diff --git a/lib/src/records.dart b/lib/src/records.dart
index 603d360..34202e6 100644
--- a/lib/src/records.dart
+++ b/lib/src/records.dart
@@ -61,6 +61,9 @@ class R24 {
   /// do NOT apply a correction factor here: a scale derived from one corpus
   /// would be a fabricated calibration, and it would silently move every value
   /// downstream code has already stored.
+  ///
+  /// EMPTY on v25 — that layout has no accelerometer field we can read (see
+  /// [_parseV25]). Empty means absent; it does not mean a still wrist.
   final List<double> accelG;
 
   /// The u8 at inner[51] (frame[55]).
@@ -231,7 +234,8 @@ const int kMinRrMs = 200;
 const int kMaxRrMs = 2500;
 
 /// The historical-record layout versions this package has a field map for.
-/// v25 has its own dedicated layout ([_parseV25]); the rest share the v24 field
+/// v25 has its own dedicated layout ([_parseV25], timestamp only); the rest
+/// share the v24 field
 /// map with a per-version HR offset ([_hrOffsetByVersion]). Used by live.dart's
 /// `decodeRecord` so every version [parseR24] can decode is actually routed to
 /// it (v12 in particular is real firmware and used to fall through to null,
@@ -250,37 +254,35 @@ double _pow10(int n) {
 ByteData _view(Uint8List b) =>
     b.buffer.asByteData(b.offsetInBytes, b.lengthInBytes);
 
-double _gravI16(Uint8List inner, int offset) {
-  final v = _view(inner).getInt16(offset, Endian.little);
-  return _round(v / 16384.0, 4);
-}
-
-/// Decode a v25 (PPG-derived) historical record. Unlike v24, the v25 field map
-/// comes purely from our own device captures — it has no independent
-/// cross-reference, so it is LOWER-CONFIDENCE: we decode timing + gravity and
-/// deliberately leave HR at 0 (v25 carries no honest 1 Hz beat). It is gated on
-/// a gravity-magnitude sanity check to avoid emitting garbage.
+/// Decode a v25 historical record. v25 is NOT a 1 Hz biometric record: it is a
+/// raw ~24 Hz PPG waveform, sent in 13–27 s bursts roughly every 20 minutes.
+/// The only field solved in it is the unix second at inner[7:11]; the waveform
+/// itself is left whole in [R24.rawTail].
+///
+/// So everything else comes back absent, and that is deliberate:
+///
+///   - HR stays 0 — v25 carries no beat. Running a peak/autocorrelation pass
+///     over concatenated bursts manufactures a ~60 bpm peak out of the burst
+///     cadence itself.
+///   - accelG is EMPTY, not (0,0,0). We used to read a "gravity" vector at
+///     inner[69/71/73] as i16/16384 and gate it on |g| ≈ 1. It is not gravity.
+///     Measured over the real v25 corpus: the "z" axis takes THREE distinct
+///     values in the whole sample, the "x" is the upper half of an f32 that
+///     starts two bytes earlier, and per-axis correlation against the v24
+///     gravity vector for the SAME second is 0.16 / 0.22 / 0.06. A vector that
+///     near-constant reads downstream as a perfectly still wrist, so emitting
+///     it fabricated immobility — for hours a day, every day. Do not put it
+///     back, and do not go looking for a different accel offset here.
+///   - the optical fields stay 0 — no red/IR scalar offset is solved, and no
+///     true SpO2% decode exists.
+///
+/// Respiratory rate, HRV, SpO2 and perfusion index were each tested against
+/// this waveform and each failed. Nothing derived belongs in this function.
 R24? _parseV25(Uint8List inner) {
+  // Only inner[3:11] is read below, but a real v25 burst record is ~76 bytes;
+  // anything shorter is a truncated frame, and this floor keeps them out.
   if (inner.length < 75) return null;
   final view = _view(inner);
-  // The documented v25 layout uses frame-absolute offsets. Our decoder
-  // receives the inner record starting at packet type 0x2f, so subtract the
-  // 4-byte transport prefix: unix 11->7 and gravity 73/75/77->69/71/73.
-  //
-  // What is KNOWN for v25:
-  //   - unix is stable at inner[7:11]
-  //   - gravity is stable at inner[69/71/73] as i16/16384
-  // What is NOT solved here:
-  //   - exact red/IR scalar offsets inside the v25 optical region
-  //   - any true SpO2% decode
-  //
-  // So v25 intentionally surfaces motion + time only and leaves the optical
-  // fields zero rather than inventing a bogus decode.
-  final gx = _gravI16(inner, 69);
-  final gy = _gravI16(inner, 71);
-  final gz = _gravI16(inner, 73);
-  final mag = (gx * gx + gy * gy + gz * gz);
-  if (mag < 0.25 || mag > 2.25) return null; // ~0.5g..1.5g
   return R24(
     histVersion: 25,
     tsEpoch: view.getUint32(7, Endian.little),
@@ -291,7 +293,7 @@ R24? _parseV25(Uint8List inner) {
     rrIntervalsMs: const [],
     ppgGreen: 0,
     ppgRedIr: 0,
-    accelG: [gx, gy, gz],
+    accelG: const [],
     skinContact: 0,
     spo2RedRaw: 0,
     spo2IrRaw: 0,
@@ -335,7 +337,8 @@ bool _physiologicallyPlausible(List<double> accelG, int hr) {
 /// packet-type byte.
 ///
 /// Routing:
-///   - v25 → the PPG-derived layout ([_parseV25]).
+///   - v25 → the 24 Hz PPG-waveform layout ([_parseV25]): timestamp only, with
+///     HR 0 and accelG empty.
 ///   - v24 / v12 → our hardware-validated field map, returned as-is.
 ///   - v18 / v9 / v7 → the SAME field map but with the HR byte read at that
 ///     version's offset (only the HR offset is independently confirmed for
diff --git a/test/v25_test.dart b/test/v25_test.dart
index 8a38da5..1bfe588 100644
--- a/test/v25_test.dart
+++ b/test/v25_test.dart
@@ -1,5 +1,3 @@
-import 'dart:math' as math;
-
 import 'package:openstrap_protocol/openstrap_protocol.dart';
 import 'package:test/test.dart';
 
@@ -13,7 +11,7 @@ void main() {
       'aa50000c2f1900026800007fff2a6a38390900729103003608a2fd0104850d4f1bd21aa60f080d850edb116b0f160b7d063f06ab04d5041704a4045f04f003f5ffd7ff7efe73ffa8b2333e9003010000fa54e5e9',
     ];
 
-    test('parseR24 auto-routes v25 and decodes unix + gravity', () {
+    test('parseR24 auto-routes v25 and decodes the unix second only', () {
       final parsed = [
         for (final hex in records) parseR24(hexToBytes(_innerHex(hex)))
       ];
@@ -28,15 +26,25 @@ void main() {
         expect(r.rrIntervalsMs, isEmpty);
         expect(r.spo2RedRaw, 0);
         expect(r.spo2IrRaw, 0);
-        final mag = math.sqrt(
-          r.accelG[0] * r.accelG[0] +
-              r.accelG[1] * r.accelG[1] +
-              r.accelG[2] * r.accelG[2],
-        );
-        expect(mag, inInclusiveRange(0.8, 1.2));
+        // No accel on v25. The old inner[69/71/73] read looked like a ~1 g
+        // vector on exactly these records, which is why it survived so long:
+        // across the real corpus its "z" has three distinct values, its "x" is
+        // half an f32, and it correlates 0.16/0.22/0.06 with the v24 gravity
+        // for the same second. Absent, NOT (0,0,0) — a zero vector is a still
+        // wrist to every motion consumer.
+        expect(r.accelG, isEmpty);
       }
     });
 
+    test('the waveform payload and its timestamp are untouched', () {
+      final r = parseR24(hexToBytes(_innerHex(records[0])))!;
+      // rawTail is inner[13:], i.e. the PPG burst itself. Compare the WHOLE
+      // slice, not `endsWith` — that also passes if the decoder silently drops
+      // leading waveform bytes, which is the failure worth catching.
+      expect(r.rawTail, isNotEmpty);
+      expect(r.rawTail, _innerHex(records[0]).substring(13 * 2));
+    });
+
     test('decodeRecord surfaces v25 timestamps', () {
       final decoded = [for (final hex in records) decodeRecord(_innerHex(hex))];
       expect(decoded.every((r) => r != null), isTrue);