustering and encrypted on-device storage for long-term memory. Status: Pending release. - [Beyond the Feed - whitepaper](https://www.algn.io/research#beyond-the-feed): Privacy-preserving matching, graduated identity disclosure and consent-based contact in a non-engagement social architecture. Status: Pending release. - [The Counterpart Architecture - research](https://www.algn.io/research#the-counterpart-architecture): What the evidence demands from a voice-first system built for pattern literacy. Status: Pending release. - [The Behavioural Philosophy of algn - research](https://www.algn.io/research#behavioural-philosophy): How a voice-first conversational system inverts the industry's default behaviours. Status: Pending release. - [The Algorithmic Self - research](https://www.algn.io/research#the-algorithmic-self): Research on how AI shapes identity, introspection and agency, and how algn architecture responds. Status: Pending release. ## Blog - [Blog](https://www.algn.io/blog): Evidence-led answers to common questions on attention span, close friendship, loneliness, overwhelm and doomscrolling. Each post opens with a sourced short answer, gives key figures by dataset and year, and states where evidence is thin, contested or untested. Datasets include Pew Research Center trust in government (1958-2025), Gallup (1990) and Survey Center on American Life (2021, 2024) friendship surveys, the American Time Use Survey via the US Surgeon General (2023), Gloria Mark's screen-switching studies (2004 to 2021), and papers in Science, PNAS and JAMA Pediatrics. Most long-run data is from the US. First cluster: It's not just you (pillar page and eight-minute film), built on *The Pattern Beneath the Symptoms* (September 2026), which rates every claim for strength; *The Invisible Cage* develops the report's reading further. - [What is AI sycophancy?](https://www.algn.io/blog/what-is-ai-sycophancy): Defines AI sycophancy: model responses that match user beliefs over truthful ones (Sharma et al., ICLR 2024, which found five state-of-the-art AI assistants consistently sycophantic across four free-form tasks, and that agreement with a user's views is one of the strongest predictors of a response being preferred in human preference data; likely driven in part by human preference judgements). Sharma, McCain, Douglas and Duvenaud (arXiv:2601.19062, 2026) analysed 1.5 million consumer conversations with one AI assistant from one week in December 2025: severe disempowerment potential in fewer than one in a thousand; around 8% moderate or severe in relationships and lifestyle against under 1% in software development; patterns included validation of persecution narratives in emphatic, flattering language. In more than 500,000 rated interactions (late 2024 to late 2025), flagged interactions received a higher share of thumbs-up than baseline. Limits: one provider, one week, country mix not reported, imperfect classifiers, no tracking of individual users over time, generalisability to other platforms and later models unknown. - [Do we really have a goldfish attention span?](https://www.algn.io/blog/goldfish-attention-span): Myth check on the "eight-second human attention span". Traces the figure (12 seconds in 2000 to eight in 2013) to Statistic Brain, cited in a 2015 Microsoft Canada report; when the BBC traced the sources (Simon Maybin, More or Less, 2017), none backed it. The goldfish comparison has no research either. Measured alternative: Gloria Mark's studies of how long office workers stay on one screen before switching (Attention Span, 2023): about 150 seconds (2004), about 75 (2012), about 47 (2016 to 2021), from separate studies with similar methods. That is screen-switching at work, not attention capacity. Whether people can focus less is an open question; a brain-level link is plausible but not tested directly. - [How many close friends do most people have?](https://www.algn.io/blog/how-many-close-friends): No single number: Robin Dunbar's research (from 1993) suggests layers of roughly five people in the innermost circle, then about 15, 50 and 150, though the exact numbers are debated. Lindenfors et al. (Biology Letters, 2021) found the data behind 150 too wide to support a single number. Three US national surveys show direction: Americans reporting no close friends rose from 3% (Gallup, 1990) to 12% (Survey Center on American Life, 2021) and 17% (Cox and Pressler, Disconnected, 2024); ten or more fell from 33% to 13% (1990 to 2021). In 2024, 24% with a high-school education or less had none, against 10% of university graduates. The 1990 and 2021 surveys were run differently, which may exaggerate the change; its size is uncertain. The decline predates smartphones; no one has proved which cause matters most. US data only. - [Is there really a loneliness epidemic?](https://www.algn.io/blog/is-there-really-a-loneliness-epidemic): Tests the "loneliness epidemic" label (US Surgeon General, 2023) against long-run data, mostly from the US. Finding: the decline is real but uneven, and "epidemic" overstates what the data show. Americans with no close friends: 3% (Gallup, 1990), 12% (2021), 17% (2024, Survey Center on American Life); different methods may make the change look bigger. In 2024, 24% with a high-school education or less had none, versus 10% of university graduates (Cox and Pressler). Americans aged 15 to 24 spent about 70% less time with friends in person in 2020 than 2003. Health evidence is separate and strong: loneliness linked to 26% higher risk of early death (Holt-Lunstad et al., 2015). These figures measure friendship, not loneliness itself. - [What is doomscrolling?](https://www.algn.io/blog/what-is-doomscrolling): Defines doomscrolling: scrolling through bad news or upsetting content long after meaning to stop (first known use 2020, Merriam-Webster). The feed design dates to 2006 to 2012: Facebook's News Feed (2006), Like button (2009) and ranking by likely reaction; in 2017 an angry reaction counted five times a Like (Haugen documents, 2021). Bedtime screen use is linked to roughly double the odds of not getting enough sleep in children and teenagers (Carter et al., JAMA Pediatrics, 2016). Each step of the stress and sleep loop has evidence; the full loop has not been tested directly. Ten days of feed re-ranking moved hostility towards the other party both ways, comparable to about three years of national change (Piccardi et al., Science, 2025); one experiment on one platform, still to be repeated. - [Why does everyone feel overwhelmed all the time?](https://www.algn.io/blog/why-does-everyone-feel-overwhelmed): Reads reported overwhelm against several long-run measures that moved the same way over the same decades (the report's reading). Anchor: 74% of UK adults felt overwhelmed or unable to cope at some point in the past year (Mental Health Foundation, 2018); one survey, not a trend. US series: share earning more than their parents at 30, about 92% (born 1940) to about 50% (born 1980s) (Chetty et al., Science, 2017); no close friends, 3% to 17% (1990 to 2024); trust in the federal government, 77% (1964) to 17% (September 2025, Pew). Proposed loop: allostatic load (McEwen and Stellar, 1993) and sleep loss; each step evidenced, the population loop untested. No dataset measures overwhelm itself; causation is not shown.