Q3 2026 allocation: 2 PRO units remaining Order window closes in — Reserve Now →
Ground Truth Labeling Machine · Precision Biosignal Capture

Your labels
are the
bottleneck.

GTLM — Ground Truth Labeling Machine — is a controlled-environment biometric capture booth that automatically labels physiological and behavioral data at the resolution research has always needed, but never had.

Words are a lossy codec. Synchronized biosignals are the original file.

32+ Sensor channels
μs Sync precision
4× Label hierarchy
0 Manual annotations
Reserve Your Unit → See Sample Output

Research institutions currently on waitlist

H: 2200mm W: 1200mm RGB-D DEPTH CAMERA 4-MIC SPATIAL ARRAY EEG HEADSET ARRAY ECG + HRV CHEST SENSOR GSR / EDA WRIST BAND FACIAL ACTION UNIT CAM EYE-TRACKING OCULOMETER RESPIRATION BELT EMG ELECTRODES POSTURAL LOAD SENSORS FRONT VIEW · SENSOR ARRAY
BUILD / A la carte · every sensor, dimension, material · Configure →
EEG · ECG · HRV · GSR · EMG · PROSODY · OCULOMETRY · PROXEMICS · RESPIRATION · FACIAL ACTION UNITS · POSTURAL ANALYSIS · SPEECH SENTIMENT · MICRO-EXPRESSIONS · BLINK RATE · PUPIL DILATION · VOCAL TREMOR · SKIN CONDUCTANCE · EEG · ECG · HRV · GSR · EMG · PROSODY · OCULOMETRY · PROXEMICS · RESPIRATION · FACIAL ACTION UNITS · POSTURAL ANALYSIS · SPEECH SENTIMENT · MICRO-EXPRESSIONS · BLINK RATE · PUPIL DILATION · VOCAL TREMOR · SKIN CONDUCTANCE · EEG · ECG · HRV · GSR · EMG · PROSODY · OCULOMETRY · PROXEMICS · RESPIRATION · FACIAL ACTION UNITS · POSTURAL ANALYSIS · SPEECH SENTIMENT · MICRO-EXPRESSIONS · BLINK RATE · PUPIL DILATION · VOCAL TREMOR · SKIN CONDUCTANCE ·

Data quality

What changes when the label is the measurement

Most affective datasets use self-report or rater annotation. That label is a compressed, delayed, language-encoded reconstruction of a physiological event. It is not the thing. It is a description of a memory of the thing.

Standard workflow With GTLM
Label source Human rater or self-report, post-session Concurrent physiological measurement
Temporal resolution 1-30 seconds per annotation Sub-millisecond, all channels
Label type Categorical (e.g. arousal 1-7) Continuous, 32+ channels simultaneously
Inter-rater agreement κ ≈ 0.4-0.6 N/A — deterministic
Annotation time 3-10× session duration Zero — real time
Label entropy ~3 bits per annotation ~256 kbps continuous stream
What the model trains on Description of a memory of a state The state itself, measured directly

Sample output · single frame · t = 12.847 s into session

{
  "session_id":  "VRM-2024-0847",
  "t_unix_us":   1718432847392,

  "eeg":         { "channels": 32, "sample_rate_hz": 1000, "frame": [ /* 32 × float32 */ ] },
  "ecg":         { "bpm": 74.3, "hrv_rmssd_ms": 42.1, "r_peak": true },
  "gsr":         { "conductance_us": 4.82, "tonic": 4.61, "phasic": 0.21 },
  "respiration": { "rate_bpm": 15.2, "phase": "exhale", "depth_norm": 0.61 },
  "eye":         { "fixation": true, "pupil_diam_mm": 4.1, "gaze_x": 0.48, "gaze_y": 0.51 },
  "emg":         { "zygomaticus": 0.12, "corrugator": 0.04, "frontalis": 0.07 },
  "depth_cam":   { "head_pose_yaw_deg": -3.2, "au_06": 0.08, "au_12": 0.14 },

  "labels": {
    "raw":        "sympathetic_activation · eeg_alpha_suppression · resp_hold",
    "derived":    "arousal: +0.74 · valence: -0.21 · engagement: 0.68",
    "inferred":   "cognitive_load: elevated · stress: moderate",
    "behavioral": "attention: sustained · facial_affect: neutral_tense"
  }
}

Training on this means optimizing against a signal that was physically present, at the moment it was present — not against a rating someone gave 30 seconds later.

In plain language

What this is and why it exists

When researchers want to understand how someone feels, they usually ask them. But asking changes the answer. People guess, round up, misremember. The body does not have that problem.

The booth

GTLM is an enclosed space, roughly phone-booth-sized. A participant sits inside for a session. Ten sensors record what the body is doing: heart rate, brain electrical activity, sweat response, eye movement, breathing, and muscle tension. None of it requires the participant to do anything except be there.

The labeling problem

Raw physiological data is not useful on its own. Every recording needs to be annotated: what was happening at each moment, what the stimulus was, what the context was. In standard research, someone does this by hand. It takes weeks and introduces error. GTLM does it automatically, in real time, as the session runs.

What you receive

At the end of a session: one structured file. Every signal synchronized to a single clock. Each time point labeled with four layers of context. No manual cleanup, no raw wrangling. The data is ready to use.

Hardware

The Booth, Up Close

Subject seated inside GTLM booth wearing EEG headset and eye-tracking glasses during a biometric capture session
GTLM PRO, active capture session. EEG array, eye-tracking oculometer, and depth camera visible on subject and booth walls.
GTLM booth exterior, matte anodized aluminum enclosure with smoked glass door, standing in a research laboratory
Exterior: matte anodized aluminum chassis with Faraday-shielded smoked glass door. Footprint: 1.2 m × 1.2 m.

Technical Specifications

Sensor Catalogue

Every GTLM shares the same booth enclosure, Faraday shielding, LSL middleware, and auto-labeling engine. The sensor suite is fully configurable — select any combination in the Build section below.

SENSOR CATALOGUE

All sensors available for GTLM configuration — consumer and medical grade. Select any combination in the Build section below.

Audio / Visual

Device Modality Rate Resolution Signal
Logitech Brio 4K Facial capture 60 fps 4K UHD RGB video
ReSpeaker 4-Mic Array Spatial audio 48 kHz 24-bit PCM audio
Intel RealSense D455 Depth / proxemics 90 fps 848×480 RGB-D

Neurophysiological

Device Modality Rate Resolution Signal
Emotiv EPOC X EEG 128 Hz 14 channels μV
Empatica E4 GSR / EDA 4 Hz 0.01 μS μS
Polar H10 ECG / HRV 130 Hz R-R interval ms
Garmin HRM-Pro SpO₂ / PPG 1 Hz 1% %
Hexoskin belt Respiration 128 Hz chest excursion a.u.

Movement / Posture

Device Modality Rate Resolution Signal
Noraxon IMU Posture / IMU 200 Hz 6-DOF deg/s, g
Zebris FDM Floor pressure 120 Hz 4 sensors/cm² N/cm²

Eye Tracking

Device Modality Rate Resolution Signal
Tobii Eye Tracker 5 Oculometry 90 Hz 0.4° accuracy gaze XY

Neurophysiological — Medical Grade

Device Modality Rate Resolution Signal
BrainProducts LiveAmp 32 EEG 500 Hz 32 channels μV
Biopac ECG100C 12-lead ECG 1000 Hz clinical mV
Biopac EDA100C Medical GSR 2000 Hz 0.001 μS μS
Masimo EMMA Capnography 100 Hz EtCO₂ mmHg mmHg
cEEGrid array Around-ear EEG 500 Hz 10 channels μV

Movement / Posture — Medical Grade

Device Modality Rate Resolution Signal
Xsens DOT (×8) Full-body mocap 120 Hz <1° RMS deg, m
Delsys Trigno Wireless EMG (8–16ch) 2000 Hz 16-bit mV

Voice / Affect

Device Modality Rate Resolution Signal
Shure MXA310 Clinical mic array 96 kHz 24-bit PCM
On-device prosody engine Prosody F0/jitter real-time 1 Hz Hz, %

Process

From Subject to Dataset

Six steps from booth entry to structured multimodal output. Every step is standardized, reproducible, and auditable.

  1. Enter the Booth

    Subject steps in. Acoustically dampened (–42 dB), Faraday-shielded, climate controlled at 21°C ±0.5°C. No external noise, no signal bleed, no electromagnetic interference. The enclosure is the experiment's first independent variable: it holds constant every environmental factor that cannot otherwise be controlled.

  2. Sensor Initialization

    All devices boot and synchronize via GTLM's proprietary LSL-compatible middleware. Baseline calibration runs for 90 seconds of resting-state capture. Impedance checks, electrode contact verification, and sampling rate confirmation are performed automatically. Any sensor below threshold triggers an alert before the session begins.

  3. Session Begins

    Subject interacts with a presented stimulus: video, audio, structured task, interview protocol, or physical product. Session duration is configurable between 5 and 120 minutes. The experimenter monitors signal quality via the external dashboard in real time without entering the booth or disturbing the electromagnetic environment.

  4. Live Capture

    Every sensor streams simultaneously. The LSL sync layer timestamps all streams to microsecond precision, enabling true multimodal temporal alignment. Jitter between streams is below 1 ms. Data is written to NVMe storage in real time with redundant backup. No network dependency during capture.

  5. Auto-Labeling Engine

    Proprietary software analyzes each timepoint across all modalities simultaneously. The output follows a four-level label hierarchy: (a) raw signal state: measured physiological values at each sample; (b) derived physiological state: computed metrics such as HRV, SCR amplitude, EEG band power; (c) inferred affective/cognitive state: Valence-Arousal-Dominance coordinates and categorical probability distributions; (d) behavioral annotation: action units, postural events, gaze fixations, vocal events. All labels are time-indexed and timestamped to the same microsecond reference frame as the raw signals.

  6. Export

    Output is a structured dataset: time-indexed multimodal signals paired with label columns at every temporal resolution. Available export formats: CSV · Parquet · HDF5 · JSON-LD. Schema documentation is included with every export. Datasets are ready for ingestion into research pipelines without further preprocessing.

SENSORS ×10 EEG · ECG · GSR · EMG OCULO · RESP · DEPTH AUDIO · IMU · PRESSURE LSL SYNC LAYER Microsecond timestamp alignment · <1ms jitter AUTO-LABELING ENGINE Raw → Physiological → Affective → Behavioral 4-level label hierarchy STRUCTURED OUTPUT CSV · Parquet · HDF5 JSON-LD · Pipeline-ready Time-indexed · Labeled

Signal Output Preview

Live Simulation

30-second multimodal capture window. Five synchronized signal streams rendered in real time.

RAW SIGNAL STATE

GSR spike +2.3 μS  |  EEG α suppression (40%→22%)  |  AU6+AU12: active (intensity 0.9)

DERIVED PHYSIOLOGICAL STATE

Sympathetic arousal onset  |  SNS:PNS ratio 1.4:1  |  Frontal α asymmetry: L > R  |  HRV RMSSD ↓ 31ms

AFFECTIVE INFERENCE

Valence: +0.72  |  Arousal: +0.68. Probable state: DELIGHT / SURPRISE (p=0.81)

BEHAVIORAL ANNOTATION

Duchenne smile confirmed (AU6+AU12+AU25)  |  Forward postural lean: +12°  |  Blink rate suppressed (2.1/min)

Download Sample Dataset (.csv)

Methodology

Why Biosignals Are Ground Truth

Why Words Are a Lossy Codec

Language compresses rich internal states into discrete tokens that lose temporal resolution, contextual co-occurrence, and physiological ground truth. Annotation happens hours after the event, by someone other than the subject, using vocabulary that does not exist for most internal states. The mean inter-rater agreement for standard emotion annotation corpora is rarely above 0.6 Cohen's κ, a ceiling imposed by language, not by the complexity of the underlying cognition.

"A word is a lossy approximation. A synchronized biosignal array is the original file."

The Labeling Bottleneck

The quality ceiling of any model is the label quality of its training data. This constraint is particularly acute for affective computing, RLHF, and human preference modeling. Natarajan et al. (2013)¹ demonstrated that label noise induces systematic bias that degrades model accuracy non-linearly. AffectNet (Mollahosseini et al., 2017)², one of the largest affective datasets, relies on crowdsourced keyword annotations, resulting in inter-rater reliability of approximately 0.55 for compound expressions. Biosignals bypass this bottleneck by providing continuous, high-resolution, physiologically grounded labels that do not require a human annotator to agree on a word.

"The most important variable in your model is not your architecture. It is the reliability of your labels."

Controlled Environment as Scientific Instrument

Signal quality degrades catastrophically in ambulatory settings. EEG artifact rejection rates exceed 40% in mobile recordings versus under 8% in controlled environments (Gramann et al., 2011)³. Electrodermal activity in unshielded environments introduces 50/60 Hz interference that masks genuine skin conductance response events. The GTLM booth is not a limitation on ecological validity; it is a methodological requirement for scientific validity. Faraday shielding (attenuation: –60 dB at 100 kHz), acoustic dampening (–42 dB SIL), and standardized ambient lighting (200 lux ±5%) are not product features. They are the experimental controls that make the data usable.

  1. Natarajan, N., Dhillon, I. S., Ravikumar, P. K., & Tewari, A. (2013). Learning with noisy labels. Advances in Neural Information Processing Systems (NeurIPS), 26.
  2. Mollahosseini, A., Hasani, B., & Mahoor, M. H. (2017). AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild. IEEE Transactions on Affective Computing, 10(1), 18–31.
  3. Gramann, K., Gwin, J. T., Ferris, D. P., Oie, K., Jung, T.-P., Lin, C.-T., & Makeig, S. (2011). Cognition in action: imaging brain/body dynamics in mobile humans. Reviews in the Neurosciences, 22(6), 593–608.

Since you opened this page, approximately

0.0 hours of behavioral data

have been discarded by researchers using inferior annotation methods.
Every session without GTLM is signal you cannot recover.

Stop losing data. →

Configuration & Pricing

Acquire a GTLM System

Choose a pre-configured package or build your own from scratch. All systems are built to order. Q3 allocation is almost gone.

Q3 order window closes in
96d
23h
47m
12s

Units unconfirmed by Aug 31 roll to Q4 allocation. Lead time: 8–10 weeks from order confirmation.

Recent deployments
  • Mar 2026MIT Media Lab, Cambridge — PRO
  • Feb 2026Max Planck Institute, Frankfurt — PRO
  • Jan 2026King's College London — LITE
  • Dec 2025Stanford HCI Group, Palo Alto — PRO
  • Nov 2025ETH Zürich, Comp. Psychiatry Lab — LITE
LITE
€67,900

Consumer-Grade Sensor Suite

  • Complete 11-sensor hardware kit (consumer-grade)
  • LSL sync middleware license (1 year)
  • Auto-labeling engine license (1 year)
  • Cloud dashboard access
  • Remote setup support (2 sessions)
  • 2 calibration sessions (remote)
  • Standard lead time: 6–10 weeks
Reserve a GTLM Lite →
RECOMMENDED
PRO
€129,900

Medical-Grade Sensor Suite

2 of 4 units available this quarter
  • Everything in LITE, plus:
  • Medical-grade sensor suite upgrade
  • 32-channel EEG (BrainProducts LiveAmp) + 12-lead ECG + medical GSR (Biopac)
  • Delsys Trigno EMG array (8–16 channels)
  • Clinical-grade Shure MXA310 mic array
  • On-site installation + calibration (1 day, EU/US)
  • Dedicated data engineering support (10 hours)
  • Priority firmware and labeling engine updates
  • Ethics documentation package (IRB/REC ready)
Reserve a GTLM Pro →
or configure from scratch

Build Your Own GTLM

Select every variable — sensors, dimensions, materials, colour, and services. Prices shown are indicative approximations; final quote is confirmed on submission.

Base Unit Included in all configurations

  • Faraday-shielded enclosure structure (–60 dB at 100 kHz)
  • Acoustic dampening (–42 dB SIL)
  • Climate control (21°C ±0.5°C)
  • Rack-mount processing unit + LSL sync hub
  • Internal display (stimulus delivery)
  • Wiring harness + cable management system
Base enclosure €28,500

Sensors — Audio / Visual

Sensors — Neurophysiological

Sensors — Movement & Posture

Sensors — Eye Tracking

Enclosure — Dimensions

Enclosure — Exterior Material

Enclosure — Exterior Colour

Software & Services

Frequently Asked Questions

Yes. Institutional lease arrangements are available for 6- and 12-month terms. Contact us for current availability and pricing at your location. Lease terms include your configured sensor suite, middleware license, and remote technical support.

All data collected using a GTLM system remains exclusively owned by the purchasing institution. GTLM Systems collects no usage telemetry. The auto-labeling engine runs entirely on-device, with no data transmission to external servers at any point during or after capture.

The standard booth requires a floor area of 1.2 m × 1.2 m, a ceiling height of ≥2.5 m, and a 230 V / 16 A outlet. Compact and Extended configurations differ — see dimension options above. The unit is delivered flat-packed and assembled on-site; minimum doorway clearance required is 80 cm.

Yes. The Ethics Documentation Package (available as an add-on) includes a full sensor disclosure with radiation emissions data, contact materials inventory, and participant information templates formatted for IRB, REC, and equivalent ethics bodies in EU, US, and UK jurisdictions.

Typically 6–10 weeks from order confirmation, depending on sensor availability and delivery location. On-site installation is scheduled separately and typically occurs within 2 weeks of hardware delivery. Remote configuration is completed via video call within 5 business days of delivery.

About Us

Biometric research, since 2013

Founded in London, we are a multidisciplinary team of computer scientists, neuroscience experts, MDs, psychologists, researchers, and hardware and software engineers. Since incorporation we have worked alongside universities, research institutions, and corporate customers to build instruments that measure what self-report cannot.

What we build

Over a decade of research has produced a suite of commercially available products — sold under our own brand and under the brands of our customers. Every product we ship is grounded in the same principle: physiological measurement is more reliable than retrospective annotation.

Who we work with

Our team has designed, developed, and deployed products for research institutions, government agencies, and Fortune 100 companies. Clients include Pedigree, Hasbro, NIH, Vodafone, and the United Nations — projects spanning consumer insight, clinical research, human factors, and applied neuroscience.

Selected clients & partners

Pedigree Hasbro NIH Vodafone United Nations

For investors

We are open to conversations with venture capital firms and strategic investors interested in the future of affective computing, AI training data infrastructure, and precision biometric capture. Get in touch for demos, case studies, and a detailed briefing on the commercial pipeline.

Contact Us →

pedja@idguardian.co

Lab Access

Experience the Booth

Lab visits are available at our partner facilities in four cities. Each session is 90 minutes and includes a full data capture demonstration, live signal visualization, and a post-session signal debrief with a GTLM researcher.

London

51.5321°N, 0.1233°W

King's Cross Neurotech Hub, London

Available Dates

  • Tue 9 Sep 202610:00–11:30
  • Thu 18 Sep 202614:00–15:30
  • Wed 7 Oct 202610:00–11:30
  • Fri 23 Oct 202613:00–14:30
Book This Location

New York

40.7291°N, 73.9965°W

NYU Center for Biomedical Imaging, New York

Available Dates

  • Mon 15 Sep 202609:00–10:30
  • Fri 26 Sep 202614:00–15:30
  • Tue 13 Oct 202610:00–11:30
  • Thu 29 Oct 202609:00–10:30
Book This Location

Zürich

47.3769°N, 8.5417°E

ETH Zürich Neurotechnology Center

Available Dates

  • Wed 10 Sep 202610:00–11:30
  • Mon 22 Sep 202614:00–15:30
  • Thu 15 Oct 202610:00–11:30
  • Tue 3 Nov 202613:00–14:30
Book This Location

Berlin

52.5247°N, 13.3786°E

Charité Medical Innovation Lab, Berlin

Available Dates

  • Fri 12 Sep 202611:00–12:30
  • Wed 24 Sep 202609:00–10:30
  • Mon 19 Oct 202614:00–15:30
  • Thu 5 Nov 202610:00–11:30
Book This Location