diff --git a/docs/now/2026-08-22-prune-platform-baseline.md b/docs/now/2026-08-22-prune-platform-baseline.md
new file mode 100644
index 000000000..0fb49f57d
--- /dev/null
+++ b/docs/now/2026-08-22-prune-platform-baseline.md
@@ -0,0 +1,32 @@
+# NOW — a CI-passing test sat in the failing baseline
+
+Last updated: 2026-08-22
+
+## Prune four macOS-only failures from the ratchet baseline (Closes #2387)
+
+- Branch: `fix/2387-prune-platform-baseline`
+- Issue: #2387
+
+### Что легло
+
+`scripts/ci/test-baseline.txt`, 382 → 378 entries. The four `spec_first_*` tests in
+`bitnet_layer`, `bitnet_mlp`, `bitnet_mlp3` and `bitnet_neuron_nchunk` fail on macOS/arm64,
+where the baseline was generated, and **pass on the Linux runner** — the first
+`test-ratchet` run on master reported all four under "now PASS".
+
+A CI-passing test left in the baseline is invisible if it breaks: the ratchet fails only on
+names new relative to the baseline, so these four could go red on Linux and the gate would
+stay green. That is the blind spot the ratchet exists to remove, reintroduced in its own
+input. **The baseline must describe the platform that gates.**
+
+### Границы честности (BINDING)
+
+- **The cause of the platform divergence was not investigated.** Endianness, float
+  formatting, path separators and filesystem ordering are all plausible; none is confirmed.
+  That they are all `spec_first_*` inference-vs-reference comparisons hints at a numeric or
+  ordering difference — a hint, not a finding.
+- Pruning means a **macOS** developer now sees four local failures the ratchet does not know
+  about. Correct trade — CI is what gates — but a trade, noted in the file itself.
+- **This fixes nothing.** The four still fail on macOS.
+- The baseline is still generated on macOS. Regenerating it from a runner log would be
+  better and is not done here.
diff --git a/scripts/ci/test-baseline.txt b/scripts/ci/test-baseline.txt
index 871a25f8b..40cd3969d 100644
--- a/scripts/ci/test-baseline.txt
+++ b/scripts/ci/test-baseline.txt
@@ -1,7 +1,9 @@
 # Generated by scripts/ci/test_ratchet.py --emit-baseline.
 # Do not hand-edit: regenerate from a real --no-fail-fast log.
 # Each line is target<TAB>test for a test failing on master.
-# 73 targets ran; 382 failing tests (one pruned by #2384).
+# 73 targets ran; 378 failing tests on the Linux runner.
+# Four macOS-only failures pruned by #2387: the baseline describes the
+# platform that gates, and a CI-passing test left in it is invisible if it breaks.
 src/main.rs	compiler::tests_compiler_rejects::lowers_only_first_of_two_modules_characterization
 src/main.rs	compiler::tests_phase40_coverage::test_for_range_loop_unroll
 src/main.rs	compiler::tests_phase40_coverage::test_parse_for_range
@@ -15,10 +17,6 @@ src/main.rs	compiler::tests_w458::tests_w459::array_param_bound_from_test_block
 src/main.rs	compiler::tests_w458::tests_w459::test_block_emits_real_function_call
 src/main.rs	lex_conform::tests::lexer_matches_its_conformance_table
 src/main.rs	parse_conform::tests::parser_matches_its_conformance_table
-tests/bitnet_layer.rs	spec_first_layer2_packs_two_neuron_trits
-tests/bitnet_mlp.rs	spec_first_mlp2_two_layer_inference_matches_reference
-tests/bitnet_mlp3.rs	spec_first_mlp3_three_layer_inference_matches_reference
-tests/bitnet_neuron_nchunk.rs	spec_first_neuron_n_accumulates_and_quantizes
 tests/icarus_lowerable.rs	accepts_known_lowerable_witnesses
 tests/icarus_lowerable.rs	accepts_w545_primitive_scalar_array_return
 tests/icarus_lowerable.rs	accepts_w547_signed_primitive_scalar_array_return