# Changelog Every published companion build has an entry below. Android packages are laboratory builds; iOS packages require Apple signing and provisioning. Historical validation belongs to the named version and does not establish current physical-device performance. ## [0.2.32](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.32): build 36 · 2026-11-17 **Guided companion onboarding and shared visual identity** · experimental-prerelease - Unify Android and iOS around the approved large cyan emblem, ink/navy panels, a clear next action, workload Activity and progressive disclosure for advanced settings. Preserve existing enrollment and reviewed native LLM binaries. - Simplify QR-first onboarding and manual fallback. Android registers Hive v3 pairing links. iOS provides labelled bottom navigation, a direct manual-invitation sheet and accurate English, Simplified Chinese and Italian permission guidance. - Make model preparation and work consent explicit before download. Keep saved pairing, fresh contact, ownership confirmation, model readiness and actual work distinct; show the actual local or Hive connection mode. - Extend the public animated diagram from enrolled resources to request assignment, whole-model execution, returned result and contributor visibility. Public credits, payments and guaranteed service remain inactive. **Validation:** 58 Android JVM tests, Android lint and APK build, 128-state Swift onboarding checks, EN/zh-Hans/IT catalog checks and 109 release-tooling tests passed. Signed iOS simulator exercised manual invitation, explicit trust, real disposable TLS pairing, pause and saved-session resume; separate simulator tests passed invalid-invitation rejection and navigation. Development-signed build 36 passed navigation preserving enrollment on a physical iPhone 15 Pro Max before final display-only copy polish. Public iOS ARM64 build and APK embedded identities, source files, Android signer continuity and retained native LLM hashes verified. Static external review and public diagram responsive/locale checks passed. **Limits:** Android camera, installation and enrollment on physical hardware are unqualified for this release: the Mac Android emulator crashes before opening the app. iOS public archives are unsigned and require personal signing/provisioning; no Safari installation, TestFlight or App Store release. Physical iPhone navigation is not physical QR/network/inference qualification. No native-engine optimisation, distributed-memory, general acceleration, reward credit or energy-saving claim. ## [0.2.31](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.31): build 35 · 2026-11-16 **Visible mobile connection mode and coordinator** · experimental-prerelease - Show Current connection at the top of Android/iOS Home with the actual saved protocol mode, coordinator hostname/port and session state. Endpoint display excludes credentials, paths and query strings. - Explain when a device is still connected to a local coordinator instead of direct Hive. Show cluster/region as unreported when not supplied by the hub; never infer hub assignment from device locale or an IP address. - Keep the private customer confirmation code prominent for protocol-3 Hive enrollment. Preserve local pairings and unchanged native inference libraries; include English, Simplified Chinese and Italian. **Validation:** 56 Android JVM tests, lint/APK and unsigned iOS ARM64 compilation passed. Swift connection projection/secret-code vector and three-language checks passed. Exact Android signature continuity and retained Android/iOS native LLM hashes verified. Separate deployed backend: 502 host/71 Django tests; real two-account HTTPS and two resident/Hive token-equivalent pairs; browser task completed on a co-located Linux CPU worker. Those backend tests are not physical mobile qualification. **Limits:** 0.2.31 mobile connection card/installation and direct customer enrollment have not yet been physically qualified. On previous0.2.30, actual Xiaomi contact confirmed local protocol2 Mac binding, explaining absence of a direct-Hive customer code. No distributed memory, general acceleration or energy-saving claim. ## [0.2.30](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.30): build 34 · 2026-11-16 **Customer enrollment confirmation on the companion Home** · experimental-prerelease - Display an eight-character confirmation code in Home for protocol-3 Hive enrollment. The code is derived from the private worker credential, not the public node ID, and does not grant work consent or prove physical hardware identity. - Prepare the Android and unsigned iOS companion interfaces for the separately rolled-out customer workspace portal. Preserve existing saved pairing and native LLM library bytes. English, Simplified Chinese and Italian are included. **Validation:** Android 56 JVM tests, lint/APK build, unsigned iOS ARM64 compilation and English/Chinese/Italian runtime localization checks passed. Exact Android signer continuity and retained Android/iOS native-library bytes were verified. Customer backend isolation and portal tests are recorded separately from this companion source tag. **Limits:** No physical installation or camera/code display qualification yet for this build. The package is prepared for the customer portal enrollment test; no new distributed execution, shared RAM or performance gain is claimed. Existing operator-owned enrollments are not automatically transferred to a customer workspace; explicit migration is a separate operation. ## [0.2.29](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.29): build 33 · 2026-11-16 **Visible update prompts, scheduled Android downloads and clearer Hive capacity** · experimental-prerelease - Automatically present release notes when a newer mobile build is found at process launch or the seven-day deadline. Persist an exact-build deferral and support one-hour, next-day and seven-day reminders. - Android shows the changelog before download consent, verifies package length, SHA-256, identity and signer, then opens the Android installer while the update view is foregrounded. Scheduled downloads start only when the app is active at or after the chosen time; Android retains installation confirmation. - iOS presents release notes and persistent reminders, clearly explaining that the published unsigned IPA still requires personal signing and Mac installation. No App Store or TestFlight channel is claimed. - Hive offers owner sessions of nine hours, one week or thirty days, with absolute server-side expiration, persistent HttpOnly cookies, CSRF and logout/revocation checks. Existing authenticated sessions can select a duration from Account without a new login. - Separate historical sustained FP32 qualifications from fresh scheduler admission. Expose per-node and aggregate blockers such as absent contact, missing policy, expired qualification and changed runtime; historical rates never become current capacity. **Validation:** 485 host tests, 56 Android JVM tests, Android lint/build, iOS scheduling/release-metadata policies and unsigned iOS ARM64 compilation passed. Android signer continuity and exact reuse of the previous Android/iOS native LLM libraries were verified. Live portal behavior and publication receipts are recorded separately; mobile update UI and installation have not been physically qualified. **Limits:** App-active update timing only; a closed or OS-suspended app checks or services reminders after returning. Android still requires OS installation consent; iOS packages remain unsigned and are not directly installable. This release does not qualify new physical device performance, pooled model memory, acceleration or energy savings. ## [0.2.28](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.28): build 32 · 2026-11-15 **Hive-owned complete-model requests and direct mobile workers** · experimental-prerelease - Add persistent Hive-owned LLM requests, model hash admission, bounded greedy generation, fenced worker leases, incremental output, cancellation and immutable result replay. Device permission and operating windows remain separate; the one-request exception is explicit. - Add direct Android and iOS complete-model worker adapters and authenticated model streaming from the private Hive catalog. Model execution stays on the selected phone; the Mac is not an inference host or request relay. - Add the LLM tab to Hive with model/node selection, output and separate queue, first-token, prefill, generation, control-wait and service timings. English, Simplified Chinese and Italian messages are included. **Validation:** 475 host tests, 55 Android JVM tests, APK/lint and unsigned iOS ARM64 compilation passed. A separate unprivileged native Linux worker co-located with Hive completed 19 ordinary requests and 18 matched native-reference comparisons. Two injected failures preserved cancellation/expiry and rejected late output. Authenticated model streaming transferred 491,400,032 hash-matching bytes. Android installed on OPPO, but its locked screen prevented physical LLM qualification; iOS physical execution and live owner click-through remain pending. **Limits:** Complete-model placement only: no model splitting or shared model memory. One operator-provisioned default mobile model, at most 1 GiB, context up to 4096 and 256 output tokens, greedy text only. Results and device phase times are authenticated worker reports, not independent semantic attestation. Hive times use server acceptance and receipt boundaries. No acceleration or energy saving is claimed. iOS archives require personal signing and physical qualification. ## [0.2.27](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.27): build 31 · 2026-11-15 **Explain unavailable Android energy sensors and separate battery state from USB power** · experimental-prerelease - Distinguish external power flags from full/not-charging battery status. Accept discharge power only with no reported external source, a compatible status, valid voltage and a strictly negative supported instantaneous current. - Expose allowlisted raw electrical diagnostics on Hive and in the local diagnostic export: battery status, external power flags, voltage, instantaneous current and average current. Preserve exact unavailable-current and sign-conflict reasons without absolute-value conversion or nominal estimates. - Keep measurements, estimates and missing values separate. Existing pairing, credentials and native LLM library bytes are preserved. iOS receives the matching compiled package identity; its public W/J API limitation is unchanged. **Validation:** 446 host tests, 55 Android JVM tests, APK/lint and unsigned iOS ARM64 compilation passed. Android retains the same signer and native LLM bytes. Synthetic browser diagnostics preserve signed current values and keep power unavailable on contradictory evidence. Read-only physical OPPO investigation confirmed USB power; field-specific post-unplug attribution awaits the diagnostic build. **Limits:** The OPPO sensor failure cannot be attributed to a particular raw field until the diagnostic build reports it. No Android power measurement or energy saving is claimed merely from package compilation. Positive-current polarity is not automatically inverted, averages do not silently replace missing instantaneous current, and connected USB/charging readings remain excluded as device consumption. iOS artifacts are unsigned and require personal Apple signing. ## [0.2.26](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.26): build 30 · 2026-11-15 **Automatic supported power sensors and honest energy coverage** · experimental-prerelease - Collect bounded Android battery-terminal discharge samples automatically during an active node session. Exclude charging, external power and unsupported OEM readings. Continue acquisition during CPU work and upload via authenticated heartbeat/poll/progress paths. - Show per-node source, scope, freshness, unknown calibration, recent power traces and separate measured/estimated subtotals on Hive. Integrate common covered windows without extrapolation or overlap double counting. - Explicitly report the iOS public-API limit instead of deriving watts or joules from charge percentage. Preserve enrollment and existing LLM native-library bytes. Windows Node 0.7.1-preview.1 separately adds supported driver power collection with its bundled runtime. **Validation:** 96 focused host tests and 10 frontend behavior tests passed. Android 53 JVM tests, debug APK build and lint passed; signer and reused native LLM bytes match the previous 0.2.25 package. Unsigned iOS ARM64 build passed. Synthetic browser checks covered English, Simplified Chinese, Italian and 390/1440 widths. A separate physical Mac battery run recorded eight samples and integrated a 35-second window. **Limits:** Physical Android and Windows power-sensor qualification remains pending. Battery discharge is gross device consumption and unavailable as a device total on external power. NVIDIA readings cover GPU boards; RAPL remains estimated. No calibrated wall-meter comparison, application-only attribution, cross-node clock attestation, lifetime energy accounting or energy-saving claim. iOS archives require personal Apple signing and expose no live W/J sensor. ## [0.2.14](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.14): build 17 · 2026-11-05 **Find the paired computer after a local IP change** · experimental-prerelease - Add protocol-2 host identity independent of the IP address and native Bonjour/DNS-SD discovery in Android and iOS. Validate the original pinned certificate and logical DNS name before sending credentials. - Preserve node identity, downloaded drafts and settings across validated endpoint changes. Bound discovery, retry and cancellation; clear stale in-memory draft state. The host follows its selected private IPv4 route independently of QR creation. - One new QR is required to migrate a legacy pairing. Subsequent address changes use rediscovery on reachable local networks. Add English, Simplified Chinese and Italian recovery messages. **Validation:** 175 host tests, 28 Android JVM tests, Android lint (0 errors, 18 warnings), unsigned iOS ARM64 build, pairing and localization checks passed. Actual Swift TLS checks on Mac preserved node identity across two endpoints and rejected wrong certificates and DNS names. NWBrowser discovered a real local advertisement on Mac. **Limits:** Physical Android/iPhone roaming is not yet tested. Discovery needs reachable private IPv4 and permitted local multicast; client isolation can block it. The host needs a native DNS-SD publisher. A pinned certificate lasts 365 days; authenticated renewal is not implemented. Android is debug-signed; iOS is unsigned and needs Apple signing. Native engine binaries are unchanged from 0.2.13, including its limitations; later source-only patches are not included. No inference speedup is claimed. ## [0.2.13](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.13): build 16 · 2026-11-03 **Native timing and bounded Android lookahead experiments** · experimental-prerelease - Separate native prefix synchronization, decode diagnostics and worker durations. Record late and uncommitted proposals on the host. - Add an explicit one-block Android lookahead experiment with exact committed-prefix matching, sequence checks and expiry. Serial target verification remains the default. - Handle expired Android inference leases without breaking the session. Harden chat control-token handling, response allocation failures and cooperative cancellation during native model loading. **Validation:** 163 host tests, 25 Android JVM tests, Android lint (0 errors, 18 warnings), Swift protocol/transfer checks on macOS and unsigned iOS ARM64 compilation passed. The final host library passed 31 real-weight serial checks and 15 boundary checks; four additional failure-injection cases passed in a separate test build. Five real-weight TLS loopback cases preserved output. APK optimization flags and published package hashes were verified. 135 local pilot runs compare an 8B target and 1B draft, retaining output mismatches. GLM 5.3 Max and Gemini 4 Argon High reviewed the supplied design; they did not execute tests. **Limits:** Physical testing of this Android/iOS build is pending. Android has explicit bounded lookahead; iOS has timing support and does not advertise that pipeline. Batch verification can diverge from the serial reference and remains experimental. Exact-output local variants did not establish a speedup. ACK wait, failed compute and missing cleanup reports remain incompletely measured. Android is debug-signed; iOS requires Apple signing and provisioning. No distributed speedup, pooled memory or measured energy saving is claimed. ## [0.2.11](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.11): build 13 · 2026-11-02 **Model metadata and transparent companion updates** · experimental-prerelease - Show Hugging Face repository update and creation dates, task, reported parameters, context, license, languages, base models, gated access and popularity when supplied. Missing metadata stays unknown. Repository dates are not original model release dates. - Show matching release notes before an explicit direct APK or unsigned IPA download. Validate version, build, publication date and artifact metadata; invalidate stale details after failed checks. Current releases can also be reviewed. - Allow an explicitly confirmed legacy enrollment to be retired into a stable enrollment while retaining assessment and job history. Established identities are never merged by device name or IP address. **Validation:** 118 host tests, five desktop JavaScript suites, 21 Android JVM tests, Android lint (0 errors, 17 warnings), both Swift policy suites and unsigned iOS ARM64 build passed. Live Hugging Face metadata rendering and one retained legacy/stable enrollment were verified on the local dashboard. Gemini 4 Argon High supplied a textual UX review. Physical device installation, accessibility and browser download handoff remain unqualified. **Limits:** Android remains debug-signed with the existing laboratory signer. iOS requires Apple signing and provisioning; downloading an IPA cannot install it. Browser downloads are not hash-verified by the companion. No new inference speedup or shared-memory capability is claimed. HarmonyOS/NEXT builds remain deferred. ## [0.2.10](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.10): build 12 · 2026-11-01 **Backend selection fixes and synchronized companion builds** · experimental-prerelease - Publish fresh Android and iOS companion builds from the current source and shared English, Simplified Chinese and Italian catalogs. Retain the Home, Activity and Settings interface, assessment visibility and session-generation guards from 0.2.9. - Include the desktop backend catalog-loading controls, matching provider configuration, inline errors and native main/draft explanations in the release source. Verify local LM Studio text generation and public OpenRouter catalog selection separately. - Update current download links and installation documentation. Native inference libraries are unchanged; no new phone acceleration capability is claimed. **Validation:** Android build, 18 JVM tests and lint (0 errors, 17 warnings), plus unsigned iOS ARM64 build passed. Verified existing Android signer, current localization catalogs, unchanged native libraries, source parity and identical ZIP/IPA app payloads. Desktop backend controls have separate live LM Studio and public OpenRouter catalog evidence. No physical-device qualification. **Limits:** Android remains debug-signed with the existing laboratory signer. iOS packages are unsigned and require Apple signing and provisioning. Physical installation, accessibility and mobile LLM speedup are not qualified. HarmonyOS/NEXT native builds remain deferred. ## [0.2.9](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.9): build 11 · 2026-11-01 **Device assessment, session visibility and simpler companions** · experimental-prerelease - Show authenticated peer IP, enrollment dates, latest contact and observed session boundaries for devices; show host LAN address and process session start. Legacy dates remain unknown until observed. - Automatically schedule bounded sequential network and synthetic CPU assessment for newly enrolled devices, with explicit retry and classifications based on returned measurements. Low battery, severe thermal state and host inference defer automatic work. A changed peer address invalidates the prior connection classification. - Replace ambiguous contribution labels with explicit host target/verifier role, no selected run, external route, unused phone, missing accounting or measured zero. Basic diagnostics never claim LLM speedup or shared RAM. - Reorganize Android and iOS into Home, Activity and Settings, separate connection from draft readiness, expose one next action, move advanced controls into disclosure and make diagnostic export directly available. Retain identity confirmation for QR and manual pairing. - Age the connection indicator from successful authenticated responses so a stale session cannot stay marked active indefinitely. This is status reporting, not a claimed fix for the unresolved physical network timeout. **Validation:** 115 host tests, five JavaScript suites, 18 Android JVM tests, Android lint (0 errors, 17 warnings) and unsigned iOS ARM64 compilation passed. An isolated browser and synthetic TLS worker exercised enrollment and sequential assessment. Gemini 4 Argon High provided a textual UX architecture review. Physical accessibility, networking and LLM speedup remain unqualified. **Limits:** Physical Android/iOS accessibility and stable mobile inference are not qualified. Basic assessment is HTTPS plus a synthetic CPU kernel; model-specific usefulness still needs a matched native comparison. Session boundaries are inferred after a contact gap over 20 seconds. iOS packages require Apple signing. HarmonyOS/NEXT remains deferred. ## [0.2.8](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.8): build 10 · 2026-11-01 **Stable device enrollment and local dashboard recovery** · experimental-prerelease - Android and iOS store a private installation identity in Android Keystore-encrypted storage or iOS Keychain. Pairing derives a separate key for each host, so scanning another QR reuses the same node and renews its access credential. No IMEI, MAC address or hardware fingerprint is collected. - Existing enrollment is adopted only after an invitation-bound proof of the saved credential. Different phones with the same model remain separate. Interrupted pairing can recover with a new QR and the persisted identity. - The desktop offers a protected Restore local session button with localized feedback. Owner-confirmed legacy consolidation preserves diagnostic jobs and their original enrollment IDs. **Validation:** Android build, 14 JVM tests and lint (0 errors, 17 warnings), unsigned iOS ARM64 build, 10 enrollment identity tests, 24 host protocol/network tests and real Swift TLS repeat enrollment on macOS passed. Kotlin/Swift identity and proof vectors match the host. Physical-phone repeated QR qualification remains pending. **Limits:** Update the desktop host before pairing with 0.2.8. Removing Android app data or losing identity storage creates a new installation; automatic physical-device recognition after identity loss is not claimed. Older companions do not send identity proof. HarmonyOS/NEXT builds remain deferred. No inference performance claim. ## [0.2.7](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.7): build 9 · 2026-11-01 **Current-network QR pairing and the approved 3D emblem** · experimental-prerelease - Use the approved blue three-dimensional showcase emblem in Android/iOS launcher icons and native headers, the GitHub README and HarmonyOS source assets. Keep the simplified mark for small monochrome system surfaces. - The host follows its current private IPv4 route when creating a pairing QR, validates the local bind and retires stale invitations after a network move. Per-address TLS identities preserve certificate and hostname verification; companions scan a fresh QR after an address change. Enrollment and measurement history remain available. - Android routes paired-host requests through a matching connected Wi-Fi/Ethernet network when available, retaining system routing for tethering. Network timeouts and TLS identity failures have distinct localized guidance. Exported diagnostics include only allowlisted route events and available transport categories. - Desktop graph links use color and width for measured accepted draft token share. Selected running draft routes animate; completed, offline, unknown and fallback routes stay still. A motion toggle, background pause and reduced-motion support are included. - Publish a pinned Locally Uncensored source assessment with attribution and clear boundaries between independently implemented diagnostics and future planner/cancellation requirements. **Validation:** Android compilation, 13 JVM tests and lint (0 errors, 19 warnings), unsigned iOS ARM64 device build, 24 host protocol/network tests, localization and graph accounting checks passed. Browser fixture verified weighted animation and the motion toggle using explicitly synthetic data. Android signer continuity, packaged artwork, unchanged native LLM libraries and identical ZIP/IPA app payloads were verified. Physical-device reconnection and installation remain unconfirmed. **Limits:** No physical Oppo/Xiaomi installation or camera qualification in this build. Token share is not compute utilization or a speedup. iOS packages require Apple signing. HarmonyOS/NEXT source artwork is updated but native builds remain deferred. Native LLM libraries are reused; no engine performance claim. ## [0.2.6](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.6): build 8 · 2026-11-01 **Reliable controls, local diagnostics and automatic update checks** · experimental-prerelease - Replace the Android legacy scanner activity with lifecycle-bound CameraX, an offline ZXing decoder, retry/cancel controls and a camera-open timeout. Permission denial and camera errors return actionable feedback; invitation validation and identity confirmation remain required. - Android controls distinguish available, disabled, pressed, selected and busy states. Session status, engine permission and actual draft readiness are separate; saved engine permission is synchronized after service changes. - Desktop actions show request progress across polling updates and block duplicate submissions. Narrow-screen Add device labels and QR spacing are repaired. iOS controls gain explicit pressed/disabled styling and connection progress. - English, Simplified Chinese and Italian messages, Android APK and unsigned iOS packages share version 0.2.6/build 8. Native LLM libraries are reused without engine changes. - Add bounded local Android/iOS diagnostic journals with explicit export and clear, sanitized managed exception context, supported Android process exit reasons and reduced iOS MetricKit events. Previous interruption markers do not establish a crash cause. - Check official companion release metadata at launch and weekly while running, with due checks after OS suspension, offline retry and an explicit Check now action. Installation remains manual. - Use the current original orbital NodeLift mark in native headers and launcher icons. Add project acknowledgements and preserve original third-party licenses and source. **Validation:** Android app/instrumentation compilation and lint, ten JVM QR/support policy tests, unsigned iOS device compilation, Swift journal/update checks, 19 host protocol tests, localization and focused desktop control tests passed. Android signing continuity and unchanged native libraries were verified. Actual external source/design reviews are scoped in release evidence. Physical scanning, sharing and weekly scheduling on a phone were not tested. **Limits:** The reported Xiaomi crash remains unconfirmed without a device log; install and exercise the new APK. Logs are best effort and cannot capture every native crash. OS suspension delays checks until execution resumes. iOS packages require Apple signing; TestFlight/App Store distribution is planned, not available. HarmonyOS/NEXT builds remain deferred. No new inference speedup or pooled-RAM result. ## [0.2.5](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.5): build 7 · 2026-11-01 **Desktop QR pairing and native mobile scanning** · experimental-prerelease - Add device on the desktop displays an offline QR; Android and iOS companions scan it and request address/fingerprint confirmation. Manual invitation import remains available. - Five-minute, single-use invitations expire, clear after use and are revoked when explicitly closed. Camera access is requested only for scanning; pairing does not enable inference automatically. - Android retains the existing debug signer and saved pairings, with an explicit host leaf-certificate pin. Update and restart the desktop host to use the new pairing flow. - Android ARM64 APK, unsigned iOS IPA/app ZIP and SHA-256 manifests are published. Native engine libraries are unchanged from 0.2.4. **Validation:** Android and unsigned iOS device compilation, host protocol and desktop lifecycle checks, Swift parsing and loopback TLS tests passed. Independent ZXing decoding recovered the QR from a desktop screenshot. Published package hashes were verified. **Limits:** Physical Android/iOS camera scanning remains unqualified. iOS requires personal Apple signing/provisioning and cannot install directly from Safari. HarmonyOS/NEXT builds are deferred. No inference speedup or RAM-pooling claim. ## [0.2.4](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.4): build 6 · 2026-11-01 **About, references and manual release discovery** · experimental-prerelease - Desktop and companions expose authorship, installed version/build, official project links, technical references, licenses and the official GitHub release list. - Update discovery is manual; no silent update check or installation. Installed Android/iOS versions replace hardcoded telemetry values. - Host 0.3.0 preview metadata is separate from companion 0.2.4. Android signer and native libraries match 0.2.3; iOS IPA and ZIP contain the same unsigned app. **Validation:** 717-key catalog parity, JavaScript/Swift formatters, graph checks, 16 focused host protocol tests and desktop browser checks passed. GLM 5.3 Max reviewed selected About source; no Gemini review for this change. **Limits:** No physical Android/iPhone installation or UI qualification. Native HarmonyOS/NEXT remains uncompiled without a HAP or native LLM worker. No new native-engine performance result. ## [0.2.3](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.3): build 5 · 2026-11-01 **Italian UI and host integration preview** · experimental-prerelease - Italian joins default English and Simplified Chinese; 557 shared messages cover pairing, diagnostics and model/provider controls. Saved choices persist and prompts/model output are not translated. - Tagged host source includes the 0.3.0 integration preview: public Hugging Face GGUF search/download, direct import, explicit OpenRouter and loopback local-server routes. - Supplementary unsigned IPA packaging was published on 1 November with a byte-identical app payload and a separate manifest; original build artifacts were unchanged. **Validation:** Catalog/placeholder/freshness, JavaScript/Swift formatter and live browser checks passed. Android app/instrumentation APKs and unsigned iOS device app compiled; native binaries match 0.2.2. **Limits:** Physical mobile UI and Android instrumentation execution remain unverified; no ADB device was available. No new GLM/Gemini review for Italian. iOS requires signing; no HarmonyOS HAP or new speed benchmark. ## [0.2.2](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.2): build 4 · 2026-10-31 **Default English and Simplified Chinese localization** · experimental-prerelease - Desktop and companions start in English, offer Simplified Chinese and retain an explicit offline language choice across sessions; initial catalog contains 440 shared entries. - Language switching preserves user content and unapplied settings; the API-key panel clears sensitive access data. - Tagged source includes the earlier measured adaptive-control/context-lookup work. Packaged native binaries are unchanged from 0.2.1. **Validation:** 40 host tests, localization/formatting checks and live browser checks passed. Android and unsigned iOS builds compiled. GLM 5.3 Max reviewed 15 terminology entries; Gemini High was usage-limited. **Limits:** Physical localization and sustained active-node switching remain unverified. Android instrumentation did not execute because the emulator stayed offline. No physical-phone acceleration or pooled RAM established. ## [0.2.1](https://github.com/sev7enITA/NodeLift/releases/tag/v0.2.1): build 3 · 2026-10-31 **Native inference and authenticated mobile draft workers** · experimental-prerelease - Pinned llama.cpp core, greedy block verification, correction/bonus tokens, KV resynchronization, identity checks, pinned TLS leases and bounded remote fallback. - Android ARM64 CPU and iOS Metal/CPU workers, model integrity and battery/thermal/lifecycle controls; authenticated text/SSE gateway and LM Studio generator adapter. - Actual Gemini 4 High and GLM 5.3 Max design/source reviews were adjudicated. The frozen small-pair local-draft block-4 control reached 0.719× host-only speed, with all 48 paired token sequences matching. **Validation:** 26 host tests, 29 real-weight native checks, four real-weight TLS loopback fault cases, live gateway checks and three adapter stream tests passed. Android and unsigned iOS device builds compiled. **Limits:** The remote test node was emulated. Physical-phone inference/speedup, sustained thermals, capacity sharding, hybrid speculative checkpoints, HarmonyOS native worker and installed LM Studio GUI qualification remained open. The 27B upstream kernel result is not NodeLift hybrid-model support. ## Host, documentation and tooling updates ### 2026-11-16: Hive account invitations with expiry, revocation and separate SMTP status - Add Users → Invitations to the authenticated Hive owner interface, with English, Simplified Chinese and Italian controls, bounded expiry, immediate revocation, explicit replacement and paginated history. Keep account invitations separate from device pairing. - Implement approved-admin APIs and an encrypted SMTP outbox with atomic claims, lost-response idempotency, per-recipient limits and no implicit retries after uncertain delivery. Preserve operator approval and customer MFA. SMTP acceptance does not confirm inbox delivery. - Qualify account and native routes, actual browser creation/replacement/revocation, setup helper safety and three real TLS/SMTP loopback messages. The selected invites@nodelift.app mailbox is active; private SMTP configuration remains pending and external delivery is not enabled or qualified. No companion build changes. ### 2026-11-16: Hive-governed public/private model catalog and verified peer delivery - Deploy separate curated-public and MFA-protected workspace-private model catalogs, immutable file/chunk manifests, scoped storage-peer credentials, per-chunk authorization, revocation and bounded resumable HTTPS transfers. Pirate Face is credited as prior art; no code is copied and no public torrent announcement is made. - Complete 30 real verified transfer/cache operations in two separate Mac/VPS series. Final battery-series medians: direct HTTPS 19.990 s, authorized WAN peer 25.093 s, co-located peer 8.437 s, two sources 11.286 s and cache reuse 0.697 s. All 15 final Mac sensor intervals measured; VPS energy unavailable. Preserve per-series variability and failures, without inference acceleration or pooled-RAM claims. - Qualify actual standalone Model Peer 0.1.0-preview.1 build 2 on macOS ARM64, including interruption, 134,217,728 verified resumed bytes and full model/license verification. The bundled-runtime CLI is ad-hoc signed, not Developer ID signed/notarized; no mobile binary changes. - Verify 33 live public HTTPS authorization checks with isolated disposable accounts, plus scoped frontend controls and transport regression tests. Publish English documentation/figures and English, Simplified Chinese and Italian portal/showcase copy. ### 2026-11-16: Customer password recovery with private one-use links - Add Forgot password to the customer portal in English, Simplified Chinese and Italian. The pilot explicitly uses operator-assisted identity verification; no recovery email is sent. - Protect private reset links with a 15-minute lifetime, atomic single-use consumption, password/MFA-state binding, existing Authenticator verification and rate limits. Preserve account approval, clusters and device enrollment; invalidate password sessions and developer keys after success. - Verify account-service recovery and file-backed concurrent consumption, the actual shipped recovery controller and a disposable-account public HTTPS reset/replay/sign-in flow. No real customer password or MFA was changed; companion packages remain 0.2.31 (35). ### 2026-11-16: Customer account password minimum of eight characters - Accept account passwords of at least eight characters in the server and customer portal. Keep common-password, entirely numeric and account-similarity validation, with a localized rejection message in English, Simplified Chinese and Italian. - Verify rejected passwords preserve the unused invitation, accepted eight-character passwords can register and sign in, and the customer portal catalogs remain synchronized. Existing account passwords and companion packages are unchanged. ### 2026-11-15: Expose the complete Hive device pool and the missing distributed-model executor - Show all registered devices in the LLM view, including forwarded, stale and non-ready entries with explicit exclusion reasons. Complete-model admission remains restricted to fresh permitted direct workers. - Expose pinned Qwen2.5 72B (145.4 GB) and Kimi K2 (1.029 TB) checkpoint metadata for inspection only. No large weights were downloaded, no 1 GiB runtime bound removed and no Hive model partitioning claimed. - Retain six actual OPPO native/Hive token-matched comparisons, controller interruption and the final instrumentation socket error. Separate NAT device and owner request budgets without changing authentication limits; add hidden-tab polling/backoff tests. ### 2026-11-15: Automatic energy telemetry for local and remote enrolled nodes - Add unprivileged macOS battery, NVIDIA driver and Linux RAPL collectors; keep measured sensor paths, model counters and unavailable sources distinct. - Use the same authenticated sensor contract for directly enrolled hub workers and inventory-forwarded local nodes. Preserve actual hub/group attribution; planned regions do not add measured capacity or energy. - Publish portable Windows Node 0.7.1-preview.1 for x86, x64 and ARM64 with bundled runtime, automatic supported NVIDIA power probing and no new GPU compute or LLM execution claim. ### 2026-11-15: Unified Hive overview, four-region planning and instrumented energy evidence - Unify actual hub/group membership, four explicitly planned regional zones, declared geography, memory, profiled compute, historical delivered work, energy coverage and devices in the default Hive Overview. Planned zones are Americas, Europe, Asia Pacific and Middle East / Africa; regional federation remains undeployed. - Clarify worker invitations, Xbox/TV code approval and owner-console access on a phone. Compact completed onboarding, add device search/contact filters and preserve expanded details across refresh. Surface existing reported accelerator/memory-domain/API evidence without promoting it to model qualification. Localize new interface text in English, Simplified Chinese and Italian. - Accept optional authenticated/fenced, instrumented job-energy evidence with full monotonic-window coverage, sample-gap checks and retained uncertainty. Integrate W samples into J/Wh; expose per-job coverage without adding unrelated windows or overlapping sensor scopes. No automatic collectors, physical energy result, energy saving or new mobile binary is included. - Requested static GLM 5.3 Max and Gemini 4 Argon High reviews completed; the local uncensored LM Studio review returned only a partial response. Model suggestions were checked against protocol, geolocation and measurement contracts; they are not physical or visual test results. - Inspect official EdgeAI-Trust, dAIEDGE, EDGELM and CIREN sources, including the retrieved dAIEDGE D3.2 requirements report. Publish fourteen explicitly proposed NodeLift evidence gates; distinguish project requirements from applicable law and FL encrypted aggregation from encrypted LLM inference. No compliance certification is claimed. ### 2026-11-15: Desktop package identities and native CPU node previews - Name Windows worker packages by OS, architecture, CPU/CUDA backend and portable versus update delivery. Retain legacy private URLs and historical hardware attribution. Portable 0.6.2 includes Python, dependencies and the control model; Update 0.6.1 reuses an installed runtime. No Vulkan backend is included. - Publish the Windows x86, x64 (Intel/AMD) and ARM64 Node 0.7.0-preview.1 qualification packages with native launchers, bundled Python and CPU math. Windows physical startup/performance remain untested; no GPU/NPU or LLM executor is bundled in these generic previews. - Update English documentation, wiki and English/Simplified Chinese/Italian download guidance. Compilation, static checks and renaming are not acceleration evidence. ### 2026-11-14: Custom domain and global Hive gate - Activate trusted HTTPS on hive.nodelift.app, preserving the persistent hub identity and node state. Use nodelift.app for the public project site and hive.nodelift.app for authenticated enrollment and owner control. - A real HTTPS test with separate owner and phone browser contexts and a simulated native client exercised single-use authorization, enrollment, heartbeat, replay rejection and logout. This is not physical Xbox pairing or a distributed inference benchmark. ### 2026-11-13: Xbox first launch and phone pairing source update - The physical P16 compiled Xbox UWP x64 version 0.1.0 with zero errors and warnings. Owner television screenshots confirm that the interface runs on Xbox One X. The observed 1.00 GiB memory limit is the UWP application budget. - Version 0.1.1 source replaces long manual invitations with a bounded phone QR authorization flow, retains version-2 saved-session compatibility, clarifies missing qualification steps and displays the installed package version. Windows compilation and physical pairing of this update remain pending. No Xbox LLM/GPU worker is included. ### 2026-11-11: Physical P16 accelerator and CPU partition controls - The P16 completed CPU and DirectML kernel checks with provider attribution and a CPU/CUDA control-model repeat. The strict CPU/CUDA comparison failed at token 37 on one prompt; retained pre-sampling scores show a ranking flip sensitive to backend and evaluation history. No correction token is forced. - A compatible CPU/RPC profile completed physical Mac/P16 controls and held-out requests with matching 64-token sequences, real remote model buffers, remote KV placement and remote graph calls. Cutting the required transport stopped generation. The medium-model control also completed. These bounded experiments do not demonstrate a model above 64 GiB, a general speedup or a production worker. - The P16 graphical worker pilot exposes startup, file verification, actual activity, stop and failure detail. Retry handling prevents transient Windows progress-file sharing conflicts from terminating the worker. ### 2026-11-11: Modular desktop and guided mobile setup - The desktop separates modules and distinguishes recent contact, inventory, worker readiness and measured contribution. Hardware totals, process headroom and model-usable capacity retain different labels and units. Contextual help and per-node workload details are available in development source. - Private 0.2.19 Android/iOS installations add guided setup, visible model-download activity and clearer error paths. Android has a consent-based APK update route; iOS remains constrained by Apple signing and distribution. Public companion downloads remain 0.2.14. ### 2026-11-09: Desktop startup repair and capability census - The local Mac candidate embeds the native capability probe and reuses the authenticated running host on repeated launches. Foreign port conflicts produce a bounded explanation rather than an unhandled address-in-use traceback. Ad-hoc signing, one-Mac coverage and wider stability limits remain explicit. - Authenticated QR enrollment and heartbeats carry a versioned capability profile. CPU extension coverage, shared/dedicated/unknown memory domains, detected accelerators and verified microkernels inform a host-owned worker catalog. GPU/NPU model execution requires separate qualification. ### 2026-11-05: Standalone host preview 2 with authenticated network rediscovery - Publish a self-contained macOS arm64 host with embedded Python and protocol-2 discovery. On one M3 Max, startup, dashboard, branding, inventory, an eight-token native generation, a protocol-2 invitation and a running discovery publisher passed. Physical phone roaming and other Macs remain unqualified. The binary is ad-hoc signed and not notarized. ### 2026-11-03: Local Qwen review and verified request-boundary corrections - Run the pinned HauhauCS Qwen3.8 27B Q4_K_P model locally through NodeLift, retaining source packets, model identity, raw responses, truncation markers and independent finding adjudication. MTP is disabled; this is static code review, not a distributed performance result. - Reproduce and fix a provider cancellation window identified by Qwen: cancellation during connection setup must be checked before a generation POST is sent. A real loopback regression fails before and passes after. OS DNS/connect is not promised to stop instantly. - Independently correct supplementary Unicode conversion at the JNI boundary and reject null worker totals as missing telemetry without discarding valid tokens. Validation: 167 host tests, 15 real JVM/JNI cases, Android ARM64 object compilation and eight real-weight native counterchecks. The JNI correction is source-only; published companion build 16 is unchanged. - Publish the observed TLS connection-slot pressure graph and refresh current Android/iOS evidence descriptions. Physical post-fix phone performance and rapid lifecycle transitions remain unqualified. ### 2026-11-03: TLS connection-slot diagnostic and English presentation images - A physical serial Android diagnostic reached all 32 companion TLS connection slots before lease expiry. The source host now closes companion TLS responses promptly; 165 host tests pass, including more clients than the slot limit and retained local HTTP persistence. Physical post-fix validation and handshake overhead remain unmeasured. No new mobile binary is implied. - Replace the README and gallery presentation image with a real English browser capture. Preserve dated historical evidence and require English for future presentation screenshots. ### 2026-11-03: Interrupted 8B physical trials and rejection diagnostics - Transferred the compatible 1B draft to Xiaomi 0.2.13. Two 8B attempts accepted 26 draft tokens before sequence 9 expired. Host fallback preserved the final 128-token sequence; finish accounting remained incomplete. Publish individual timing charts and all partial attempts, without calling the unfinished pilots a speedup. - Add bounded host-side rejection reason codes without retaining invalid payloads or granting token credit. 164 host tests pass. No rejected response was recorded in the second failure. The underlying interruption, including the connection-accumulation hypothesis, remains unresolved. ### 2026-11-03: First physical optimized Android timing check - The Xiaomi 0.2.13 worker completed two remote 16-token Qwen requests with identical host-reference token hashes, 9 accepted draft tokens each and no generation fallback. Timing and bounded-work telemetry were received. - The bounded finish request timed out after approximately 2 seconds. Cleanup remains incomplete; the cause is unresolved. This fixed-order check, with a cold first host request, does not qualify speedup. The larger 8B/1B randomized physical pilot requires a new draft transfer. ### 2026-11-03: 8B bounded-pipeline controls and numerical diagnosis - Publish 72 NodeLift controller runs and 63 upstream HTTP control runs with raw tokens, timing records, comparative charts and independent design-review adjudication. Each harness uses its own baseline. - Reproduce a greedy token divergence with identical token prefixes by varying KV history construction and current batch shape. Preserve serial verification as the default; the specific numerical kernel remains unidentified. - Publish Android 0.2.13 build 16 with O3 native kernels and unsigned iOS timing packages. Physical qualification of the optimized Android build remains pending. ### 2026-11-02: Standalone macOS host preview and startup verification - Add an experimental Apple silicon host bundle with embedded Python, the desktop dashboard, native library, local startup checks and a launcher. System macOS frameworks and OpenSSL remain required; models are supplied separately. - Validate dashboard, branding, authenticated inventory and a native eight-token GGUF generation on an M3 Max using temporary loopback state. The first bundle failed on a missing provenance file; the corrected bundle passed. - The host is ad-hoc signed, not notarized, and has not been qualified on other Macs or for physical peer pairing. Windows/Linux downloads are separate CPU-only probes, not full hosts. - All six standalone probe platform builds passed hosted CI startup and inventory checks. The Mac Intel probe is CPU-only and not physically qualified. ### 2026-11-02: Standalone Windows and Linux node-probe previews - Publish self-contained CPU-only Windows and Linux x64/ARM64 probe archives with embedded Python, native benchmark, startup checks, licenses and hashes. Models and GPU drivers remain external. - Startup and local inventory were checked on hosted CI runners; no physical Windows/Linux, GPU or NPU device is qualified. Windows ARM64 uses a baseline Clang build; Linux packages target Ubuntu 22.04. - These are command-line probes rather than the full desktop host or distributed workers. The Mac Apple silicon preview was physically smoke-tested separately. ### 2026-11-02: Standalone node probe 0.1.1 and startup checks - Add a self-contained local probe executable containing Python, standard libraries and llama-bench. Startup checks the bundled benchmark without installing system packages, uploading telemetry or modifying GPU drivers. - Publish a separate experimental macOS Apple silicon ZIP, ad-hoc signed and not notarized; its physical M3 Max test includes inventory and a GGUF smoke test with sentinel Python commands in PATH. - Provide build automation for Windows, Linux and macOS x86/ARM variants; CI execution and physical-device qualification are distinct. A model remains user-supplied and the standalone probe is not a desktop inference worker or the NodeLift host. - No companion binary release or distributed-speedup claim is introduced. ### 2026-11-02: Synchronize showcase and comparison with physical evidence - Refresh the Android and iOS cards with the actual physical draft trials, distinguish privately signed iOS development builds from the unsigned public IPA, and retain negative paired outcomes. - Add the desktop/server probe, first physical Mac kernel controls and an explicit roadmap for physical Windows/Linux qualification and future PC/server workers. - Update only the NodeLift atlas row to the 2 November evidence; retain the 31 October source freeze for all 54 other entries and show both dates. Regenerate the matching CSV, Markdown and integrity manifest. - Correct the showcase packager to include the published node-probe package. No companion binary, distributed speedup or target-memory sharding is added by this documentation update. ### 2026-11-02: Generic node probe 0.1.0 and native model microbenchmarks - Add a reusable local hardware/software probe for macOS, Windows and Linux, including ARM64 detection and explicit unknown states for unsupported accelerators. No machine-specific Lenovo profile is required. - Publish a portable Python 3.10+ package, authenticated desktop inventory/export controls and English, Simplified Chinese and Italian feedback. No model, native runtime or automatic upload is bundled. - Add bounded llama-bench invocation with artifact hashes, retained repetitions, failure records and same-model kernel comparisons. TTFT, energy, connection-specific acceleration and automatic remote-worker enrollment remain unqualified. - Validate physical Apple M3 Max inventory and CPU/requested-offload kernel profiles; Windows/Linux/ARM adapters have fixture tests only. The selected release source passed 151 host tests and seven desktop JavaScript suites in an isolated checkout. See docs/NODE_PROBE.md and research/node-probe-2026-11-02/README.md. - Credit linuxhw/hw-probe and llama.cpp. This is a host/tooling preview, not a new companion binary release; public mobile downloads remain 0.2.11/build 13. ### 2026-11-02: Desktop, workstation and server deployment scope - Extend the two-path acceleration/capacity architecture to Windows/Linux PCs, GPU workstations, macOS nodes and dedicated servers, with the Lenovo P16 as the first proposed non-Mac qualification device. - Define dedicated, shared-workstation and site-local pools; separate coordinator and executor, approved resource reservations, interruption policies and project access boundaries. Benchmark Mac alone, P16 alone and combined execution against the strongest applicable standalone baseline. - Document backend candidates and office-pilot gates. This is a design update: no Windows/Linux worker, Lenovo benchmark, enterprise deployment or new companion build is claimed. ### 2026-11-02: Phone performance diagnosis and Android native optimization correction - Confirm that the distributed Android 0.2.8 native ELF matches an unoptimized Debug build. Enable -O3 for Android Debug native kernels while preserving symbols, assertions and the generic ARM ISA. Cross-compile the corrected AArch64 library and add a compile-command audit with five passing unit tests. - Complete nine new physical iPhone paired comparisons across draft widths 1, 4 and 8. All primary outputs match and no fallback occurs, but median paired duration remains 4.843x to 5.312x the Mac baseline. Width adjustment alone does not demonstrate acceleration. - Publish retained raw sessions, source hashes, a reproducible cost decomposition, research acknowledgements and external-review adjudication. Gemini High completed two reviews; GLM 5.3 Max returned a capacity error and has no completed review. - The asynchronous architecture remains a proposal. No optimized Android phone benchmark, new APK/IPA, measured energy result or phone-assisted speedup is claimed. ### 2026-11-02: Visual phone benchmarks in English, Simplified Chinese and Italian - Publish four reproducible chart families from existing physical Xiaomi and iPhone raw sessions: fixed-delegation duration ratios, added first-token delay, adaptive fallback timing, and proposal acceptance with candidate-request outcomes. - Show individual paired observations, sample sizes, median definitions, the logarithmic ratio axis and an explicitly labelled TTFT detail scale. Retain invalid and missing candidate requests; do not convert absent measurements to zero. - Add a localized expandable gallery to the public evidence section and embed canonical English figures in the research documentation. Exact rows, source hashes and the rendering script are versioned. - Presentation-only update. No new inference runs or companion builds; no confidence, energy, sustained thermal or general hardware-ranking claim is introduced. ### 2026-11-02: Physical iPhone results and private iOS build 0.2.12 (15) - Complete 18 original iPhone-versus-Mac pairs with matching outputs. Fixed phone delegation at a 32-token budget has a median paired 4.657x slowdown, +1040.652 ms total and +109.299 ms TTFT, with 128/246 proposals accepted. The iPhone route has a smaller observed penalty than the earlier Xiaomi route, but the backend and topology differ. - Both adaptive cells fall back in every request, with no accepted phone tokens. A signed total-time improvement in one short cell is not phone acceleration. The fixed-128 cell stops after connection drift; retain one invalid and five missing primary attempts. - Fix iOS handling of obsolete draft-result rejection so expected expired or superseded output is discarded without ending the session. Add a local diagnostic event and regression tests. Sign and install private development build 0.2.12 (15); six separate post-fix pairs complete with matching output and four obsolete proposals discarded without session loss. Fixed-128 remains 4.562x slower with 204/328 proposals accepted. - Publish the frozen plan, raw sessions, failure record and descriptive comparison in research/iphone-tradeoffs-2026-11-02/REPORT.md. Energy, sustained thermals, independent confidence intervals and larger-target benefit remain unqualified. Public companion downloads remain 0.2.11 (13). ### 2026-11-02: iOS draft setup: private development build 0.2.12 (14) - Add the missing iOS download-from-Mac action, progress, cancellation and an explicit development-transfer import action in English, Simplified Chinese and Italian. Setup verifies the paired host hash before enabling proposals; resume the session to load the engine. - Use pinned authenticated TLS, fixed endpoints, bounded streamed transfer and integrity checks before replacement. Add integration coverage for transfer failure, certificate mismatch and cancellation cleanup. - Sign and install private iOS build 0.2.12 (14) over 0.2.8 on the owner's physical iPhone. Copy and import the selected draft. Native inference readiness and performance results remain separate from installation and import. - Public downloads remain 0.2.11 (13). This development installation is not a public, TestFlight or App Store release. See docs/IOS_DRAFT_SETUP.md for setup steps, validation and limitations. ### 2026-11-02: Physical Xiaomi trade-offs: latency loss and adaptive containment - Run 18 new physical Xiaomi-versus-Mac pairs plus 3 local-draft control pairs with matching outputs. Fixed phone drafting at a 32-token budget showed a median 134.949x slowdown in this model/hotspot scope. - Adaptive phone probing returned to the host in all 12 adaptive requests. Median paired overhead was 43.738 ms at a 32-token budget and 87.604 ms at 128 tokens; no phone tokens were used. This contains loss and does not demonstrate acceleration. - Retain policy/width fingerprints, decision blocks and contextual telemetry. Fix the collector clock boundary and show fallback-inclusive completed-service timing separately in Measurements. - A separate post-hoc network diagnostic timed out without samples. Energy, sustained thermals, larger targets and other phones/transports remain unqualified. Full protocol, raw records and trade-off matrix: research/phone-tradeoffs-2026-11-02/REPORT.md. - Validation: 138 host tests, seven desktop JavaScript suites, generated localization checks and 95-page documentation validation passed. All saved reports reproduce from raw sessions; fallback-inclusive timing is visible in the live dashboard. - Host tooling and desktop update only. Native inference engine and companion binaries are unchanged; public Android/iOS packages remain 0.2.11/build 13. ### 2026-11-02: NodeLift Impact v1: scoped time, TTFT, reliability, energy and uncertainty - Add an explicit paired native SSE collector, local impact-analysis API and Measurements import/export panel in English, Simplified Chinese and Italian. Reports retain fallback, missing and mismatched attempts for one model/device/connection scope. - Implement whole-session bootstrap uncertainty, full-coverage instrumented-energy integration and explicit qualification gates. Missing energy remains unavailable; effective offload, residency and independent-session attestation remain incomplete, so automatic routing promotion is disabled. - Methodology reviewed through GLM 5.3 Max and Gemini 4 Argon High. A three-pair Xiaomi hotspot pilot completed with matching outputs, one fallback and no observed speed benefit; one session cannot establish confidence intervals. See docs/IMPACT_BENCHMARK.md and research/impact-2026-11-02/REPORT.md. - Validation: 136 host tests, seven JavaScript suites, generated localization/release checks and 92-page documentation validation passed. The real dashboard and authenticated analysis endpoint rendered the physical pilot with explicit missing energy and no qualified recommendation. - Host, analysis and desktop update only. Native engine and Android/iOS binaries remain unchanged at companion 0.2.11/build 13. ### 2026-11-02: Desktop contextual help and explicit engine capability boundaries - Add contextual explanations beside desktop controls and a browser-local Options preference to disable or restore them. Essential limitations remain visible. - Document native generation, local and phone drafting, prompt lookup, adaptive control and external text routes separately from unimplemented capacity sharding and unqualified physical-phone speedup. See docs/ENGINE_STATUS.md in the repository. - Desktop and documentation update only. Android and iOS packages remain 0.2.11/build 13; no new native inference benchmark is claimed. ### 2026-11-02: Execution backend catalog controls and model-selection guidance - Add catalog loading and matching provider configuration next to Execution backend, with inline errors, empty-catalog guidance and immediate model-selection readiness. Native target/draft controls are disabled for external execution. - Explain that Main model lists discovered/imported GGUF files, not recommended or certified-compatible models, and that draft models are optional token proposers for drafting strategies. - Verify actual OpenRouter catalog selection without cloud generation and an end-to-end text response through an existing LM Studio local server. Focused control, model-library, graph and localization suites pass. Host/UI update only; mobile packages remain 0.2.9/build 11. ### 2026-11-01: Host functional QA with GLM and Gemini, fixes and physical-device limits - Fix cancelled-download error feedback, canonical model provenance and small-model size labels. Localize additional native/validation errors and expose the recorded reason when a run falls back to the computer. - Execute real browser model upload, cancellation, native run/stop/reconfiguration, gateway and localization checks with GLM 5.3 Max and Gemini 4 High providing mediated test plans and reviews. Publish a scoped audit and sanitized receipts; external models did not operate localhost. - A physical Xiaomi completed draft loading and CPU calibration, but two network jobs timed out and no stable remote inference route was qualified. Preserve these failures and the unresolved legacy enrollment duplication. No mobile speedup or pooled RAM claim; companion binaries remain 0.2.8. ### 2026-11-01: Host dashboard: Model library rendering fix and functional audit - Fix Hugging Face result buttons that incorrectly reused device-action state and raised an undefined-message error. Add independent model-operation busy states, explicit search/inspection feedback, upload cancellation and session-aware controls. - Verify real Hugging Face search, inspection, a 19 MB download with SHA-256, cancellation, path import and chunked HTTP upload. Verify a 16-token native Qwen run and the live public OpenRouter catalog. Provider configuration uses an explicit local test fixture; no cloud-generation or installed third-party GUI qualification is claimed. - Add combined dashboard/model-library UI regression checks to CI and publish docs/evidence/model-library-2026-11-01/REPORT.md. Companion binaries remain 0.2.8. ### 2026-11-01: Network recovery guides and unused legacy icon fallback - Update installation/QR support guides for automatic current-network invitations without resetting state. - Replace the unused pre-API-26 legacy drawable in source after 0.2.7 publication. Supported Android versions (API 26+) use the approved 3D adaptive launcher in the published APK. Original 0.2.7 binaries and source tag remain unchanged. ### 2026-11-01: Project acknowledgements and editorial consistency - Add a dedicated showcase Credits chapter and repository acknowledgements with direct links to DistAI and other studied projects, actual dependencies, integration providers and all 54 external comparison entries. - Name contributors and distinguish incorporated software from research references. Preserve original licenses, source snapshots, benchmark records and archived reviewer text. - Remove em dashes from NodeLift-authored prose and UI placeholders, including release-history generation, while preserving verbatim third-party evidence. ### 2026-11-01: English documentation and synchronized release history - Repository guides, historical assessments and research dossiers are maintained in English; application/showcase localization remains English, Simplified Chinese and Italian. - One versioned history generates the repository changelog, portal downloadable notes and rendered release notes. Build and CI checks reject missing build entries or stale generated history. - Wiki, how-to hub, gallery, diagnostics, release-integrity workflow and REA source study document implemented behavior and qualification limits. No mobile binary is rebuilt for this documentation update. Publication dates are UTC. Companion versions and the host preview version are separate. Original published binaries and tags are not rewritten for documentation corrections.