diff --git a/crypto_kem/firesaber/META.yml b/crypto_kem/firesaber/META.yml
index 6cd4342b..3d1a0a41 100644
--- a/crypto_kem/firesaber/META.yml
+++ b/crypto_kem/firesaber/META.yml
@@ -14,9 +14,9 @@ principal-submitters:
   - Frederik Vercauteren
 implementations:
     - name: clean
-      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/33e5ed62/saber
+      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/90d072e4/saber
     - name: avx2
-      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/33e5ed62/saber
+      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/90d072e4/saber
       supported_platforms:
           - architecture: x86_64
             operating_systems:
diff --git a/crypto_kem/firesaber/avx2/pack_unpack.c b/crypto_kem/firesaber/avx2/pack_unpack.c
index 41b9747a..82f5a3f0 100644
--- a/crypto_kem/firesaber/avx2/pack_unpack.c
+++ b/crypto_kem/firesaber/avx2/pack_unpack.c
@@ -8,23 +8,24 @@ void PQCLEAN_FIRESABER_AVX2_POLT2BS(uint8_t bytes[SABER_SCALEBYTES_KEM], const p
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & 0x3f) | ((in[1] & 0x03) << 6);
-        out[1] = ((in[1] >> 2) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[2] = ((in[2] >> 4) & 0x03) | ((in[3] & 0x3f) << 2);
+        out[0] = (in[0] & 0x3f) | (in[1] << 6);
+        out[1] = ((in[1] >> 2) & 0x0f) | (in[2] << 4);
+        out[2] = ((in[2] >> 4) & 0x03) | (in[3] << 2);
         in += 4;
         out += 3;
     }
 }
 
 void PQCLEAN_FIRESABER_AVX2_BS2POLT(poly *data, const uint8_t bytes[SABER_SCALEBYTES_KEM]) {
+    /* This function does not reduce its output mod T */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = in[0] & 0x3f;
-        out[1] = ((in[0] >> 6) & 0x03) | ((in[1] & 0x0f) << 2);
-        out[2] = ((in[1] & 0xff) >> 4) | ((in[2] & 0x03) << 4);
-        out[3] = ((in[2] & 0xff) >> 2);
+        out[0] = in[0];
+        out[1] = (in[0] >> 6) | (in[1] << 2);
+        out[2] = (in[1] >> 4) | (in[2] << 4);
+        out[3] = (in[2] >> 2);
         in += 3;
         out += 4;
     }
@@ -35,37 +36,38 @@ static void POLq2BS(uint8_t bytes[SABER_POLYBYTES], const poly *data) {
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x1f) | ((in[1] & 0x07) << 5);
-        out[2] = ((in[1] >> 3) & 0xff);
-        out[3] = ((in[1] >> 11) & 0x03) | ((in[2] & 0x3f) << 2);
-        out[4] = ((in[2] >> 6) & 0x7f) | ((in[3] & 0x01) << 7);
-        out[5] = ((in[3] >> 1) & 0xff);
-        out[6] = ((in[3] >> 9) & 0x0f) | ((in[4] & 0x0f) << 4);
-        out[7] = ((in[4] >> 4) & 0xff);
-        out[8] = ((in[4] >> 12) & 0x01) | ((in[5] & 0x7f) << 1);
-        out[9] = ((in[5] >> 7) & 0x3f) | ((in[6] & 0x03) << 6);
-        out[10] = ((in[6] >> 2) & 0xff);
-        out[11] = ((in[6] >> 10) & 0x07) | ((in[7] & 0x1f) << 3);
-        out[12] = ((in[7] >> 5) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x1f) | (in[1] << 5);
+        out[2] = in[1] >> 3;
+        out[3] = ((in[1] >> 11) & 0x03) | (in[2] << 2);
+        out[4] = ((in[2] >> 6) & 0x7f) | (in[3] << 7);
+        out[5] = in[3] >> 1;
+        out[6] = ((in[3] >> 9) & 0x0f) | (in[4] << 4);
+        out[7] = in[4] >> 4;
+        out[8] = ((in[4] >> 12) & 0x01) | (in[5] << 1);
+        out[9] = ((in[5] >> 7) & 0x3f) | (in[6] << 6);
+        out[10] = in[6] >> 2;
+        out[11] = ((in[6] >> 10) & 0x07) | (in[7] << 3);
+        out[12] = in[7] >> 5;
         in += 8;
         out += 13;
     }
 }
 
 static void BS2POLq(poly *data, const uint8_t bytes[SABER_POLYBYTES]) {
+    /* This function does not reduce its output mod Q */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x1f) << 8);
-        out[1] = (in[1] >> 5 & (0x07)) | ((in[2] & 0xff) << 3) | ((in[3] & 0x03) << 11);
-        out[2] = (in[3] >> 2 & (0x3f)) | ((in[4] & 0x7f) << 6);
-        out[3] = (in[4] >> 7 & (0x01)) | ((in[5] & 0xff) << 1) | ((in[6] & 0x0f) << 9);
-        out[4] = (in[6] >> 4 & (0x0f)) | ((in[7] & 0xff) << 4) | ((in[8] & 0x01) << 12);
-        out[5] = (in[8] >> 1 & (0x7f)) | ((in[9] & 0x3f) << 7);
-        out[6] = (in[9] >> 6 & (0x03)) | ((in[10] & 0xff) << 2) | ((in[11] & 0x07) << 10);
-        out[7] = (in[11] >> 3 & (0x1f)) | ((in[12] & 0xff) << 5);
+        out[0] = (in[0]) | (in[1] << 8);
+        out[1] = (in[1] >> 5) | (in[2] << 3) | (in[3] << 11);
+        out[2] = (in[3] >> 2) | (in[4] << 6);
+        out[3] = (in[4] >> 7) | (in[5] << 1) | (in[6] << 9);
+        out[4] = (in[6] >> 4) | (in[7] << 4) | (in[8] << 12);
+        out[5] = (in[8] >> 1) | (in[9] << 7);
+        out[6] = (in[9] >> 6) | (in[10] << 2) | (in[11] << 10);
+        out[7] = (in[11] >> 3) | (in[12] << 5);
         in += 13;
         out += 8;
     }
@@ -76,11 +78,11 @@ static void POLp2BS(uint8_t bytes[SABER_POLYCOMPRESSEDBYTES], const poly *data)
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x03) | ((in[1] & 0x3f) << 2);
-        out[2] = ((in[1] >> 6) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[3] = ((in[2] >> 4) & 0x3f) | ((in[3] & 0x03) << 6);
-        out[4] = ((in[3] >> 2) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x03) | (in[1] << 2);
+        out[2] = ((in[1] >> 6) & 0x0f) | (in[2] << 4);
+        out[3] = ((in[2] >> 4) & 0x3f) | (in[3] << 6);
+        out[4] = in[3] >> 2;
         in += 4;
         out += 5;
     }
@@ -91,10 +93,10 @@ static void BS2POLp(poly *data, const uint8_t bytes[SABER_POLYCOMPRESSEDBYTES])
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x03) << 8);
-        out[1] = ((in[1] >> 2) & (0x3f)) | ((in[2] & 0x0f) << 6);
-        out[2] = ((in[2] >> 4) & (0x0f)) | ((in[3] & 0x3f) << 4);
-        out[3] = ((in[3] >> 6) & (0x03)) | ((in[4] & 0xff) << 2);
+        out[0] = in[0] | (in[1] << 8);
+        out[1] = (in[1] >> 2) | (in[2] << 6);
+        out[2] = (in[2] >> 4) | (in[3] << 4);
+        out[3] = (in[3] >> 6) | (in[4] << 2);
         in += 5;
         out += 4;
     }
diff --git a/crypto_kem/firesaber/clean/pack_unpack.c b/crypto_kem/firesaber/clean/pack_unpack.c
index ec2f1263..91ffd723 100644
--- a/crypto_kem/firesaber/clean/pack_unpack.c
+++ b/crypto_kem/firesaber/clean/pack_unpack.c
@@ -8,23 +8,24 @@ void PQCLEAN_FIRESABER_CLEAN_POLT2BS(uint8_t bytes[SABER_SCALEBYTES_KEM], const
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & 0x3f) | ((in[1] & 0x03) << 6);
-        out[1] = ((in[1] >> 2) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[2] = ((in[2] >> 4) & 0x03) | ((in[3] & 0x3f) << 2);
+        out[0] = (in[0] & 0x3f) | (in[1] << 6);
+        out[1] = ((in[1] >> 2) & 0x0f) | (in[2] << 4);
+        out[2] = ((in[2] >> 4) & 0x03) | (in[3] << 2);
         in += 4;
         out += 3;
     }
 }
 
 void PQCLEAN_FIRESABER_CLEAN_BS2POLT(poly *data, const uint8_t bytes[SABER_SCALEBYTES_KEM]) {
+    /* This function does not reduce its output mod T */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = in[0] & 0x3f;
-        out[1] = ((in[0] >> 6) & 0x03) | ((in[1] & 0x0f) << 2);
-        out[2] = ((in[1] & 0xff) >> 4) | ((in[2] & 0x03) << 4);
-        out[3] = ((in[2] & 0xff) >> 2);
+        out[0] = in[0];
+        out[1] = (in[0] >> 6) | (in[1] << 2);
+        out[2] = (in[1] >> 4) | (in[2] << 4);
+        out[3] = (in[2] >> 2);
         in += 3;
         out += 4;
     }
@@ -35,37 +36,38 @@ static void POLq2BS(uint8_t bytes[SABER_POLYBYTES], const poly *data) {
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x1f) | ((in[1] & 0x07) << 5);
-        out[2] = ((in[1] >> 3) & 0xff);
-        out[3] = ((in[1] >> 11) & 0x03) | ((in[2] & 0x3f) << 2);
-        out[4] = ((in[2] >> 6) & 0x7f) | ((in[3] & 0x01) << 7);
-        out[5] = ((in[3] >> 1) & 0xff);
-        out[6] = ((in[3] >> 9) & 0x0f) | ((in[4] & 0x0f) << 4);
-        out[7] = ((in[4] >> 4) & 0xff);
-        out[8] = ((in[4] >> 12) & 0x01) | ((in[5] & 0x7f) << 1);
-        out[9] = ((in[5] >> 7) & 0x3f) | ((in[6] & 0x03) << 6);
-        out[10] = ((in[6] >> 2) & 0xff);
-        out[11] = ((in[6] >> 10) & 0x07) | ((in[7] & 0x1f) << 3);
-        out[12] = ((in[7] >> 5) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x1f) | (in[1] << 5);
+        out[2] = in[1] >> 3;
+        out[3] = ((in[1] >> 11) & 0x03) | (in[2] << 2);
+        out[4] = ((in[2] >> 6) & 0x7f) | (in[3] << 7);
+        out[5] = in[3] >> 1;
+        out[6] = ((in[3] >> 9) & 0x0f) | (in[4] << 4);
+        out[7] = in[4] >> 4;
+        out[8] = ((in[4] >> 12) & 0x01) | (in[5] << 1);
+        out[9] = ((in[5] >> 7) & 0x3f) | (in[6] << 6);
+        out[10] = in[6] >> 2;
+        out[11] = ((in[6] >> 10) & 0x07) | (in[7] << 3);
+        out[12] = in[7] >> 5;
         in += 8;
         out += 13;
     }
 }
 
 static void BS2POLq(poly *data, const uint8_t bytes[SABER_POLYBYTES]) {
+    /* This function does not reduce its output mod Q */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x1f) << 8);
-        out[1] = (in[1] >> 5 & (0x07)) | ((in[2] & 0xff) << 3) | ((in[3] & 0x03) << 11);
-        out[2] = (in[3] >> 2 & (0x3f)) | ((in[4] & 0x7f) << 6);
-        out[3] = (in[4] >> 7 & (0x01)) | ((in[5] & 0xff) << 1) | ((in[6] & 0x0f) << 9);
-        out[4] = (in[6] >> 4 & (0x0f)) | ((in[7] & 0xff) << 4) | ((in[8] & 0x01) << 12);
-        out[5] = (in[8] >> 1 & (0x7f)) | ((in[9] & 0x3f) << 7);
-        out[6] = (in[9] >> 6 & (0x03)) | ((in[10] & 0xff) << 2) | ((in[11] & 0x07) << 10);
-        out[7] = (in[11] >> 3 & (0x1f)) | ((in[12] & 0xff) << 5);
+        out[0] = (in[0]) | (in[1] << 8);
+        out[1] = (in[1] >> 5) | (in[2] << 3) | (in[3] << 11);
+        out[2] = (in[3] >> 2) | (in[4] << 6);
+        out[3] = (in[4] >> 7) | (in[5] << 1) | (in[6] << 9);
+        out[4] = (in[6] >> 4) | (in[7] << 4) | (in[8] << 12);
+        out[5] = (in[8] >> 1) | (in[9] << 7);
+        out[6] = (in[9] >> 6) | (in[10] << 2) | (in[11] << 10);
+        out[7] = (in[11] >> 3) | (in[12] << 5);
         in += 13;
         out += 8;
     }
@@ -76,11 +78,11 @@ static void POLp2BS(uint8_t bytes[SABER_POLYCOMPRESSEDBYTES], const poly *data)
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x03) | ((in[1] & 0x3f) << 2);
-        out[2] = ((in[1] >> 6) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[3] = ((in[2] >> 4) & 0x3f) | ((in[3] & 0x03) << 6);
-        out[4] = ((in[3] >> 2) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x03) | (in[1] << 2);
+        out[2] = ((in[1] >> 6) & 0x0f) | (in[2] << 4);
+        out[3] = ((in[2] >> 4) & 0x3f) | (in[3] << 6);
+        out[4] = in[3] >> 2;
         in += 4;
         out += 5;
     }
@@ -91,10 +93,10 @@ static void BS2POLp(poly *data, const uint8_t bytes[SABER_POLYCOMPRESSEDBYTES])
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x03) << 8);
-        out[1] = ((in[1] >> 2) & (0x3f)) | ((in[2] & 0x0f) << 6);
-        out[2] = ((in[2] >> 4) & (0x0f)) | ((in[3] & 0x3f) << 4);
-        out[3] = ((in[3] >> 6) & (0x03)) | ((in[4] & 0xff) << 2);
+        out[0] = in[0] | (in[1] << 8);
+        out[1] = (in[1] >> 2) | (in[2] << 6);
+        out[2] = (in[2] >> 4) | (in[3] << 4);
+        out[3] = (in[3] >> 6) | (in[4] << 2);
         in += 5;
         out += 4;
     }
diff --git a/crypto_kem/lightsaber/META.yml b/crypto_kem/lightsaber/META.yml
index d3d7bf13..263db2e0 100644
--- a/crypto_kem/lightsaber/META.yml
+++ b/crypto_kem/lightsaber/META.yml
@@ -14,9 +14,9 @@ principal-submitters:
   - Frederik Vercauteren
 implementations:
     - name: clean
-      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/33e5ed62/saber
+      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/90d072e4/saber
     - name: avx2
-      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/33e5ed62/saber
+      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/90d072e4/saber
       supported_platforms:
           - architecture: x86_64
             operating_systems:
diff --git a/crypto_kem/lightsaber/avx2/pack_unpack.c b/crypto_kem/lightsaber/avx2/pack_unpack.c
index a9f866ae..a154d24c 100644
--- a/crypto_kem/lightsaber/avx2/pack_unpack.c
+++ b/crypto_kem/lightsaber/avx2/pack_unpack.c
@@ -8,27 +8,28 @@ void PQCLEAN_LIGHTSABER_AVX2_POLT2BS(uint8_t bytes[SABER_SCALEBYTES_KEM], const
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & 0x7) | ((in[1] & 0x7) << 3) | ((in[2] & 0x3) << 6);
-        out[1] = ((in[2] >> 2) & 0x01) | ((in[3] & 0x7) << 1) | ((in[4] & 0x7) << 4) | (((in[5]) & 0x01) << 7);
-        out[2] = ((in[5] >> 1) & 0x03) | ((in[6] & 0x7) << 2) | ((in[7] & 0x7) << 5);
+        out[0] = (in[0] & 0x7) | ((in[1] & 0x7) << 3) | (in[2] << 6);
+        out[1] = ((in[2] >> 2) & 0x01) | ((in[3] & 0x7) << 1) | ((in[4] & 0x7) << 4) | (in[5] << 7);
+        out[2] = ((in[5] >> 1) & 0x03) | ((in[6] & 0x7) << 2) | (in[7] << 5);
         in += 8;
         out += 3;
     }
 }
 
 void PQCLEAN_LIGHTSABER_AVX2_BS2POLT(poly *data, const uint8_t bytes[SABER_SCALEBYTES_KEM]) {
+    /* This function does not reduce its output mod T */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0]) & 0x07;
-        out[1] = ((in[0]) >> 3) & 0x07;
-        out[2] = (((in[0]) >> 6) & 0x03) | (((in[1]) & 0x01) << 2);
-        out[3] = ((in[1]) >> 1) & 0x07;
-        out[4] = ((in[1]) >> 4) & 0x07;
-        out[5] = (((in[1]) >> 7) & 0x01) | (((in[2]) & 0x03) << 1);
-        out[6] = ((in[2] >> 2) & 0x07);
-        out[7] = ((in[2] >> 5) & 0x07);
+        out[0] = in[0];
+        out[1] = in[0] >> 3;
+        out[2] = (in[0] >> 6) | (in[1] << 2);
+        out[3] = in[1] >> 1;
+        out[4] = in[1] >> 4;
+        out[5] = (in[1] >> 7) | (in[2] << 1);
+        out[6] = in[2] >> 2;
+        out[7] = in[2] >> 5;
         in += 3;
         out += 8;
     }
@@ -39,37 +40,38 @@ static void POLq2BS(uint8_t bytes[SABER_POLYBYTES], const poly *data) {
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x1f) | ((in[1] & 0x07) << 5);
-        out[2] = ((in[1] >> 3) & 0xff);
-        out[3] = ((in[1] >> 11) & 0x03) | ((in[2] & 0x3f) << 2);
-        out[4] = ((in[2] >> 6) & 0x7f) | ((in[3] & 0x01) << 7);
-        out[5] = ((in[3] >> 1) & 0xff);
-        out[6] = ((in[3] >> 9) & 0x0f) | ((in[4] & 0x0f) << 4);
-        out[7] = ((in[4] >> 4) & 0xff);
-        out[8] = ((in[4] >> 12) & 0x01) | ((in[5] & 0x7f) << 1);
-        out[9] = ((in[5] >> 7) & 0x3f) | ((in[6] & 0x03) << 6);
-        out[10] = ((in[6] >> 2) & 0xff);
-        out[11] = ((in[6] >> 10) & 0x07) | ((in[7] & 0x1f) << 3);
-        out[12] = ((in[7] >> 5) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x1f) | (in[1] << 5);
+        out[2] = in[1] >> 3;
+        out[3] = ((in[1] >> 11) & 0x03) | (in[2] << 2);
+        out[4] = ((in[2] >> 6) & 0x7f) | (in[3] << 7);
+        out[5] = in[3] >> 1;
+        out[6] = ((in[3] >> 9) & 0x0f) | (in[4] << 4);
+        out[7] = in[4] >> 4;
+        out[8] = ((in[4] >> 12) & 0x01) | (in[5] << 1);
+        out[9] = ((in[5] >> 7) & 0x3f) | (in[6] << 6);
+        out[10] = in[6] >> 2;
+        out[11] = ((in[6] >> 10) & 0x07) | (in[7] << 3);
+        out[12] = in[7] >> 5;
         in += 8;
         out += 13;
     }
 }
 
 static void BS2POLq(poly *data, const uint8_t bytes[SABER_POLYBYTES]) {
+    /* This function does not reduce its output mod Q */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x1f) << 8);
-        out[1] = (in[1] >> 5 & (0x07)) | ((in[2] & 0xff) << 3) | ((in[3] & 0x03) << 11);
-        out[2] = (in[3] >> 2 & (0x3f)) | ((in[4] & 0x7f) << 6);
-        out[3] = (in[4] >> 7 & (0x01)) | ((in[5] & 0xff) << 1) | ((in[6] & 0x0f) << 9);
-        out[4] = (in[6] >> 4 & (0x0f)) | ((in[7] & 0xff) << 4) | ((in[8] & 0x01) << 12);
-        out[5] = (in[8] >> 1 & (0x7f)) | ((in[9] & 0x3f) << 7);
-        out[6] = (in[9] >> 6 & (0x03)) | ((in[10] & 0xff) << 2) | ((in[11] & 0x07) << 10);
-        out[7] = (in[11] >> 3 & (0x1f)) | ((in[12] & 0xff) << 5);
+        out[0] = (in[0]) | (in[1] << 8);
+        out[1] = (in[1] >> 5) | (in[2] << 3) | (in[3] << 11);
+        out[2] = (in[3] >> 2) | (in[4] << 6);
+        out[3] = (in[4] >> 7) | (in[5] << 1) | (in[6] << 9);
+        out[4] = (in[6] >> 4) | (in[7] << 4) | (in[8] << 12);
+        out[5] = (in[8] >> 1) | (in[9] << 7);
+        out[6] = (in[9] >> 6) | (in[10] << 2) | (in[11] << 10);
+        out[7] = (in[11] >> 3) | (in[12] << 5);
         in += 13;
         out += 8;
     }
@@ -80,11 +82,11 @@ static void POLp2BS(uint8_t bytes[SABER_POLYCOMPRESSEDBYTES], const poly *data)
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x03) | ((in[1] & 0x3f) << 2);
-        out[2] = ((in[1] >> 6) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[3] = ((in[2] >> 4) & 0x3f) | ((in[3] & 0x03) << 6);
-        out[4] = ((in[3] >> 2) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x03) | (in[1] << 2);
+        out[2] = ((in[1] >> 6) & 0x0f) | (in[2] << 4);
+        out[3] = ((in[2] >> 4) & 0x3f) | (in[3] << 6);
+        out[4] = in[3] >> 2;
         in += 4;
         out += 5;
     }
@@ -95,10 +97,10 @@ static void BS2POLp(poly *data, const uint8_t bytes[SABER_POLYCOMPRESSEDBYTES])
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x03) << 8);
-        out[1] = ((in[1] >> 2) & (0x3f)) | ((in[2] & 0x0f) << 6);
-        out[2] = ((in[2] >> 4) & (0x0f)) | ((in[3] & 0x3f) << 4);
-        out[3] = ((in[3] >> 6) & (0x03)) | ((in[4] & 0xff) << 2);
+        out[0] = in[0] | (in[1] << 8);
+        out[1] = (in[1] >> 2) | (in[2] << 6);
+        out[2] = (in[2] >> 4) | (in[3] << 4);
+        out[3] = (in[3] >> 6) | (in[4] << 2);
         in += 5;
         out += 4;
     }
diff --git a/crypto_kem/lightsaber/clean/pack_unpack.c b/crypto_kem/lightsaber/clean/pack_unpack.c
index f64c4143..c1c8666c 100644
--- a/crypto_kem/lightsaber/clean/pack_unpack.c
+++ b/crypto_kem/lightsaber/clean/pack_unpack.c
@@ -8,27 +8,28 @@ void PQCLEAN_LIGHTSABER_CLEAN_POLT2BS(uint8_t bytes[SABER_SCALEBYTES_KEM], const
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & 0x7) | ((in[1] & 0x7) << 3) | ((in[2] & 0x3) << 6);
-        out[1] = ((in[2] >> 2) & 0x01) | ((in[3] & 0x7) << 1) | ((in[4] & 0x7) << 4) | (((in[5]) & 0x01) << 7);
-        out[2] = ((in[5] >> 1) & 0x03) | ((in[6] & 0x7) << 2) | ((in[7] & 0x7) << 5);
+        out[0] = (in[0] & 0x7) | ((in[1] & 0x7) << 3) | (in[2] << 6);
+        out[1] = ((in[2] >> 2) & 0x01) | ((in[3] & 0x7) << 1) | ((in[4] & 0x7) << 4) | (in[5] << 7);
+        out[2] = ((in[5] >> 1) & 0x03) | ((in[6] & 0x7) << 2) | (in[7] << 5);
         in += 8;
         out += 3;
     }
 }
 
 void PQCLEAN_LIGHTSABER_CLEAN_BS2POLT(poly *data, const uint8_t bytes[SABER_SCALEBYTES_KEM]) {
+    /* This function does not reduce its output mod T */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0]) & 0x07;
-        out[1] = ((in[0]) >> 3) & 0x07;
-        out[2] = (((in[0]) >> 6) & 0x03) | (((in[1]) & 0x01) << 2);
-        out[3] = ((in[1]) >> 1) & 0x07;
-        out[4] = ((in[1]) >> 4) & 0x07;
-        out[5] = (((in[1]) >> 7) & 0x01) | (((in[2]) & 0x03) << 1);
-        out[6] = ((in[2] >> 2) & 0x07);
-        out[7] = ((in[2] >> 5) & 0x07);
+        out[0] = in[0];
+        out[1] = in[0] >> 3;
+        out[2] = (in[0] >> 6) | (in[1] << 2);
+        out[3] = in[1] >> 1;
+        out[4] = in[1] >> 4;
+        out[5] = (in[1] >> 7) | (in[2] << 1);
+        out[6] = in[2] >> 2;
+        out[7] = in[2] >> 5;
         in += 3;
         out += 8;
     }
@@ -39,37 +40,38 @@ static void POLq2BS(uint8_t bytes[SABER_POLYBYTES], const poly *data) {
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x1f) | ((in[1] & 0x07) << 5);
-        out[2] = ((in[1] >> 3) & 0xff);
-        out[3] = ((in[1] >> 11) & 0x03) | ((in[2] & 0x3f) << 2);
-        out[4] = ((in[2] >> 6) & 0x7f) | ((in[3] & 0x01) << 7);
-        out[5] = ((in[3] >> 1) & 0xff);
-        out[6] = ((in[3] >> 9) & 0x0f) | ((in[4] & 0x0f) << 4);
-        out[7] = ((in[4] >> 4) & 0xff);
-        out[8] = ((in[4] >> 12) & 0x01) | ((in[5] & 0x7f) << 1);
-        out[9] = ((in[5] >> 7) & 0x3f) | ((in[6] & 0x03) << 6);
-        out[10] = ((in[6] >> 2) & 0xff);
-        out[11] = ((in[6] >> 10) & 0x07) | ((in[7] & 0x1f) << 3);
-        out[12] = ((in[7] >> 5) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x1f) | (in[1] << 5);
+        out[2] = in[1] >> 3;
+        out[3] = ((in[1] >> 11) & 0x03) | (in[2] << 2);
+        out[4] = ((in[2] >> 6) & 0x7f) | (in[3] << 7);
+        out[5] = in[3] >> 1;
+        out[6] = ((in[3] >> 9) & 0x0f) | (in[4] << 4);
+        out[7] = in[4] >> 4;
+        out[8] = ((in[4] >> 12) & 0x01) | (in[5] << 1);
+        out[9] = ((in[5] >> 7) & 0x3f) | (in[6] << 6);
+        out[10] = in[6] >> 2;
+        out[11] = ((in[6] >> 10) & 0x07) | (in[7] << 3);
+        out[12] = in[7] >> 5;
         in += 8;
         out += 13;
     }
 }
 
 static void BS2POLq(poly *data, const uint8_t bytes[SABER_POLYBYTES]) {
+    /* This function does not reduce its output mod Q */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x1f) << 8);
-        out[1] = (in[1] >> 5 & (0x07)) | ((in[2] & 0xff) << 3) | ((in[3] & 0x03) << 11);
-        out[2] = (in[3] >> 2 & (0x3f)) | ((in[4] & 0x7f) << 6);
-        out[3] = (in[4] >> 7 & (0x01)) | ((in[5] & 0xff) << 1) | ((in[6] & 0x0f) << 9);
-        out[4] = (in[6] >> 4 & (0x0f)) | ((in[7] & 0xff) << 4) | ((in[8] & 0x01) << 12);
-        out[5] = (in[8] >> 1 & (0x7f)) | ((in[9] & 0x3f) << 7);
-        out[6] = (in[9] >> 6 & (0x03)) | ((in[10] & 0xff) << 2) | ((in[11] & 0x07) << 10);
-        out[7] = (in[11] >> 3 & (0x1f)) | ((in[12] & 0xff) << 5);
+        out[0] = (in[0]) | (in[1] << 8);
+        out[1] = (in[1] >> 5) | (in[2] << 3) | (in[3] << 11);
+        out[2] = (in[3] >> 2) | (in[4] << 6);
+        out[3] = (in[4] >> 7) | (in[5] << 1) | (in[6] << 9);
+        out[4] = (in[6] >> 4) | (in[7] << 4) | (in[8] << 12);
+        out[5] = (in[8] >> 1) | (in[9] << 7);
+        out[6] = (in[9] >> 6) | (in[10] << 2) | (in[11] << 10);
+        out[7] = (in[11] >> 3) | (in[12] << 5);
         in += 13;
         out += 8;
     }
@@ -80,11 +82,11 @@ static void POLp2BS(uint8_t bytes[SABER_POLYCOMPRESSEDBYTES], const poly *data)
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x03) | ((in[1] & 0x3f) << 2);
-        out[2] = ((in[1] >> 6) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[3] = ((in[2] >> 4) & 0x3f) | ((in[3] & 0x03) << 6);
-        out[4] = ((in[3] >> 2) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x03) | (in[1] << 2);
+        out[2] = ((in[1] >> 6) & 0x0f) | (in[2] << 4);
+        out[3] = ((in[2] >> 4) & 0x3f) | (in[3] << 6);
+        out[4] = in[3] >> 2;
         in += 4;
         out += 5;
     }
@@ -95,10 +97,10 @@ static void BS2POLp(poly *data, const uint8_t bytes[SABER_POLYCOMPRESSEDBYTES])
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x03) << 8);
-        out[1] = ((in[1] >> 2) & (0x3f)) | ((in[2] & 0x0f) << 6);
-        out[2] = ((in[2] >> 4) & (0x0f)) | ((in[3] & 0x3f) << 4);
-        out[3] = ((in[3] >> 6) & (0x03)) | ((in[4] & 0xff) << 2);
+        out[0] = in[0] | (in[1] << 8);
+        out[1] = (in[1] >> 2) | (in[2] << 6);
+        out[2] = (in[2] >> 4) | (in[3] << 4);
+        out[3] = (in[3] >> 6) | (in[4] << 2);
         in += 5;
         out += 4;
     }
diff --git a/crypto_kem/saber/META.yml b/crypto_kem/saber/META.yml
index 32b4b964..319f4ebf 100644
--- a/crypto_kem/saber/META.yml
+++ b/crypto_kem/saber/META.yml
@@ -14,9 +14,9 @@ principal-submitters:
   - Frederik Vercauteren
 implementations:
     - name: clean
-      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/33e5ed62/saber
+      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/90d072e4/saber
     - name: avx2
-      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/33e5ed62/saber
+      version: https://github.com/KULeuven-COSIC/SABER/tree/509cc5ec3a7e12a751ccdd2ef5bd6e54e00bd350 via https://github.com/jschanck/package-pqclean/tree/90d072e4/saber
       supported_platforms:
           - architecture: x86_64
             operating_systems:
diff --git a/crypto_kem/saber/avx2/pack_unpack.c b/crypto_kem/saber/avx2/pack_unpack.c
index 9bb46acb..a055b7e5 100644
--- a/crypto_kem/saber/avx2/pack_unpack.c
+++ b/crypto_kem/saber/avx2/pack_unpack.c
@@ -8,19 +8,20 @@ void PQCLEAN_SABER_AVX2_POLT2BS(uint8_t bytes[SABER_SCALEBYTES_KEM], const poly
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 2; j++) {
-        out[0] = (in[0] & 0x0f) | ((in[1] & 0x0f) << 4);
+        out[0] = (in[0] & 0x0f) | (in[1] << 4);
         in += 2;
         out += 1;
     }
 }
 
 void PQCLEAN_SABER_AVX2_BS2POLT(poly *data, const uint8_t bytes[SABER_SCALEBYTES_KEM]) {
+    /* This function does not reduce its output mod T */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 2; j++) {
-        out[0] = in[0] & 0x0f;
-        out[1] = (in[0] >> 4) & 0x0f;
+        out[0] = in[0];
+        out[1] = in[0] >> 4;
         in += 1;
         out += 2;
     }
@@ -31,37 +32,38 @@ static void POLq2BS(uint8_t bytes[SABER_POLYBYTES], const poly *data) {
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x1f) | ((in[1] & 0x07) << 5);
-        out[2] = ((in[1] >> 3) & 0xff);
-        out[3] = ((in[1] >> 11) & 0x03) | ((in[2] & 0x3f) << 2);
-        out[4] = ((in[2] >> 6) & 0x7f) | ((in[3] & 0x01) << 7);
-        out[5] = ((in[3] >> 1) & 0xff);
-        out[6] = ((in[3] >> 9) & 0x0f) | ((in[4] & 0x0f) << 4);
-        out[7] = ((in[4] >> 4) & 0xff);
-        out[8] = ((in[4] >> 12) & 0x01) | ((in[5] & 0x7f) << 1);
-        out[9] = ((in[5] >> 7) & 0x3f) | ((in[6] & 0x03) << 6);
-        out[10] = ((in[6] >> 2) & 0xff);
-        out[11] = ((in[6] >> 10) & 0x07) | ((in[7] & 0x1f) << 3);
-        out[12] = ((in[7] >> 5) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x1f) | (in[1] << 5);
+        out[2] = in[1] >> 3;
+        out[3] = ((in[1] >> 11) & 0x03) | (in[2] << 2);
+        out[4] = ((in[2] >> 6) & 0x7f) | (in[3] << 7);
+        out[5] = in[3] >> 1;
+        out[6] = ((in[3] >> 9) & 0x0f) | (in[4] << 4);
+        out[7] = in[4] >> 4;
+        out[8] = ((in[4] >> 12) & 0x01) | (in[5] << 1);
+        out[9] = ((in[5] >> 7) & 0x3f) | (in[6] << 6);
+        out[10] = in[6] >> 2;
+        out[11] = ((in[6] >> 10) & 0x07) | (in[7] << 3);
+        out[12] = in[7] >> 5;
         in += 8;
         out += 13;
     }
 }
 
 static void BS2POLq(poly *data, const uint8_t bytes[SABER_POLYBYTES]) {
+    /* This function does not reduce its output mod Q */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x1f) << 8);
-        out[1] = (in[1] >> 5 & (0x07)) | ((in[2] & 0xff) << 3) | ((in[3] & 0x03) << 11);
-        out[2] = (in[3] >> 2 & (0x3f)) | ((in[4] & 0x7f) << 6);
-        out[3] = (in[4] >> 7 & (0x01)) | ((in[5] & 0xff) << 1) | ((in[6] & 0x0f) << 9);
-        out[4] = (in[6] >> 4 & (0x0f)) | ((in[7] & 0xff) << 4) | ((in[8] & 0x01) << 12);
-        out[5] = (in[8] >> 1 & (0x7f)) | ((in[9] & 0x3f) << 7);
-        out[6] = (in[9] >> 6 & (0x03)) | ((in[10] & 0xff) << 2) | ((in[11] & 0x07) << 10);
-        out[7] = (in[11] >> 3 & (0x1f)) | ((in[12] & 0xff) << 5);
+        out[0] = (in[0]) | (in[1] << 8);
+        out[1] = (in[1] >> 5) | (in[2] << 3) | (in[3] << 11);
+        out[2] = (in[3] >> 2) | (in[4] << 6);
+        out[3] = (in[4] >> 7) | (in[5] << 1) | (in[6] << 9);
+        out[4] = (in[6] >> 4) | (in[7] << 4) | (in[8] << 12);
+        out[5] = (in[8] >> 1) | (in[9] << 7);
+        out[6] = (in[9] >> 6) | (in[10] << 2) | (in[11] << 10);
+        out[7] = (in[11] >> 3) | (in[12] << 5);
         in += 13;
         out += 8;
     }
@@ -72,11 +74,11 @@ static void POLp2BS(uint8_t bytes[SABER_POLYCOMPRESSEDBYTES], const poly *data)
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x03) | ((in[1] & 0x3f) << 2);
-        out[2] = ((in[1] >> 6) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[3] = ((in[2] >> 4) & 0x3f) | ((in[3] & 0x03) << 6);
-        out[4] = ((in[3] >> 2) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x03) | (in[1] << 2);
+        out[2] = ((in[1] >> 6) & 0x0f) | (in[2] << 4);
+        out[3] = ((in[2] >> 4) & 0x3f) | (in[3] << 6);
+        out[4] = in[3] >> 2;
         in += 4;
         out += 5;
     }
@@ -87,10 +89,10 @@ static void BS2POLp(poly *data, const uint8_t bytes[SABER_POLYCOMPRESSEDBYTES])
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x03) << 8);
-        out[1] = ((in[1] >> 2) & (0x3f)) | ((in[2] & 0x0f) << 6);
-        out[2] = ((in[2] >> 4) & (0x0f)) | ((in[3] & 0x3f) << 4);
-        out[3] = ((in[3] >> 6) & (0x03)) | ((in[4] & 0xff) << 2);
+        out[0] = in[0] | (in[1] << 8);
+        out[1] = (in[1] >> 2) | (in[2] << 6);
+        out[2] = (in[2] >> 4) | (in[3] << 4);
+        out[3] = (in[3] >> 6) | (in[4] << 2);
         in += 5;
         out += 4;
     }
diff --git a/crypto_kem/saber/clean/pack_unpack.c b/crypto_kem/saber/clean/pack_unpack.c
index 2542f0b0..1b5bed81 100644
--- a/crypto_kem/saber/clean/pack_unpack.c
+++ b/crypto_kem/saber/clean/pack_unpack.c
@@ -8,19 +8,20 @@ void PQCLEAN_SABER_CLEAN_POLT2BS(uint8_t bytes[SABER_SCALEBYTES_KEM], const poly
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 2; j++) {
-        out[0] = (in[0] & 0x0f) | ((in[1] & 0x0f) << 4);
+        out[0] = (in[0] & 0x0f) | (in[1] << 4);
         in += 2;
         out += 1;
     }
 }
 
 void PQCLEAN_SABER_CLEAN_BS2POLT(poly *data, const uint8_t bytes[SABER_SCALEBYTES_KEM]) {
+    /* This function does not reduce its output mod T */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 2; j++) {
-        out[0] = in[0] & 0x0f;
-        out[1] = (in[0] >> 4) & 0x0f;
+        out[0] = in[0];
+        out[1] = in[0] >> 4;
         in += 1;
         out += 2;
     }
@@ -31,37 +32,38 @@ static void POLq2BS(uint8_t bytes[SABER_POLYBYTES], const poly *data) {
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x1f) | ((in[1] & 0x07) << 5);
-        out[2] = ((in[1] >> 3) & 0xff);
-        out[3] = ((in[1] >> 11) & 0x03) | ((in[2] & 0x3f) << 2);
-        out[4] = ((in[2] >> 6) & 0x7f) | ((in[3] & 0x01) << 7);
-        out[5] = ((in[3] >> 1) & 0xff);
-        out[6] = ((in[3] >> 9) & 0x0f) | ((in[4] & 0x0f) << 4);
-        out[7] = ((in[4] >> 4) & 0xff);
-        out[8] = ((in[4] >> 12) & 0x01) | ((in[5] & 0x7f) << 1);
-        out[9] = ((in[5] >> 7) & 0x3f) | ((in[6] & 0x03) << 6);
-        out[10] = ((in[6] >> 2) & 0xff);
-        out[11] = ((in[6] >> 10) & 0x07) | ((in[7] & 0x1f) << 3);
-        out[12] = ((in[7] >> 5) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x1f) | (in[1] << 5);
+        out[2] = in[1] >> 3;
+        out[3] = ((in[1] >> 11) & 0x03) | (in[2] << 2);
+        out[4] = ((in[2] >> 6) & 0x7f) | (in[3] << 7);
+        out[5] = in[3] >> 1;
+        out[6] = ((in[3] >> 9) & 0x0f) | (in[4] << 4);
+        out[7] = in[4] >> 4;
+        out[8] = ((in[4] >> 12) & 0x01) | (in[5] << 1);
+        out[9] = ((in[5] >> 7) & 0x3f) | (in[6] << 6);
+        out[10] = in[6] >> 2;
+        out[11] = ((in[6] >> 10) & 0x07) | (in[7] << 3);
+        out[12] = in[7] >> 5;
         in += 8;
         out += 13;
     }
 }
 
 static void BS2POLq(poly *data, const uint8_t bytes[SABER_POLYBYTES]) {
+    /* This function does not reduce its output mod Q */
     size_t j;
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 8; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x1f) << 8);
-        out[1] = (in[1] >> 5 & (0x07)) | ((in[2] & 0xff) << 3) | ((in[3] & 0x03) << 11);
-        out[2] = (in[3] >> 2 & (0x3f)) | ((in[4] & 0x7f) << 6);
-        out[3] = (in[4] >> 7 & (0x01)) | ((in[5] & 0xff) << 1) | ((in[6] & 0x0f) << 9);
-        out[4] = (in[6] >> 4 & (0x0f)) | ((in[7] & 0xff) << 4) | ((in[8] & 0x01) << 12);
-        out[5] = (in[8] >> 1 & (0x7f)) | ((in[9] & 0x3f) << 7);
-        out[6] = (in[9] >> 6 & (0x03)) | ((in[10] & 0xff) << 2) | ((in[11] & 0x07) << 10);
-        out[7] = (in[11] >> 3 & (0x1f)) | ((in[12] & 0xff) << 5);
+        out[0] = (in[0]) | (in[1] << 8);
+        out[1] = (in[1] >> 5) | (in[2] << 3) | (in[3] << 11);
+        out[2] = (in[3] >> 2) | (in[4] << 6);
+        out[3] = (in[4] >> 7) | (in[5] << 1) | (in[6] << 9);
+        out[4] = (in[6] >> 4) | (in[7] << 4) | (in[8] << 12);
+        out[5] = (in[8] >> 1) | (in[9] << 7);
+        out[6] = (in[9] >> 6) | (in[10] << 2) | (in[11] << 10);
+        out[7] = (in[11] >> 3) | (in[12] << 5);
         in += 13;
         out += 8;
     }
@@ -72,11 +74,11 @@ static void POLp2BS(uint8_t bytes[SABER_POLYCOMPRESSEDBYTES], const poly *data)
     const uint16_t *in = data->coeffs;
     uint8_t *out = bytes;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff));
-        out[1] = ((in[0] >> 8) & 0x03) | ((in[1] & 0x3f) << 2);
-        out[2] = ((in[1] >> 6) & 0x0f) | ((in[2] & 0x0f) << 4);
-        out[3] = ((in[2] >> 4) & 0x3f) | ((in[3] & 0x03) << 6);
-        out[4] = ((in[3] >> 2) & 0xff);
+        out[0] = in[0];
+        out[1] = ((in[0] >> 8) & 0x03) | (in[1] << 2);
+        out[2] = ((in[1] >> 6) & 0x0f) | (in[2] << 4);
+        out[3] = ((in[2] >> 4) & 0x3f) | (in[3] << 6);
+        out[4] = in[3] >> 2;
         in += 4;
         out += 5;
     }
@@ -87,10 +89,10 @@ static void BS2POLp(poly *data, const uint8_t bytes[SABER_POLYCOMPRESSEDBYTES])
     const uint8_t *in = bytes;
     uint16_t *out = data->coeffs;
     for (j = 0; j < SABER_N / 4; j++) {
-        out[0] = (in[0] & (0xff)) | ((in[1] & 0x03) << 8);
-        out[1] = ((in[1] >> 2) & (0x3f)) | ((in[2] & 0x0f) << 6);
-        out[2] = ((in[2] >> 4) & (0x0f)) | ((in[3] & 0x3f) << 4);
-        out[3] = ((in[3] >> 6) & (0x03)) | ((in[4] & 0xff) << 2);
+        out[0] = in[0] | (in[1] << 8);
+        out[1] = (in[1] >> 2) | (in[2] << 6);
+        out[2] = (in[2] >> 4) | (in[3] << 4);
+        out[3] = (in[3] >> 6) | (in[4] << 2);
         in += 5;
         out += 4;
     }
diff --git a/test/duplicate_consistency/firesaber_avx2.yml b/test/duplicate_consistency/firesaber_avx2.yml
index 0ff04768..f5240334 100644
--- a/test/duplicate_consistency/firesaber_avx2.yml
+++ b/test/duplicate_consistency/firesaber_avx2.yml
@@ -6,7 +6,6 @@ consistency_checks:
       - api.h
       - cbd.h
       - pack_unpack.h
-      - kem.h
       - SABER_indcpa.h
       - SABER_params.h
       - verify.h
diff --git a/test/duplicate_consistency/firesaber_clean.yml b/test/duplicate_consistency/firesaber_clean.yml
index 5537ea62..bcfed7c0 100644
--- a/test/duplicate_consistency/firesaber_clean.yml
+++ b/test/duplicate_consistency/firesaber_clean.yml
@@ -5,7 +5,6 @@ consistency_checks:
       files:
       - api.h
       - cbd.h
-      - poly_mul.h
       - pack_unpack.h
       - SABER_indcpa.h
       - SABER_params.h
diff --git a/test/duplicate_consistency/lightsaber_avx2.yml b/test/duplicate_consistency/lightsaber_avx2.yml
index ed9ea4b8..dd6e9fed 100644
--- a/test/duplicate_consistency/lightsaber_avx2.yml
+++ b/test/duplicate_consistency/lightsaber_avx2.yml
@@ -6,7 +6,6 @@ consistency_checks:
       - api.h
       - cbd.h
       - pack_unpack.h
-      - kem.h
       - SABER_indcpa.h
       - SABER_params.h
       - verify.h
@@ -20,7 +19,6 @@ consistency_checks:
       files:
       - cbd.h
       - pack_unpack.h
-      - kem.h
       - SABER_indcpa.h
       - verify.h
       - kem.c
@@ -30,7 +28,6 @@ consistency_checks:
         implementation: avx2
       files:
       - cbd.h
-      - kem.h
       - pack_unpack.h
       - poly.h
       - SABER_indcpa.h
@@ -46,7 +43,6 @@ consistency_checks:
       files:
       - cbd.h
       - pack_unpack.h
-      - kem.h
       - SABER_indcpa.h
       - verify.h
       - kem.c
@@ -56,7 +52,6 @@ consistency_checks:
         implementation: avx2
       files:
       - cbd.h
-      - kem.h
       - pack_unpack.h
       - poly.h
       - SABER_indcpa.h
diff --git a/test/duplicate_consistency/lightsaber_clean.yml b/test/duplicate_consistency/lightsaber_clean.yml
index 8146b7d0..2f36ec86 100644
--- a/test/duplicate_consistency/lightsaber_clean.yml
+++ b/test/duplicate_consistency/lightsaber_clean.yml
@@ -5,7 +5,6 @@ consistency_checks:
       files:
       - api.h
       - cbd.h
-      - poly_mul.h
       - pack_unpack.h
       - SABER_indcpa.h
       - SABER_params.h
@@ -34,7 +33,6 @@ consistency_checks:
         implementation: avx2
       files:
       - cbd.h
-      - poly_mul.h
       - pack_unpack.h
       - SABER_indcpa.h
       - verify.h
@@ -60,7 +58,6 @@ consistency_checks:
         implementation: avx2
       files:
       - cbd.h
-      - poly_mul.h
       - pack_unpack.h
       - SABER_indcpa.h
       - verify.h
diff --git a/test/duplicate_consistency/saber_avx2.yml b/test/duplicate_consistency/saber_avx2.yml
index 4a04951d..0b4b60d7 100644
--- a/test/duplicate_consistency/saber_avx2.yml
+++ b/test/duplicate_consistency/saber_avx2.yml
@@ -6,7 +6,6 @@ consistency_checks:
       - api.h
       - cbd.h
       - pack_unpack.h
-      - kem.h
       - SABER_indcpa.h
       - SABER_params.h
       - verify.h
@@ -20,7 +19,6 @@ consistency_checks:
       files:
       - cbd.h
       - pack_unpack.h
-      - kem.h
       - SABER_indcpa.h
       - verify.h
       - kem.c
@@ -30,7 +28,6 @@ consistency_checks:
         implementation: avx2
       files:
       - cbd.h
-      - kem.h
       - pack_unpack.h
       - poly.h
       - SABER_indcpa.h
diff --git a/test/duplicate_consistency/saber_clean.yml b/test/duplicate_consistency/saber_clean.yml
index a2700ea8..7f5ba121 100644
--- a/test/duplicate_consistency/saber_clean.yml
+++ b/test/duplicate_consistency/saber_clean.yml
@@ -5,7 +5,6 @@ consistency_checks:
       files:
       - api.h
       - cbd.h
-      - poly_mul.h
       - pack_unpack.h
       - SABER_indcpa.h
       - SABER_params.h
@@ -34,7 +33,6 @@ consistency_checks:
         implementation: avx2
       files:
       - cbd.h
-      - poly_mul.h
       - pack_unpack.h
       - SABER_indcpa.h
       - verify.h