diff --git a/src/algorithms/PVT/libs/rtklib_solver.cc b/src/algorithms/PVT/libs/rtklib_solver.cc
index b0b0a5445..f4a3a4c57 100644
--- a/src/algorithms/PVT/libs/rtklib_solver.cc
+++ b/src/algorithms/PVT/libs/rtklib_solver.cc
@@ -2850,20 +2850,20 @@ bool Rtklib_Solver::get_PVT(const std::map<int, Gnss_Synchro> &gnss_observables_
                     // "not used" rather than as missing/no-az-el (which
                     // otherwise looks identical to a tracking problem).
                     //
-                    // std::isfinite() guards against corrupt ephemeris/almanac
-                    // (e.g. an out-of-range eccentricity) propagating NaN through
-                    // eph2pos()/alm2pos()/satazel() -- none of RTKLIB's own
-                    // internal sanity checks (A <= 0, Kepler iteration overflow,
-                    // geodist()'s Earth-radius check, the elevation mask) catch a
-                    // NaN, since ordinary numeric comparisons are vacuously false
-                    // against it. Without this, a NaN azel would actually pass
-                    // the old != 0.0 check (NaN != 0.0 is true) and get reported
-                    // into the monitor as a literal NaN azimuth/elevation.
+                    // Deliberately NOT filtered for NaN here: a corrupt ephemeris/
+                    // almanac (e.g. an out-of-range eccentricity) can propagate NaN
+                    // through eph2pos()/alm2pos()/satazel(), and that NaN is reported
+                    // as-is rather than silently dropping the satellite from the
+                    // monitor -- consumers (gnss-sdr-CtrlApp) are expected to handle
+                    // a NaN azimuth_deg/elevation_deg explicitly (e.g. no sky plot
+                    // entry, "NaN" printed in the signal table) rather than have it
+                    // hidden here. What NaN must NOT do is influence the position fix
+                    // itself -- see rescode()'s prange() isfinite guard in
+                    // rtklib_pntpos.cc for where that's actually enforced.
                     d_monitor_pvt.tracked_satellites.clear();
                     for (int sat_idx = 0; sat_idx < MAXSAT; sat_idx++)
                         {
-                            const bool has_azel = std::isfinite(pvt_ssat[sat_idx].azel[0]) && std::isfinite(pvt_ssat[sat_idx].azel[1]) &&
-                                                  ((pvt_ssat[sat_idx].azel[0] != 0.0) || (pvt_ssat[sat_idx].azel[1] != 0.0));
+                            const bool has_azel = (pvt_ssat[sat_idx].azel[0] != 0.0) || (pvt_ssat[sat_idx].azel[1] != 0.0);
                             if (!has_azel)
                                 {
                                     continue;
diff --git a/src/algorithms/libs/rtklib/rtklib_pntpos.cc b/src/algorithms/libs/rtklib/rtklib_pntpos.cc
index f1cb6626b..f47eafecb 100644
--- a/src/algorithms/libs/rtklib/rtklib_pntpos.cc
+++ b/src/algorithms/libs/rtklib/rtklib_pntpos.cc
@@ -736,14 +736,26 @@ int rescode(int iter, const obsd_t *obs, int n, const double *rs,
                     continue;
                 }
             double elaux = satazel(pos, e, azel + i * 2);
-            if (elaux < opt->elmin)
+            /* !isfinite(elaux): corrupt ephemeris/almanac (e.g. an out-of-range
+             * eccentricity) can propagate NaN through eph2pos()/alm2pos() into
+             * azel -- azel itself is intentionally left as computed (including
+             * NaN) for callers that report it (e.g. the monitor's per-satellite
+             * az/el), but a NaN elevation must not be allowed into the position
+             * solve below: elaux < opt->elmin is a no-op against NaN (every
+             * relational comparison against NaN is false), so it would
+             * otherwise fall through as if elmin had been satisfied. */
+            if (elaux < opt->elmin || !std::isfinite(elaux))
                 {
                     trace(4, "satazel error. el = %lf , elmin = %lf\n", elaux, opt->elmin);
                     continue;
                 }
             /* psudorange with code bias correction */
             double iono_scale = 1.0;
-            if ((P = prange(obs + i, nav, azel + i * 2, iter, opt, &vmeas, &iono_scale)) == 0.0)
+            P = prange(obs + i, nav, azel + i * 2, iter, opt, &vmeas, &iono_scale);
+            /* Same NaN concern as above, for prange()'s own corrections (iono/
+             * tropo/code bias) -- P == 0.0 is prange()'s own "no valid pseudorange"
+             * sentinel and doesn't catch NaN either. */
+            if (P == 0.0 || !std::isfinite(P))
                 {
                     trace(4, "prange error\n");
                     continue;