Building the Truth Detector 3000
with Claude

Designing a 5-state machine: S_FINGER → S_CALIBRATE → S_QUESTION → S_VERDICT → S_THUMB_OFF.

Screen layout (240×320 portrait):

• Flashing header: "!! TELL THE TRUTH !!" ↔ "DETECTING LIES..."
• Live scrolling BPM line graph with green glow effect
• Orange BPM bar that shifts red as pulse rises
• Question text with multi-line wrapping
• YES (green) / NO (red) touch buttons

Writing sketches/lie_detector/lie_detector.ino + flash-sketch-003-lie-detector.sh…

First two questions always register as truth to build false confidence. From question 3 onward: pure 50/50 coin flip.

Four main screens, each designed for maximum drama:

Verdict screens fill the entire display in red or green:

Root cause found. The original updateHR() was calling hr.getIR() which reads exactly one sample from the FIFO per call. During TFT graph redraws (~300ms), multiple sensor samples pile up in the FIFO and go unread — so the inter-beat delta spans 2 or more actual beats, halving the reported BPM.

Fix: drain the entire FIFO on every call using the FIFO API:

void updateHR() {
  hr.check();               // fill library buffer from FIFO
  while (hr.available()) {
    long ir = hr.getFIFOIR();
    fingerOn = (ir >= 50000);
    if (fingerOn && checkForBeat(ir)) {
      // compute BPM from inter-beat delta
    }
    hr.nextSample();
  }
}

Also bumped LED brightness from 60 → 0x7F for better signal through a thumb (which is thicker than a fingertip).

Added S_THUMB_OFF state. Any state (calibrating, question, verdict) that loses the IR signal immediately transitions to a full-screen flashing red alert:

THUMB OFF
 SENSOR!

Put your thumb back on
the sensor
(red dot, left side)

Removing your thumb is
very suspicious.

Header alternates between bright red and dark red every 400ms. When IR ≥ 50,000 returns, it resumes the previous state — or restarts calibration cleanly if that's where you left off.

Key fix: resume condition checks fingerOn only (not beatAvg > 20) — BPM resets to 0 when the thumb comes off so a BPM requirement would lock the user out permanently.

Root cause: Arduino's map() doesn't clamp — it extrapolates linearly. If elapsed overshoots CALIB_MS while waiting for a valid BPM reading, the mapped pixel width grows larger than barW and fillRect runs off the right edge of the screen.

// Before — unclamped, could overflow screen
int pw = map(elapsed, 0, CALIB_MS, 0, barW);

// After — safe
int pw = constrain(map(elapsed, 0, CALIB_MS, 0, barW), 0, barW);

Also removed the && calibCount > 0 gate on the calibration completion check — calibration now always exits after 3 seconds, using a 72 BPM default baseline if no beats were detected.

Two separate problems — both fixed:

Problem 1 — The loop ran too long. i2s_channel_write had a 300ms timeout per chunk. With ~172 chunks of audio, worst-case duration was 172 × 300ms = 51 seconds. Switching to a millis()-based exit with a 20ms I2S timeout lets the loop check wall time on every iteration and exit after exactly 2000ms.

Problem 2 — DMA buffer kept playing after the loop. Even after playBuzzer() returned, audio queued in the DMA buffer continued. Fix: write 8 chunks of silence after the loop to immediately overwrite and silence the DMA output.

unsigned long endMs = millis() + 2000;
while (millis() < endMs) {
  // generate + write chunk (20ms timeout)
}
// Flush DMA — silence kills residual audio
memset(buf, 0, sizeof(buf));
for (int f = 0; f < 8; f++) writeI2S(buf, CHUNK);

Added a live signal strength indicator updated every 150ms — shows the raw IR value as a colour-coded bar:

Raw IR value shown in small text at the bottom so you can see exactly what the sensor reads and whether repositioning the thumb helps.

Also exposed gLastIR as a global updated on every FIFO drain, and removed the old beatAvg > 20 transition gate — the app now moves to calibration as soon as the IR signal is strong enough, without waiting for BPM to compute first.

Architecture highlights:

• MAX30102 FIFO drained on every loop tick — no missed beats during screen redraws
• I2S audio on FreeRTOS Core 0 — UI never blocks waiting for sound to finish
• millis()-gated audio with DMA silence flush — exact 2-second cutoff
• 5-state machine handles every edge case: no thumb, calibrating, questioning, verdict, thumb removed mid-interrogation
• Scrolling BPM graph with green glow, 8-entry rolling average, 150ms refresh
• First 2 questions always truth; questions 3+ are 50/50 — theatrical not scientific