github 0x94t3z/fieldatlas v1.3.0
Field Atlas 1.3.0

latest releases: v1.4.10, v1.4.9, v1.4.8...
one month ago

Field Atlas 1.3.0

A public release for offline use without a connection: emergency guidance from published guides,
more file types, open-now place answers and faster answers on low-end phones, checked on a Redmi
13C in airplane mode. It is not a claim that answer quality or bounty acceptance criteria have
been met; the known limitations below still apply. Notes for the previous version:
1.2.1.

Changes

  • Emergency questions are answered from published guides, offline. Ask Research what to do
    (“someone is choking”, “I got bitten by a snake, what should I do?”, “is it safe to drink
    stream water?”) and the model answers from the matching bundled guide only, citing it, with
    the guide's key steps shown word for word above its answer the moment the question is asked.
    41 guides cover first aid, heat and cold, outdoors and survival, and disasters, from CDC,
    NIOSH, NIH, the National Weather Service, the National Park Service, FEMA and the U.S. Fire
    Administration (public domain) and four Wikipedia articles (CC BY-SA 4.0). They ship inside
    the app, so they work before any download; if the model cannot run, the guide is shown as
    published. See Emergency knowledge.
  • Everyday accounts of an emergency find their guide. “A snake bit my friend, what do we
    do?”, “my friend fell in a cold river and is shivering” or a long description of what happened
    now open the matching guide, dictated text without punctuation included. Before, some got the
    model alone, which on the snakebite question advised ice and raising the limb, both against the
    guide. When an injury question mentions a red-flag sign, the guide's when-to-get-care section
    leads. Ordinary questions that share the words (bit rates, a cat that bites) stay in research.
  • Questions that need no model no longer wait for it. Near-me lookups, and emergency guides
    when the model is not ready, start immediately, even while the model is loading or if it
    cannot load on a low-memory phone.
  • Near-me help reaches farther. For hospitals, clinics, pharmacies, police and drinking water,
    the search widens from 15 km to 50 km and then 100 km when little is close.
  • “Open now” is checked against the phone's clock. Asked which places are open now, the app
    reads each place's mapped opening hours, marks it open (and until when) or closed at the
    phone's current time, and lists open places first. A near-me search keeps widening while
    everything close is closed, so a 24-hour pharmacy farther away is found at night. Hours it
    cannot read are not guessed, and for a named city the answer says it assumed the phone is set
    to that city's time. Recorded hours can still be out of date. “Pharmacy open now in Berlin” is a
    lookup like “which pharmacies are open now”, and hours with commas between days
    (“Mo-Fr 08:00-20:00, Sa 09:00-19:00”) are read as separate rules.
  • Museum questions list museums. “Which museums are listed in London?” lists only places
    mapped as museums or galleries in that city, so a zoo, or a museum in Oxford, no longer
    appears; a trip question asking for a food stop and a museum gets both.
  • Places without an address show coordinates. Most Tokyo listings in the map data have no
    address; they now give coordinates that any offline map app finds, and a Japanese address
    without a street name is shown whole on its card.
  • More files can be attached. Word (.docx), PowerPoint (.pptx), OpenDocument (.odt, .odp) and
    EPUB books, HTML as readable text, HEIC/GIF/BMP photos, UTF-16 without a byte-order mark and
    Windows-1252 text. Older Office formats, spreadsheets and archives get a message saying how to
    convert them.
  • PDFs are read from their text layer on every Android version. Before, Android 13–14 ran text
    recognition on every page (an 8-page PDF now reads in about 3 seconds on a Redmi 13C), and
    Android 15+ appended recognized text to the text layer, nearly doubling it. Recognition now runs
    only on pages without real text, at most 30 per file.
  • Long files are read in part instead of refused. Up to 50 MB, 300,000 characters and 1,000
    PDF pages; what was not read is listed in the preview and in the answer.
  • Answers about files use better excerpts. Passages break at paragraphs and sentences, and are
    ranked by the question's distinctive words (part codes such as K-9 and K9 match). Summary
    questions take passages from across the whole file. When an answer used only part of a file,
    the app itself adds a line saying how many passages it used and from which pages.
  • The amount of file text fits the phone. Prompt size is now measured in tokens with an
    estimate calibrated against the real tokenizer (it never undercounted in testing), not in bytes.
    The app learns how fast the phone reads a prompt and sizes file evidence so the wait before the
    first word stays near 40 seconds: fast phones get up to about 1,800 tokens of file text, slow
    phones the best few passages. Instructions that cannot apply to a question are left out.
  • Attached files survive the app being closed. If Android closes Field Atlas while a picker or
    the camera is open, the files are restored and read again with the question.
  • Answers are easier to read and pass on. Citations are small numbered marks attached to the
    words they support. Copy and Share sit at the top of every answer and include the sources;
    sharing works without a connection (Bluetooth, nearby share, SMS). A place answer opens with its
    result in one sentence, with the caveats beneath it, and each place card shows open or closed,
    diet and distance as labels. History can be searched and each saved answer shared.
  • Answers start about 25% sooner on 8-core phones. The model now uses six threads instead of
    four. On a Redmi 13C, the same question's first word came at 71.7 s instead of 94.7 s, and
    writing went from 3.69 to 4.06 tokens a second; eight threads were slower at writing, so two
    cores stay free for the phone. See the measurements.
  • Answer details show where the time went. “Prompt reading” gives the tokens the model read
    and how long that took; “Writing speed” now counts the writing only (it divided by the whole
    wait, showing 1.6 tokens a second for a model writing at 4.4).
  • “Tell me Japan history” starts with its source. Everyday phrasings (“tell me …”, “what do you
    know about …”, “give me an overview of …”) and titles such as “History of Japan” for “Japan
    history” now open the answer with the encyclopedia's own words, cited, while the model's
    explanation is still being written.
  • Research no longer looks stuck while the model reads and writes. The reading bar moves between
    the engine's block reports at the speed measured so far, never ahead of it, with the time left
    ("about 20 s left"); the live draft follows the newest text instead of freezing once an answer
    grew past its first lines.
  • New 3D app icon and opening animation. The launcher icon is the 3D compass coin, and on
    launch the coin turns once and settles face-on while the app starts (about 0.75 s, no added
    delay; cold start on the Redmi 13C stayed at 0.7–0.75 s). Themed icons keep the flat outline.
  • More library questions find their overview passage. A question such as “Explain autophagy
    and distinguish …” now uses the saved passage that defines the topic, as “Explain autophagy and
    how …” already did, instead of being answered by the model alone.
  • Fixes. Stopping voice input while the recogniser was still decoding could crash the app (it freed the recogniser under the recording thread; found with Android's malloc debug, the kind of bug a hardened allocator such as GrapheneOS's turns into a crash). A very large file could crash the app (out of memory); a camera photo could not be
    retried after a storage error; every photo was named “Camera photo.jpg”; a full disk said “Could
    not read this file”; the text preview stuttered on long files; nested list steps showed flat; the live draft could
    show a list marker as text (“…- Do not drive”).

Verification

  • 574 JVM unit tests pass (10 opt-in desktop tests skipped), including real-tokenizer counts,
    generated DOCX/PPTX/ODT/EPUB files, a zip bomb, emergency-question routing against all bundled
    guides, questions that must not be routed to a guide, and opening-hours rules (split shifts,
    past midnight, holidays, open-ended closing times, unreadable hours, comma-separated days).
    139 Python tests pass (57 scripts, 12 packtool, 6 benchmark, 64 builders).
  • Desktop, the 24-question release suite (same model, seed and collections as 1.2.0): 15 pass,
    8 partial, 1 fail; judged about as useful as an online frontier answer on 16½ of 23 (72%),
    up from 65% before the travel and retrieval fixes. The reviewer is the author's assistant and
    the questions were used to find the problems, so this is not a held-out score. See
    Evaluation.
  • Desktop, Qwen3.5 2B, 7 attachment questions × 5 seeds: 33/35 passed the fixture checks (1.2.1:
    32/35). Keyword checks, not an accuracy score.
  • Desktop, Qwen3.5 2B, 6 emergency questions × 2 seeds: 10 of 12 model answers were clean; two
    contained one garbled sentence each, beside the verbatim source excerpt.
  • Redmi 13C, Android 13, airplane mode, installed over 1.2.1 with data kept:
    • snakebite question: the CDC/NIOSH source excerpt on screen in 4 s, cited model answer in 1 min 4 s;
    • 8-page text PDF ready in about 3 s; answer cited to page 6 in 53 s once the phone's speed was
      learned (2 min 31 s before the evidence budget was sized to the phone);
    • DOCX 39 s, EPUB (cited to chapter 3) 44 s, PPTX (cited to slides 2 and 3) 41 s, photo of a
      printed sign read with OCR 36 s, all correct; nearest-hospital lookup about 1 s;
    • after the travel fixes, installed over the earlier 1.3.0 build: “Which pharmacies near me are
      open now?” in 2 s, widening to 50 km in a rural area and listing two 24/7 pharmacies and one
      open until 21:30 first, then the three nearer pharmacies with no recorded hours; “Which vegan
      restaurants in Berlin are open now?” marked two of six open at the phone's 10:39 and stated
      the Berlin-time assumption. Each card showed its open or closed status;
    • final build, installed over the earlier 1.3.0 build: “A snake bit my friend. What do we do?”
      showed the CDC/NIOSH excerpt within seconds and the cited answer in 58 s (52 s on the build
      before the last two fixes);
      “Pharmacy open now in Berlin” was answered as a lookup in 43 s, with no model (before the fix:
      3 min 1 s and uncited); CDC's one-page “Make Water Safe During an Emergency” PDF answered “How
      long do I boil the water?” in 39 s, correctly, citing page 1.
  • Device tests on the same phone: all 68 instrumented tests pass, including the 9 attachment-reader
    tests (a 3-page text PDF reads in 852 ms). Eleven of them had been failing
    on 1.2.1 as well because they expected older wording and layouts (for example licence details
    now in a pack's About sheet, and citations that open a preview before the full source), the
    microphone test never granted the microphone, and the offline test still forbade the internet
    permission that downloads use; they were updated to the current interface and to the policy in
    scripts/verify_offline.sh. One real defect found on the way (the original-file viewer could
    keep its loading spinner after a file was read) was fixed by publishing the loaded file on the
    main thread.
  • Memory-safety check for GrapheneOS-class hardware, on the same phone: the device suite and a
    full model answer were run with Android's malloc debug (guard bytes, freed-memory tracking,
    pointer checks) on every native library: llama.cpp, Tesseract, Vosk and SQLite. It found the
    voice crash above and nothing else; after the fix the suite passed with no reports.
  • Release lint and the offline packaging audit pass. APK size 108.4 MB (1.2.1: 100.7 MB); most of
    the increase is the PDF text library, its fonts and the BouncyCastle crypto library it uses for encrypted PDFs.
  • Public APK: built from the tagged source with the project release key (certificate SHA-256
    131127512c99a625acd0dd4baf20e1c7cd240b0fd573449197070000e3d6b666); signature, offline
    packaging, ZIP alignment and all 22 ARM64 ELF alignment checks pass. The phone checks above
    used a locally signed build of the same source; the public-key APK was not installed on a phone.

Known limitations

  • The guides are information, not training, and many are written for the United States. Severe
    bleeding has no bundled tourniquet guidance; the choking and drowning guides are Wikipedia text.
    These need review by a clinician. The model's part of an emergency answer can misstate a step;
    the verbatim source excerpt above it keeps the guide's own wording. See Emergency knowledge.
  • On a slow phone a long file is answered from a few passages; the added coverage line says so.
  • A named-city question about essentials (“pharmacy open now in Berlin”) is answered from a
    keyword search and can list only one or two places, some in the surrounding area (the Berlin
    answer above listed one pharmacy in Potsdam). Near-me questions search by distance and are not
    affected.
  • The 1.2.1 limitations on citation repair and model reasoning still apply.

Installation

Download fieldatlas.apk and check fieldatlas.apk.sha256. Requires Android 13+ and an arm64
device. It installs over 1.2.1 with data kept. Model and knowledge packs are installed
separately; queries run locally after setup. Downloads and catalog refresh require a connection.
The emergency guides ship inside the app.

App version: 1.3.0 (17). GitHub channel: release (v1.3.0).

APK SHA-256: 28036d47f8a62e9e40b720b948d5206f4a53d58e4a62cbb4d0df05b3da42b267.

Release assets

  • App release v1.3.0: fieldatlas.apk, fieldatlas.apk.sha256.
  • No collection changed in this release; each stays on its existing knowledge release.

Don't miss a new fieldatlas release

NewReleases is sending notifications on new releases.