Real hardware → real numbers → full export. Magnet, mic, camera, clock, and net are not simulated. The word line is the agreed readout (hash into the pool you pick). Text will always feel meaningful to a reader — the log neither confirms nor denies contact. Sensor-forward puts physics on the main line and the token underneath; Free phrases + Corpus-stratified slots reduce authored grammar. The app records the run for your protocols and baselines.
ITC / bins: Each live sensor is z-scored in a rolling window, quantized to 8 bins, then fed to FNV so e.g. audio and network land on comparable bin semantics; cold start uses the numeric fallback range in Detection until enough samples exist. Pool = communicative English + names (or your corpus preset); stream clarity and fused sweep shape timing.
Browse corpus & data files (data/ on this host).
NULLFIELD reads real sensors on this device (magnetometer, motion, mic, camera light, network timing, clock jitter, and optional extras). Those readings are turned into numbers, binned or smoothed as shown in the controls, then hashed with a fixed rule (FNV-1a) into an index in the word list you chose: communicative English, a fixed ~30k+ line “Ovilus-5 metaphor” pool (30k frequency-ranked words plus bundled names, US places, and spoken phrases), or the full dictionary. Same sensor history → same index → same word — that is the whole “pick” — so sessions are replayable from the export.
The running phrase is separate: it collects the same tokens and can flush as a line of text depending on pause, spikes, or structured phrase rules. Nothing is sent to a server for word selection; the page is the instrument.
Closest word match (the menu labeled below) only fills the extra transcript line “Sounds like:” — phonetic suggestions (Metaphone + optional Soundex / rhyme-ish tail). They are not a second official pick; they are there when you want “could it have been heard / spelled as something more everyday?” without changing the logged word.
Ovilus-class devices (spirit boxes, branded ITC apps) usually combine a word list, some environmental or motion input, and a scan or sweep that times how often words appear. Exact algorithms and pools are often closed / undocumented, so you cannot reconstruct why a given word appeared from a published log alone.
NULLFIELD is the opposite design goal for the word line: documented hashing, your sensor readings, your chosen pool and phrase rules, and a full JSON export (raw numbers + instrument metadata + optional stats flags). The ITC sweep here is a timer-driven slice that still uses only live seeds (no server-side RNG picking your words). If you want “mystery box” feel you still can — but the bias is toward inspectability and baseline runs you can compare.
For a preset tuned like a spirit box but with a much larger word bank than commercial Ovilus-class gear, choose Ovilus-style + huge bank (full ~370k lexicon) or Ovilus-5 metaphor — ~30k+ lines (30k frequency slice + names + US places + phrases) with fused sweep so mag, mic, camera, net, etc. all feed one ITC seed when those sources are live.
Not a claim about spirits, contact, or science — only about what the software does on the record.
If something external is modulating the channel, you still have to separate that from the way any text readout can arrange into language-shaped output for a human eye — without assuming either way what caused it. The risk is lost detectability — a modulation you care about drowning in a stream that already reads like speech.
Use logs like an RF bench: look for structure that survives persistence, constraint, adaptation, and lower entropy than your control runs.
wordFrequency, sentence timestamps, and runs in the transcript (not single poetic flushes).observation (diversity, repeats, source mix) and sentence priorPhraseHits / observation.crossSessionRegistry for cross-session echoes.session_start metadata across JSON files.
Does not start a session. Loads the named bundle into the controls above — you can still tweak corpus, mic, room, then Start.
Instrumental Transcommunication (ITC) in research culture explicitly includes radios, TV, phones, and software: the idea that contact might ride whatever channel is available has been part of that discourse for a long time. Multi-sensor rigs (environmental + RNG + triggers) are also a known pattern.
Mainstream science mostly treats apparent ITC/EVP hits as noise, expectation, and system artifacts, with weak replication under tight controls. NULLFIELD does not resolve that debate; it logs so you can compare structure (below + export) to your own baselines.
What “evidence” can mean here pairs intuition with things you can write down: repetition, skewed source mix, χ² vs uniform on the used vocabulary, cross-session echoes. None of that settles the big questions by itself — it’s extra grip on the record. The live Session structure panel uses the same math as export (interpretive-style flags for display only; word picks unchanged).
The official word is always the hash pick. This menu only controls the extra transcript line that lists words that are phonetically similar (so you can ask “did we mean something more common?” without changing the instrument output). Session presets set a default here; pick Balanced or read the first section in Plain English preset guide → Closest word match.