diff --git a/docs/eval-bench.sh b/docs/eval-bench.sh
new file mode 100755
index 0000000..8c4e70c
--- /dev/null
+++ b/docs/eval-bench.sh
@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# Ralph eval-time benchmark.
+# Usage: ralph-bench.sh <runs>
+# Primary metric: median cpuTime (NIX_SHOW_STATS) — stable, isolates eval work.
+# Secondary: median wall-clock (noisy, IO-bound).
+# Eval cache disabled; fetcher cache warm via persistent $HOME.
+set -uo pipefail
+
+RUNS="${1:-5}"
+export HOME=/tmp/ralph-evalhome
+mkdir -p "$HOME"
+FLAKE="${FLAKE:-/home/srid/code/nixos-config/.worktrees/ralph}"
+
+declare -A ATTR
+ATTR[pureintent]="$FLAKE#nixosConfigurations.pureintent.config.system.build.toplevel.drvPath"
+ATTR[zest]="$FLAKE#homeConfigurations.\"srid@zest\".activationPackage.drvPath"
+
+median() { sort -n | awk '{a[NR]=$1} END{ if(NR%2){print a[(NR+1)/2]} else {print (a[NR/2]+a[NR/2+1])/2} }'; }
+now_ms() { date +%s%3N; }
+
+for target in pureintent zest; do
+  attr="${ATTR[$target]}"
+  cpus=(); walls=()
+  for i in $(seq 1 "$RUNS"); do
+    start=$(now_ms)
+    NIX_SHOW_STATS=1 nix eval --raw "$attr" --option eval-cache false \
+      >/dev/null 2>/tmp/ralph-stats.json
+    rc=$?
+    end=$(now_ms)
+    if [ $rc -ne 0 ]; then echo "$target run $i FAILED (rc=$rc)"; tail -8 /tmp/ralph-stats.json; exit 1; fi
+    cpu=$(grep -o '"cpuTime": [0-9.]*' /tmp/ralph-stats.json | head -1 | grep -o '[0-9.]*')
+    cpus+=("$cpu"); walls+=("$((end - start))")
+  done
+  medcpu=$(printf '%s\n' "${cpus[@]}" | median)
+  medwall=$(printf '%s\n' "${walls[@]}" | median)
+  medwall_s=$(awk "BEGIN{printf \"%.2f\", $medwall/1000}")
+  printf '%-11s cpuTime=%ss (median of %d) | wall=%ss | cpu=[%s] wall_ms=[%s]\n' \
+    "$target" "$medcpu" "$RUNS" "$medwall_s" "${cpus[*]}" "${walls[*]}"
+done
diff --git a/docs/eval-time-ralph-report.md b/docs/eval-time-ralph-report.md
new file mode 100644
index 0000000..8271172
--- /dev/null
+++ b/docs/eval-time-ralph-report.md
@@ -0,0 +1,54 @@
+# Nix Eval-Time Optimization (Ralph)
+
+Iterative measurement-driven reduction of Nix **evaluation** time for the two
+configs that get evaluated most often on this machine:
+
+- `pureintent` — the NixOS system (`nixosConfigurations.pureintent`)
+- `zest` — the home-manager config (`homeConfigurations."srid@zest"`)
+
+## Methodology
+
+Primary metric: **`cpuTime`** from `NIX_SHOW_STATS=1` — it isolates pure
+evaluation work and is stable run-to-run (~3% spread). Wall-clock is reported
+too but is IO/cache-bound and noisy (±15%), so it is *not* used for commit
+decisions.
+
+```sh
+# docs/eval-bench.sh <runs>   (default 5)
+# - persistent $HOME=/tmp/ralph-evalhome keeps the fetcher/git cache warm
+# - --option eval-cache false forces a full re-eval every run
+# - reports median cpuTime + median wall over N runs, per target
+nix eval --raw .#nixosConfigurations.pureintent.config.system.build.toplevel.drvPath --option eval-cache false
+nix eval --raw '.#homeConfigurations."srid@zest".activationPackage.drvPath'        --option eval-cache false
+```
+
+No `nix store gc` between runs (per request). The eval cache is disabled rather
+than cleared, so the warm fetcher cache removes input-fetch noise.
+
+**Commit rule:** commit only if median cpuTime improves >3% for a target with no
+regression on the other. Behaviour may change *only* by dropping inputs/features
+first confirmed unused.
+
+## Baseline (median of 5)
+
+| Target     | cpuTime (s) | wall (s) | cpuTime runs |
+|------------|-------------|----------|--------------|
+| pureintent | **23.17**   | 38.91    | 21.99 / 23.45 / 23.76 / 23.17 / 22.74 |
+| zest       | **14.93**   | 22.09    | 14.93 / 14.50 / 13.86 / 15.49 / 14.96 |
+
+Env: nix 2.34.7, x86_64-linux. Stats at baseline (pureintent): nrThunks 20.4M,
+nrFunctionCalls 13.0M, sets.bytes 868MB, values.bytes 503MB, gcFraction 0.058.
+
+## Optimization log
+
+| Cycle | Change | pureintent | zest | Verdict |
+|-------|--------|------------|------|---------|
+| 0 | baseline | 23.17 | 14.93 | — |
+
+## Dead ends
+
+_(investigated, no improvement — recorded so we don't retry)_
+
+## Key findings
+
+_(filled in at wrap-up)_