diff --git a/lib/prolog/conformance.sh b/lib/prolog/conformance.sh
new file mode 100755
index 00000000..4dd949c8
--- /dev/null
+++ b/lib/prolog/conformance.sh
@@ -0,0 +1,106 @@
+#!/usr/bin/env bash
+# Run every Prolog test suite via sx_server and refresh scoreboard.{json,md}.
+# Exit 0 if all green, 1 if any failures.
+set -euo pipefail
+
+HERE="$(cd "$(dirname "$0")" && pwd)"
+ROOT="$(cd "$HERE/../.." && pwd)"
+SX="${SX_SERVER:-/root/rose-ash/hosts/ocaml/_build/default/bin/sx_server.exe}"
+
+if [[ ! -x "$SX" ]]; then
+  echo "sx_server not found at $SX (set SX_SERVER env to override)" >&2
+  exit 2
+fi
+
+cd "$ROOT"
+
+# name : test-file : runner-fn
+SUITES=(
+  "parse:lib/prolog/tests/parse.sx:pl-parse-tests-run!"
+  "unify:lib/prolog/tests/unify.sx:pl-unify-tests-run!"
+  "clausedb:lib/prolog/tests/clausedb.sx:pl-clausedb-tests-run!"
+  "solve:lib/prolog/tests/solve.sx:pl-solve-tests-run!"
+  "append:lib/prolog/tests/programs/append.sx:pl-append-tests-run!"
+  "reverse:lib/prolog/tests/programs/reverse.sx:pl-reverse-tests-run!"
+  "member:lib/prolog/tests/programs/member.sx:pl-member-tests-run!"
+  "nqueens:lib/prolog/tests/programs/nqueens.sx:pl-nqueens-tests-run!"
+  "family:lib/prolog/tests/programs/family.sx:pl-family-tests-run!"
+)
+
+SCRIPT='(epoch 1)
+(load "lib/prolog/tokenizer.sx")
+(load "lib/prolog/parser.sx")
+(load "lib/prolog/runtime.sx")'
+for entry in "${SUITES[@]}"; do
+  IFS=: read -r _ file _ <<< "$entry"
+  SCRIPT+=$'\n(load "'"$file"$'")'
+done
+for entry in "${SUITES[@]}"; do
+  IFS=: read -r _ _ fn <<< "$entry"
+  SCRIPT+=$'\n(eval "('"$fn"$')")'
+done
+
+OUTPUT="$(printf '%s\n' "$SCRIPT" | "$SX" 2>&1)"
+
+mapfile -t LINES < <(printf '%s\n' "$OUTPUT" | grep -E '^\{:failed')
+
+if [[ ${#LINES[@]} -ne ${#SUITES[@]} ]]; then
+  echo "Expected ${#SUITES[@]} suite results, got ${#LINES[@]}" >&2
+  echo "---- raw output ----" >&2
+  printf '%s\n' "$OUTPUT" >&2
+  exit 3
+fi
+
+TOTAL_PASS=0
+TOTAL_FAIL=0
+TOTAL=0
+JSON_SUITES=""
+MD_ROWS=""
+
+for i in "${!SUITES[@]}"; do
+  IFS=: read -r name _ _ <<< "${SUITES[$i]}"
+  line="${LINES[$i]}"
+  passed=$(grep -oE ':passed [0-9]+' <<< "$line" | grep -oE '[0-9]+')
+  total=$(grep -oE ':total [0-9]+' <<< "$line" | grep -oE '[0-9]+')
+  failed=$(grep -oE ':failed [0-9]+' <<< "$line" | grep -oE '[0-9]+')
+  TOTAL_PASS=$((TOTAL_PASS + passed))
+  TOTAL_FAIL=$((TOTAL_FAIL + failed))
+  TOTAL=$((TOTAL + total))
+  status="ok"
+  [[ "$failed" -gt 0 ]] && status="FAIL"
+  [[ -n "$JSON_SUITES" ]] && JSON_SUITES+=","
+  JSON_SUITES+="\"$name\":{\"passed\":$passed,\"total\":$total,\"failed\":$failed}"
+  MD_ROWS+="| $name | $passed | $total | $status |"$'\n'
+done
+
+WHEN="$(date -Iseconds 2>/dev/null || date)"
+
+cat > "$HERE/scoreboard.json" <<JSON
+{
+  "total_passed": $TOTAL_PASS,
+  "total_failed": $TOTAL_FAIL,
+  "total": $TOTAL,
+  "suites": {$JSON_SUITES},
+  "generated": "$WHEN"
+}
+JSON
+
+cat > "$HERE/scoreboard.md" <<MD
+# Prolog scoreboard
+
+**$TOTAL_PASS / $TOTAL passing** ($TOTAL_FAIL failure(s)).
+Generated $WHEN.
+
+| Suite | Passed | Total | Status |
+|-------|--------|-------|--------|
+$MD_ROWS
+Run \`bash lib/prolog/conformance.sh\` to refresh. Override the binary
+with \`SX_SERVER=path/to/sx_server.exe bash …\`.
+MD
+
+if [[ "$TOTAL_FAIL" -gt 0 ]]; then
+  echo "$TOTAL_FAIL failure(s) across $TOTAL tests" >&2
+  exit 1
+fi
+
+echo "All $TOTAL tests pass."
diff --git a/lib/prolog/scoreboard.json b/lib/prolog/scoreboard.json
new file mode 100644
index 00000000..44d60788
--- /dev/null
+++ b/lib/prolog/scoreboard.json
@@ -0,0 +1,7 @@
+{
+  "total_passed": 183,
+  "total_failed": 0,
+  "total": 183,
+  "suites": {"parse":{"passed":25,"total":25,"failed":0},"unify":{"passed":47,"total":47,"failed":0},"clausedb":{"passed":14,"total":14,"failed":0},"solve":{"passed":62,"total":62,"failed":0},"append":{"passed":6,"total":6,"failed":0},"reverse":{"passed":6,"total":6,"failed":0},"member":{"passed":7,"total":7,"failed":0},"nqueens":{"passed":6,"total":6,"failed":0},"family":{"passed":10,"total":10,"failed":0}},
+  "generated": "2026-04-25T06:19:36+00:00"
+}
diff --git a/lib/prolog/scoreboard.md b/lib/prolog/scoreboard.md
new file mode 100644
index 00000000..8d21270b
--- /dev/null
+++ b/lib/prolog/scoreboard.md
@@ -0,0 +1,19 @@
+# Prolog scoreboard
+
+**183 / 183 passing** (0 failure(s)).
+Generated 2026-04-25T06:19:36+00:00.
+
+| Suite | Passed | Total | Status |
+|-------|--------|-------|--------|
+| parse | 25 | 25 | ok |
+| unify | 47 | 47 | ok |
+| clausedb | 14 | 14 | ok |
+| solve | 62 | 62 | ok |
+| append | 6 | 6 | ok |
+| reverse | 6 | 6 | ok |
+| member | 7 | 7 | ok |
+| nqueens | 6 | 6 | ok |
+| family | 10 | 10 | ok |
+
+Run `bash lib/prolog/conformance.sh` to refresh. Override the binary
+with `SX_SERVER=path/to/sx_server.exe bash …`.
diff --git a/plans/prolog-on-sx.md b/plans/prolog-on-sx.md
index faadb9e0..19610a0e 100644
--- a/plans/prolog-on-sx.md
+++ b/plans/prolog-on-sx.md
@@ -61,7 +61,7 @@ Representation choices (finalise in phase 1, document here):
   - [x] `member.pl` — generate all solutions via backtracking — `lib/prolog/tests/programs/member.{pl,sx}`. Classic 2-clause `member(X, [X|_])` + `member(X, [_|T]) :- member(X, T)`. 7 tests cover bound-element hit/miss, empty list, generator (count = list length), first-solution binding, duplicate matches counted twice, anonymous head-cell unification.
   - [x] `nqueens.pl` — 8-queens — `lib/prolog/tests/programs/nqueens.{pl,sx}`. Permute-and-test formulation: `queens(L, Qs) :- permute(L, Qs), safe(Qs)` + `select` + `safe` + `no_attack`. Tested at N=1 (1), N=2 (0), N=3 (0), N=4 (2), N=5 (10) plus first-solution check at N=4 = `[2, 4, 1, 3]`. N=8 omitted — interpreter is too slow (40320 perms); add once compiled clauses or constraint-style placement land. `range/3` skipped pending arithmetic-comparison built-ins (`>/2` etc.).
   - [x] `family.pl` — facts + rules (parent/ancestor) — `lib/prolog/tests/programs/family.{pl,sx}`. 5 parent facts + male/female + derived `father`/`mother`/`ancestor`/`sibling`. 10 tests cover direct facts, fact count, transitive ancestor through 3 generations, descendant counting, gender-restricted father/mother, sibling via shared parent + `\=`.
-- [ ] `lib/prolog/conformance.sh` + runner, `scoreboard.json` + `scoreboard.md`
+- [x] `lib/prolog/conformance.sh` + runner, `scoreboard.json` + `scoreboard.md` — bash script feeds load + eval epoch script to sx_server, parses each suite's `{:failed N :passed N :total N :failures (...)}` line, writes JSON (machine) + MD (human) scoreboards. Exit non-zero on any failure. `SX_SERVER` env var overrides binary path. First scoreboard: 183 / 183.
 - [x] Target: all 5 classic programs passing — append (6) + reverse (6) + member (7) + nqueens (6) + family (10) = 35 program tests, all green. Phase 3 architecturally complete bar the conformance harness/scoreboard.
 
 ### Phase 4 — operator table + more built-ins (next run)
@@ -88,6 +88,7 @@ Representation choices (finalise in phase 1, document here):
 
 _Newest first. Agent appends on every commit._
 
+- 2026-04-25 — Conformance harness landed. `lib/prolog/conformance.sh` runs all 9 suites in one sx_server epoch, parses the `{:failed/:passed/:total/:failures}` summary lines, and writes `scoreboard.json` + `scoreboard.md`. `SX_SERVER` env var overrides the binary path; default points at the main-repo build. Phase 3 fully complete: 183 / 183 passing across parse/unify/clausedb/solve/append/reverse/member/nqueens/family.
 - 2026-04-25 — `family.pl` fifth classic program — completes the 5-program target. 5-fact pedigree + male/female + derived father/mother/ancestor/sibling. 10 tests cover fact lookup + count, transitive ancestor through 3 generations, descendant counting (5), gender-restricted derivations, sibling via shared parent guarded by `\=`. Total 183 (+10). All 5 classic programs ticked; Phase 3 needs only conformance harness + scoreboard left.
 - 2026-04-25 — `nqueens.pl` fourth classic program. Permute-and-test variant exercises every Phase-3 feature: lists with `[H|T]` cons sugar, multi-clause backtracking, recursive `permute`/`select`/`safe`/`no_attack`, `is/2` arithmetic on diagonals, `\=/2` for diagonal-conflict check. 6 tests at N ∈ {1,2,3,4,5} with expected counts {1,0,0,2,10} + first-solution `[2,4,1,3]`. N=5 takes ~30s (120 perms × safe-check); N=8 omitted as it would be ~thousands of seconds. Total 173 (+6).
 - 2026-04-25 — `member.pl` third classic program. Standard 2-clause definition; 7 tests cover bound-element hit/miss, empty-list fail, generator-count = list length, first-solution binding (X=11), duplicate elements matched twice on backtrack, anonymous-head unification (`member(a, [X, b, c])` binds X=a). Total 167 (+7).