diff --git a/models/models.toml b/models/models.toml
index 9479ee9..6cc4c47 100644
--- a/models/models.toml
+++ b/models/models.toml
@@ -24,11 +24,10 @@ reference_source = "MAP-LIB_ReferenceResults (Dymola, branch v3.2.3)"
 #                  atol/reltol do not apply to it. Use solverReltol /
 #                  solverAtol to loosen the integrator (e.g. near-singular
 #                  Jacobians that force dt below eps under default tolerances).
-#   observedFilter = ["regex", ...]                 ([] = OM.simulate default)
-#                  Retain alias/observed variables whose eliminated name matches
-#                  any pattern through structural_simplify, so validation can read
-#                  signals MTK would otherwise prune. Threaded to
-#                  OM.simulate(...; observedFilter = [...]).
+#   observedFilter = ["regex", ...]                 ([] = no filter)
+#                  Keep only the observed and alias variables whose name matches
+#                  a pattern (with what they, the system and the events read).
+#                  Threaded to OM.translate and OM.simulate(...; observedFilter = [...]).
 #
 # Use cases for initializealg:
 #   - DAE with rank-deficient algebraic Jacobian at t=0 (e.g. flow-state
diff --git a/src/comparison.jl b/src/comparison.jl
index f2e7de0..eb2ef8e 100644
--- a/src/comparison.jl
+++ b/src/comparison.jl
@@ -383,8 +383,7 @@ function validate_against_reference(sol, spec,
     end
     #= Every listed signal is compared: one the solution lacks fails the validation (it was
        skipped, and a model validated on a subset: IdealGasH2O on 1 of 3 signals, MSL 4
-       CCCV_Stack on 1 of 13). The harness keeps the reference signals observable
-       (_reference_signal_filter). =#
+       CCCV_Stack on 1 of 13). =#
     if n_skipped > 0
         @warn "Validation: $n_skipped/$(length(signal_names)) reference signals are not in the solution"
     end
diff --git a/src/runner.jl b/src/runner.jl
index 6b75ccf..2ff2abe 100644
--- a/src/runner.jl
+++ b/src/runner.jl
@@ -285,40 +285,6 @@ function _resolve_solver(name::String)
     return NamedTuple()
 end
 
-# Escape regex metacharacters so signal names containing `.`, `[`, `]`, etc.
-# match literally inside OMBackend's `occursin(Regex(p), name)` filter.
-_escape_regex(s::AbstractString) =
-    replace(s, r"[\\.\[\]\(\)\+\*\?\^\$\|]" => sm -> "\\" * sm)
-
-# Read the reference CSV header and return regex patterns (anchored,
-# underscore-form) for every non-`time` column so OMBackend's observedFilter
-# preserves the alias-map entries the validate phase will look up.
-function _reference_signal_filter(ref_dir::String,
-                                   reference_file::String,
-                                   signal_mapping::Dict{String, String})::Vector{String}
-    isempty(reference_file) && return String[]
-    local csv_path = joinpath(ref_dir, "csv", reference_file * ".csv")
-    isfile(csv_path) || return String[]
-    local header_line = ""
-    try
-        open(csv_path) do io; header_line = readline(io); end
-    catch
-        return String[]
-    end
-    local patterns = String[]
-    local seen = Set{String}()
-    for col in split_csv_header(header_line)
-        local raw = strip(col, ['"', ' ', '\t', '\r', '\n'])
-        (isempty(raw) || raw == "time") && continue
-        local under = haskey(signal_mapping, String(raw)) ? signal_mapping[String(raw)] : modelica_to_omjl_name(String(raw))
-        if !(under in seen)
-            push!(seen, under)
-            push!(patterns, string("^", _escape_regex(under), "\$"))
-        end
-    end
-    return patterns
-end
-
 #= Cap the cost of `showerror`. Some exceptions (notably MTK's
    UnsolvableCallbackError / ExtraEquationsSystemException) render the entire
    equation / callback system — O(system size) — which can take many minutes to
@@ -400,6 +366,10 @@ function _run_model_phases(model_name::String,
 
     has_csv_ref = !isempty(reference_file)
     has_inline_ref = !isempty(reference)
+    #= A registry filter goes to the translate and the simulate alike: a simulate by name with another filter
+       translates again (OM._reusableTranslation). The reference header's signals as the filter, passed to the
+       simulate only, never applied (the backend phase's translation was reused) and none was missing. =#
+    local _of::NamedTuple = isempty(observed_filter) ? NamedTuple() : (; observedFilter = observed_filter)
 
     for phase_int in phase_ints
         if phase_int == Int(VALIDATE) && !has_csv_ref && !has_inline_ref
@@ -415,7 +385,7 @@ function _run_model_phases(model_name::String,
             if phase_int == Int(FRONTEND)
                 OM.flatten(model_name; msl...)
             elseif phase_int == Int(BACKEND)
-                OM.translate(model_name; msl...,
+                OM.translate(model_name; msl..., _of...,
                              checkSimCode = check_sim_code)
             elseif phase_int == Int(SIMULATE)
                 # Build kwargs: solver + dtmax (if non-zero) + initializealg (if specified).
@@ -423,11 +393,6 @@ function _run_model_phases(model_name::String,
                 local _sa = isempty(solver_name) ? NamedTuple() : _resolve_solver(solver_name)
                 local _extra = dtmax > 0.0 ? (; dtmax = dtmax) : NamedTuple()
                 local _ia = isempty(init_alg) ? NamedTuple() : _resolve_init_alg(init_alg)
-                #= The reference's signals stay observable (the backend drops eliminated array
-                   elements' aliases by default): a registry filter, else the reference header. =#
-                local _filter = isempty(observed_filter) ?
-                    _reference_signal_filter(ref_dir, reference_file, signal_mapping) : observed_filter
-                local _of = isempty(_filter) ? NamedTuple() : (; observedFilter = _filter)
                 local _mi = (; maxiters = maxiters > 0.0 ? round(Int, maxiters) : DEFAULT_MAXITERS)
                 local _tol = NamedTuple()
                 if solver_reltol > 0.0
diff --git a/src/types.jl b/src/types.jl
index 61432e6..4323658 100644
--- a/src/types.jl
+++ b/src/types.jl
@@ -80,9 +80,9 @@ struct ModelSpec
     # validation tolerances for trajectory comparison. 0.0 = use SciML default.
     solverAtol::Float64
     solverReltol::Float64
-    # OMBackend observedFilter regex patterns. Non-empty: passed to OM.simulate to retain
-    # matching alias/observed variables through structural_simplify so validation can read
-    # them. Empty = OM.simulate default (no explicit retention).
+    # OMBackend observedFilter regex patterns. Non-empty: passed to OM.translate and OM.simulate
+    # (the translation's observed equations and alias observations: those the patterns match and
+    # what they, the system and the events read). Empty = no filter.
     observedFilter::Vector{String}
     # Hard cap on integrator iterations, passed to OM.simulate as `maxiters`. 0.0 = use
     # the SciML default (1e5). A bounded value (e.g. 50000) lets a stiff / chattering