Datasets:

drawais
/

circuitqa

#!/usr/bin/env python3
"""Shared scorer for CircuitQA.
Takes raw model predictions and prints the metrics per category and overall. Pure standard library: no installs, so it
runs identically on every lab machine.
    python score.py --pred preds/claude-opus-5.5.B.jsonl \
                    --ref corpus/eval_sample.jsonl \
                    --model "Claude Opus 5.5" --condition B
Prediction file: one JSON object per line, {"question_hash": str, "prediction": str}.
Every percentage is printed next to the integer count it came from (k/n), so any
row in the final table can be audited. A percentage that is not k/n*100 for
integer k did not come from this script.
    python score.py --verify-csv results_template_per_category.csv
checks a filled-in table for exactly that.
"""
import argparse, json, math, re, sys, collections, hashlib, os

CATS = ["structural_analysis","design_quality","functional_behavior",
        "code_understanding","optimization","cross_stage"]

# ---------------------------------------------------------------- normalisation
UNITS = [
    (r'\bnanoseconds?\b|\bnsec\b', 'ns'), (r'\bpicoseconds?\b|\bpsec\b', 'ps'),
    (r'\bmicroseconds?\b|\busec\b', 'us'), (r'\bmilliseconds?\b|\bmsec\b', 'ms'),
    (r'\bmilliwatts?\b', 'mw'),  (r'\bmicrowatts?\b', 'uw'),
    (r'\bnanowatts?\b', 'nw'),   (r'\bwatts?\b', 'w'),
    (r'\bmegahertz\b', 'mhz'),   (r'\bgigahertz\b', 'ghz'), (r'\bkilohertz\b', 'khz'),
    (r'\bsquare micrometer?s?\b|\bmicrometer\^?2\b|\bum\^?2\b|\bµm\^?2\b', 'um2'),
    (r'\bflip[- ]?flops?\b|\bffs?\b', 'ff'),
    (r'\bregisters?\b', 'reg'), (r'\bwires?\b', 'wire'), (r'\bparameters?\b', 'param'),
    (r'\bgates?\b', 'gate'),    (r'\bcells?\b', 'cell'),
]

def squash(s):
    """Lowercase, strip punctuation noise, collapse whitespace."""
    s = (s or "").strip().lower()
    s = s.replace('−', '-').replace('µ', 'u')
    s = s.replace('²', '2').replace('³', '3')   # um2 / um3 superscripts
    s = re.sub(r'[`*_"\']', '', s)
    s = re.sub(r'\s*=\s*', '=', s)                        # "a = b" -> "a=b"
    s = re.sub(r'\s*,\s*', ' ', s)                        # commas are not content
    s = re.sub(r'\band\b', '', s)                         # list conjunction
    s = re.sub(r'\s+', ' ', s)
    return s.strip(' .')

def normalise(s):
    """squash + canonical units + canonical number formatting.
    Makes '2.5ns', '2.50 ns' and '2.5 nanoseconds' identical, and 0.0020 == 0.002.
    """
    s = squash(s)
    for pat, rep in UNITS:
        s = re.sub(pat, rep, s)
    s = re.sub(r'(\d)\s+([a-z]{1,4}\d?)\b', r'\1\2', s)      # "2.5 ns" -> "2.5ns"
    s = re.sub(r'(?<![\w.])(\d+\.?\d*)', lambda m: fmtnum(m.group(1)), s)
    s = re.sub(r'\s+', ' ', s)
    return s.strip(' .')

def fmtnum(t):
    try: v = float(t)
    except ValueError: return t
    if v == int(v) and abs(v) < 1e15: return str(int(v))
    return f"{v:.6f}".rstrip('0').rstrip('.')

def numbers(s):
    return [float(x) for x in re.findall(r'-?\d+\.?\d*(?:[eE][-+]?\d+)?', s or "")]

def toks(s):
    return re.findall(r'[a-z0-9_.\[\]:]+', normalise(s))

# ---------------------------------------------------------------------- metrics
def num_match(pred, ref, rtol=1e-3):
    """Do the reference's numbers all appear in the prediction (within rtol)?"""
    rn, pn = numbers(ref), numbers(pred)
    if not rn: return None                     # no numeric content: not applicable
    for r in rn:
        if not any(abs(p - r) <= rtol * max(1.0, abs(r)) for p in pn): return False
    return True

def f1(pred, ref):
    p, r = toks(pred), toks(ref)
    if not p or not r: return 0.0
    common = collections.Counter(p) & collections.Counter(r)
    ov = sum(common.values())
    if ov == 0: return 0.0
    prec, rec = ov/len(p), ov/len(r)
    return 2*prec*rec/(prec+rec)

def partial(pred, ref, thr=0.60):
    return f1(pred, ref) >= thr

def bleu4(pred, ref):
    p, r = toks(pred), toks(ref)
    if not p: return 0.0
    logs, nonzero = 0.0, 0
    for n in (1,2,3,4):
        pg = collections.Counter(tuple(p[i:i+n]) for i in range(len(p)-n+1))
        rg = collections.Counter(tuple(r[i:i+n]) for i in range(len(r)-n+1))
        tot = sum(pg.values())
        if tot == 0: continue
        clip = sum(min(c, rg[g]) for g, c in pg.items())
        if clip == 0: return 0.0
        logs += math.log(clip/tot); nonzero += 1
    if nonzero == 0: return 0.0
    bp = 1.0 if len(p) > len(r) else math.exp(1 - len(r)/max(1, len(p)))
    return 100 * bp * math.exp(logs/nonzero)

def rouge_l(pred, ref):
    p, r = toks(pred), toks(ref)
    if not p or not r: return 0.0
    dp = [[0]*(len(r)+1) for _ in range(len(p)+1)]
    for i in range(1, len(p)+1):
        for j in range(1, len(r)+1):
            dp[i][j] = dp[i-1][j-1]+1 if p[i-1]==r[j-1] else max(dp[i-1][j], dp[i][j-1])
    l = dp[-1][-1]
    if l == 0: return 0.0
    prec, rec = l/len(p), l/len(r)
    return 100 * 2*prec*rec/(prec+rec)

STOP = set("a an the is are was were of to in on at for with and or by from "
           "this that these those it its as be been has have had does do did "
           "module modules can could should would will shall may might".split())

def content_toks(s):
    """Content words of an answer: no stopwords, no pure punctuation."""
    return [t for t in re.findall(r"[a-z0-9_.]+", squash(s)) if t not in STOP]

def recall(pred, gold):
    """Fraction of the gold answer's content tokens present in the prediction.
    Verbose predictions are not penalised. This measures whether the model
    said what the gold answer says, not whether it said only that.
    """
    g = content_toks(gold)
    if not g: return None
    p = set(content_toks(pred))
    return sum(1 for t in g if t in p) / len(g)

def slot_match(pred, gold, module):
    """Template-slot accuracy: every number in the gold answer, plus the module
    name, appears in the prediction.
    Gold answers in this corpus are a small set of fixed templates with the
    module name and a few numbers substituted in. Slot match asks whether the
    model recovered those substituted values -- the part that carries the
    information -- and ignores the template wording, which exact match scores
    and which no model can guess without having been trained on the generator.
    """
    gn = numbers(gold)
    if not gn: return None
    pn = numbers(pred)
    if not all(x in pn for x in gn): return False
    if module and squash(module) not in squash(pred): return False
    return True

def wilson(k, n, z=1.96):
    if n == 0: return (0.0, 0.0)
    ph = k/n; d = 1 + z*z/n
    c = (ph + z*z/(2*n))/d
    h = z*math.sqrt(ph*(1-ph)/n + z*z/(4*n*n))/d
    return (100*max(0, c-h), 100*min(1, c+h))

# ------------------------------------------------------------------------- main
def score(pred_path, ref_path, model, condition, out_json=None, bitvec=None):
    ref = {}
    for line in open(ref_path):
        d = json.loads(line); ref[d["question_hash"]] = d
    sha = hashlib.sha256(open(ref_path,'rb').read()).hexdigest()

    preds, dup = {}, 0
    for ln, line in enumerate(open(pred_path), 1):
        line = line.strip()
        if not line: continue
        try: d = json.loads(line)
        except json.JSONDecodeError:
            sys.exit(f"FATAL {pred_path}:{ln}: not valid JSON")
        h = d.get("question_hash")
        if h is None: sys.exit(f"FATAL {pred_path}:{ln}: missing question_hash")
        if h in preds: dup += 1
        preds[h] = d.get("prediction", "")

    missing = set(ref) - set(preds)
    extra   = set(preds) - set(ref)
    print(f"reference : {ref_path}")
    print(f"            {len(ref)} samples, sha256 {sha[:16]}…")
    print(f"prediction: {pred_path}")
    print(f"            {len(preds)} unique hashes, {dup} duplicate lines")
    if extra:   print(f"  WARNING {len(extra)} predictions not in the reference (ignored)")
    if missing:
        print(f"  WARNING {len(missing)} reference samples have NO prediction.")
        print(f"          Scored as incorrect. Rerun those before reporting.")
    print()

    per = {c: collections.defaultdict(int) for c in CATS}
    agg = collections.defaultdict(float)
    correct = {}
    nall = 0
    for h, r in ref.items():
        cat, gold = r["category"], r["answer"]
        p = preds.get(h, "")
        per[cat]["n"] += 1; nall += 1
        ok = squash(p) == squash(gold)
        correct[h] = ok
        if ok:                              per[cat]["em_strict"] += 1
        if normalise(p) == normalise(gold): per[cat]["em_norm"]   += 1
        if partial(p, gold):                per[cat]["partial"]   += 1
        nm = num_match(p, gold)
        if nm is not None:
            per[cat]["num_n"] += 1
            if nm: per[cat]["num_ok"] += 1
        sm = slot_match(p, gold, r.get("module_name", ""))
        if sm is not None:
            per[cat]["slot_n"] += 1
            if sm: per[cat]["slot_ok"] += 1
        rc = recall(p, gold)
        if rc is not None:
            per[cat]["rec_n"] += 1
            per[cat]["rec_sum"] += rc
            if rc >= 0.8: per[cat]["rec80"] += 1
        agg["bleu"] += bleu4(p, gold); agg["rouge"] += rouge_l(p, gold); agg["f1"] += f1(p, gold)

    print("PRIMARY   slot = every gold number + the module name present in the prediction")
    print("          rec80 = >=80% of the gold answer's content words present")
    print("SECONDARY EM = byte-identical after normalisation; scores template wording,")
    print("          not content, and reads ~0 for any model not trained on the generator")
    print()
    print(f"{'category':22s} {'n':>4} {'slot':>14} {'rec80':>14} {'mean rec':>9} {'EM':>14}")
    macro_s = macro_n = 0.0; micro_s = micro_n = 0; present = 0
    for c in CATS:
        n = per[c]["n"]
        if not n: continue
        present += 1
        ks, kn = per[c]["em_strict"], per[c]["em_norm"]
        lo, hi = wilson(ks, n)
        sn, sk = per[c]["slot_n"], per[c]["slot_ok"]
        rn, r8 = per[c]["rec_n"],  per[c]["rec80"]
        mrec = per[c]["rec_sum"]/rn if rn else 0.0
        slo, shi = wilson(sk, sn) if sn else (0.0, 0.0)
        print(f"{c:22s} {n:>4} "
              f"{(f'{sk:>3}/{sn:<3} {100*sk/sn:5.1f}%' if sn else '      n/a     '):>14} "
              f"{(f'{r8:>3}/{rn:<3} {100*r8/rn:5.1f}%' if rn else '      n/a     '):>14} "
              f"{100*mrec:8.1f}% "
              f"{ks:>3}/{n:<3} {100*ks/n:5.1f}%"
              f"   slot95[{slo:.0f}, {shi:.0f}]")
        macro_s += 100*ks/n; macro_n += 100*kn/n
        micro_s += ks; micro_n += kn
        agg["slot_n"] += sn; agg["slot_ok"] += sk
        agg["rec_n"]  += rn; agg["rec80"]   += r8; agg["rec_sum"] += per[c]["rec_sum"]

    lo, hi = wilson(micro_s, nall)
    print()
    print(f"MACRO avg  EM strict {macro_s/present:5.2f}%   EM normalised {macro_n/present:5.2f}%")
    print(f"MICRO avg  EM strict {micro_s}/{nall} = {100*micro_s/nall:5.2f}%  [95% CI {lo:.1f}, {hi:.1f}]")
    print(f"           EM normalised {micro_n}/{nall} = {100*micro_n/nall:5.2f}%")
    SN, SK = int(agg["slot_n"]), int(agg["slot_ok"])
    RN, R8 = int(agg["rec_n"]),  int(agg["rec80"])
    if SN:
        lo2, hi2 = wilson(SK, SN)
        print(f"SLOT  micro {SK}/{SN} = {100*SK/SN:5.2f}%  [95% CI {lo2:.1f}, {hi2:.1f}]"
              f"   ({nall-SN} samples have no numeric slot)")
    if RN:
        print(f"REC80 micro {R8}/{RN} = {100*R8/RN:5.2f}%   mean content recall "
              f"{100*agg['rec_sum']/RN:5.2f}%")
    print(f"BLEU-4 {agg['bleu']/nall:5.2f}   ROUGE-L {agg['rouge']/nall:5.2f}   F1 {100*agg['f1']/nall:5.2f}")
    print()
    print("paste into results_template_per_category.csv:")
    cells = ",".join(f"{100*per[c]['em_strict']/per[c]['n']:.1f}" if per[c]['n'] else "" for c in CATS)
    print(f"  {model},{condition},{cells},{macro_s/present:.2f},{100*micro_s/nall:.2f}")

    if bitvec:
        # Per-sample correctness only. Contains no model output, no Verilog and
        # no question text -- one bit per evaluation sample, in sorted
        # question_hash order. Enough to recompute every number above and to run
        # McNemar between systems; not enough to reconstruct a prediction.
        order = sorted(ref)
        bits  = "".join("1" if correct.get(h) else "0" for h in order)
        osha  = hashlib.sha256("\n".join(order).encode()).hexdigest()
        with open(bitvec, "w") as f:
            f.write(f"# model        {model}\n")
            f.write(f"# condition    {condition}\n")
            f.write(f"# ref_sha256   {sha}\n")
            f.write(f"# order_sha256 {osha}\n")
            f.write(f"# n {len(order)}  correct {bits.count('1')}\n")
            f.write(bits + "\n")
        print(f"wrote {bitvec}  ({bits.count('1')}/{len(order)} correct)")

    if out_json:
        json.dump({"model":model,"condition":condition,"ref_sha256":sha,
                   "n":nall,"missing":len(missing),"duplicates":dup,
                   "per_category":{c:dict(per[c]) for c in CATS},
                   "macro_em_strict":macro_s/present,"micro_em_strict":100*micro_s/nall,
                   "macro_em_norm":macro_n/present,"micro_em_norm":100*micro_n/nall,
                   "bleu4":agg['bleu']/nall,"rouge_l":agg['rouge']/nall,"f1":100*agg['f1']/nall},
                  open(out_json,"w"), indent=2)
        print(f"\nwrote {out_json}")

def verify_csv(path):
    """Flag any cell that cannot be a count over the stated n."""
    import csv
    rows = list(csv.DictReader(open(path)))
    if not rows: sys.exit("empty csv")
    hdr = {}
    for c in CATS:
        k = next((k for k in rows[0] if k.startswith(c)), None)
        if k:
            m = re.search(r'n=(\d+)', k); hdr[c] = (k, int(m.group(1)) if m else None)
    bad = 0; checked = 0
    for r in rows:
        for c, (k, n) in hdr.items():
            v = (r.get(k) or "").strip()
            if not v or n is None: continue
            checked += 1
            kk = float(v)*n/100.0
            if abs(kk-round(kk)) > 1e-6:
                bad += 1
                print(f"  IMPOSSIBLE row {r.get('row','?'):>3} {r.get('model','?')[:26]:26s}"
                      f" {r.get('condition','?')} {c:22s} {v}% of n={n} -> {kk:.3f} correct")
    print(f"\nchecked {checked} cells, {bad} impossible.")
    print("A clean table reports 0. Anything else was not produced by score.py." if bad
          else "All cells are realisable as integer counts.")
    return 1 if bad else 0

if __name__ == "__main__":
    ap = argparse.ArgumentParser()
    ap.add_argument("--pred"); ap.add_argument("--ref", default="corpus/eval_sample.jsonl")
    ap.add_argument("--model", default="?"); ap.add_argument("--condition", default="?")
    ap.add_argument("--out-json"); ap.add_argument("--verify-csv")
    ap.add_argument("--bitvector", help="write per-sample 0/1 correctness (no model text)")
    a = ap.parse_args()
    if a.verify_csv: sys.exit(verify_csv(a.verify_csv))
    if not a.pred: ap.error("--pred is required (or use --verify-csv)")
    score(a.pred, a.ref, a.model, a.condition, a.out_json, a.bitvector)

	#!/usr/bin/env python3
	"""Shared scorer for CircuitQA.

	Takes raw model predictions and prints the metrics per category and overall. Pure standard library: no installs, so it
	runs identically on every lab machine.

	python score.py --pred preds/claude-opus-5.5.B.jsonl \
	--ref corpus/eval_sample.jsonl \
	--model "Claude Opus 5.5" --condition B

	Prediction file: one JSON object per line, {"question_hash": str, "prediction": str}.

	Every percentage is printed next to the integer count it came from (k/n), so any
	row in the final table can be audited. A percentage that is not k/n*100 for
	integer k did not come from this script.

	python score.py --verify-csv results_template_per_category.csv
	checks a filled-in table for exactly that.
	"""
	import argparse, json, math, re, sys, collections, hashlib, os

	CATS = ["structural_analysis","design_quality","functional_behavior",
	"code_understanding","optimization","cross_stage"]

	# ---------------------------------------------------------------- normalisation
	UNITS = [
	(r'\bnanoseconds?\b\|\bnsec\b', 'ns'), (r'\bpicoseconds?\b\|\bpsec\b', 'ps'),
	(r'\bmicroseconds?\b\|\busec\b', 'us'), (r'\bmilliseconds?\b\|\bmsec\b', 'ms'),
	(r'\bmilliwatts?\b', 'mw'), (r'\bmicrowatts?\b', 'uw'),
	(r'\bnanowatts?\b', 'nw'), (r'\bwatts?\b', 'w'),
	(r'\bmegahertz\b', 'mhz'), (r'\bgigahertz\b', 'ghz'), (r'\bkilohertz\b', 'khz'),
	(r'\bsquare micrometer?s?\b\|\bmicrometer\^?2\b\|\bum\^?2\b\|\bµm\^?2\b', 'um2'),
	(r'\bflip[- ]?flops?\b\|\bffs?\b', 'ff'),
	(r'\bregisters?\b', 'reg'), (r'\bwires?\b', 'wire'), (r'\bparameters?\b', 'param'),
	(r'\bgates?\b', 'gate'), (r'\bcells?\b', 'cell'),
	]

	def squash(s):
	"""Lowercase, strip punctuation noise, collapse whitespace."""
	s = (s or "").strip().lower()
	s = s.replace('−', '-').replace('µ', 'u')
	s = s.replace('²', '2').replace('³', '3') # um2 / um3 superscripts
	s = re.sub(r'[`*_"\']', '', s)
	s = re.sub(r'\s=\s', '=', s) # "a = b" -> "a=b"
	s = re.sub(r'\s,\s', ' ', s) # commas are not content
	s = re.sub(r'\band\b', '', s) # list conjunction
	s = re.sub(r'\s+', ' ', s)
	return s.strip(' .')

	def normalise(s):
	"""squash + canonical units + canonical number formatting.

	Makes '2.5ns', '2.50 ns' and '2.5 nanoseconds' identical, and 0.0020 == 0.002.
	"""
	s = squash(s)
	for pat, rep in UNITS:
	s = re.sub(pat, rep, s)
	s = re.sub(r'(\d)\s+([a-z]{1,4}\d?)\b', r'\1\2', s) # "2.5 ns" -> "2.5ns"
	s = re.sub(r'(?<![\w.])(\d+\.?\d*)', lambda m: fmtnum(m.group(1)), s)
	s = re.sub(r'\s+', ' ', s)
	return s.strip(' .')

	def fmtnum(t):
	try: v = float(t)
	except ValueError: return t
	if v == int(v) and abs(v) < 1e15: return str(int(v))
	return f"{v:.6f}".rstrip('0').rstrip('.')

	def numbers(s):
	return [float(x) for x in re.findall(r'-?\d+\.?\d*(?:[eE][-+]?\d+)?', s or "")]

	def toks(s):
	return re.findall(r'[a-z0-9_.\[\]:]+', normalise(s))

	# ---------------------------------------------------------------------- metrics
	def num_match(pred, ref, rtol=1e-3):
	"""Do the reference's numbers all appear in the prediction (within rtol)?"""
	rn, pn = numbers(ref), numbers(pred)
	if not rn: return None # no numeric content: not applicable
	for r in rn:
	if not any(abs(p - r) <= rtol * max(1.0, abs(r)) for p in pn): return False
	return True

	def f1(pred, ref):
	p, r = toks(pred), toks(ref)
	if not p or not r: return 0.0
	common = collections.Counter(p) & collections.Counter(r)
	ov = sum(common.values())
	if ov == 0: return 0.0
	prec, rec = ov/len(p), ov/len(r)
	return 2precrec/(prec+rec)

	def partial(pred, ref, thr=0.60):
	return f1(pred, ref) >= thr

	def bleu4(pred, ref):
	p, r = toks(pred), toks(ref)
	if not p: return 0.0
	logs, nonzero = 0.0, 0
	for n in (1,2,3,4):
	pg = collections.Counter(tuple(p[i:i+n]) for i in range(len(p)-n+1))
	rg = collections.Counter(tuple(r[i:i+n]) for i in range(len(r)-n+1))
	tot = sum(pg.values())
	if tot == 0: continue
	clip = sum(min(c, rg[g]) for g, c in pg.items())
	if clip == 0: return 0.0
	logs += math.log(clip/tot); nonzero += 1
	if nonzero == 0: return 0.0
	bp = 1.0 if len(p) > len(r) else math.exp(1 - len(r)/max(1, len(p)))
	return 100 * bp * math.exp(logs/nonzero)

	def rouge_l(pred, ref):
	p, r = toks(pred), toks(ref)
	if not p or not r: return 0.0
	dp = [[0]*(len(r)+1) for _ in range(len(p)+1)]
	for i in range(1, len(p)+1):
	for j in range(1, len(r)+1):
	dp[i][j] = dp[i-1][j-1]+1 if p[i-1]==r[j-1] else max(dp[i-1][j], dp[i][j-1])
	l = dp[-1][-1]
	if l == 0: return 0.0
	prec, rec = l/len(p), l/len(r)
	return 100 * 2precrec/(prec+rec)

	STOP = set("a an the is are was were of to in on at for with and or by from "
	"this that these those it its as be been has have had does do did "
	"module modules can could should would will shall may might".split())

	def content_toks(s):
	"""Content words of an answer: no stopwords, no pure punctuation."""
	return [t for t in re.findall(r"[a-z0-9_.]+", squash(s)) if t not in STOP]

	def recall(pred, gold):
	"""Fraction of the gold answer's content tokens present in the prediction.

	Verbose predictions are not penalised. This measures whether the model
	said what the gold answer says, not whether it said only that.
	"""
	g = content_toks(gold)
	if not g: return None
	p = set(content_toks(pred))
	return sum(1 for t in g if t in p) / len(g)

	def slot_match(pred, gold, module):
	"""Template-slot accuracy: every number in the gold answer, plus the module
	name, appears in the prediction.

	Gold answers in this corpus are a small set of fixed templates with the
	module name and a few numbers substituted in. Slot match asks whether the
	model recovered those substituted values -- the part that carries the
	information -- and ignores the template wording, which exact match scores
	and which no model can guess without having been trained on the generator.
	"""
	gn = numbers(gold)
	if not gn: return None
	pn = numbers(pred)
	if not all(x in pn for x in gn): return False
	if module and squash(module) not in squash(pred): return False
	return True

	def wilson(k, n, z=1.96):
	if n == 0: return (0.0, 0.0)
	ph = k/n; d = 1 + z*z/n
	c = (ph + zz/(2n))/d
	h = zmath.sqrt(ph(1-ph)/n + zz/(4n*n))/d
	return (100max(0, c-h), 100min(1, c+h))

	# ------------------------------------------------------------------------- main
	def score(pred_path, ref_path, model, condition, out_json=None, bitvec=None):
	ref = {}
	for line in open(ref_path):
	d = json.loads(line); ref[d["question_hash"]] = d
	sha = hashlib.sha256(open(ref_path,'rb').read()).hexdigest()

	preds, dup = {}, 0
	for ln, line in enumerate(open(pred_path), 1):
	line = line.strip()
	if not line: continue
	try: d = json.loads(line)
	except json.JSONDecodeError:
	sys.exit(f"FATAL {pred_path}:{ln}: not valid JSON")
	h = d.get("question_hash")
	if h is None: sys.exit(f"FATAL {pred_path}:{ln}: missing question_hash")
	if h in preds: dup += 1
	preds[h] = d.get("prediction", "")

	missing = set(ref) - set(preds)
	extra = set(preds) - set(ref)
	print(f"reference : {ref_path}")
	print(f" {len(ref)} samples, sha256 {sha[:16]}…")
	print(f"prediction: {pred_path}")
	print(f" {len(preds)} unique hashes, {dup} duplicate lines")
	if extra: print(f" WARNING {len(extra)} predictions not in the reference (ignored)")
	if missing:
	print(f" WARNING {len(missing)} reference samples have NO prediction.")
	print(f" Scored as incorrect. Rerun those before reporting.")
	print()

	per = {c: collections.defaultdict(int) for c in CATS}
	agg = collections.defaultdict(float)
	correct = {}
	nall = 0
	for h, r in ref.items():
	cat, gold = r["category"], r["answer"]
	p = preds.get(h, "")
	per[cat]["n"] += 1; nall += 1
	ok = squash(p) == squash(gold)
	correct[h] = ok
	if ok: per[cat]["em_strict"] += 1
	if normalise(p) == normalise(gold): per[cat]["em_norm"] += 1
	if partial(p, gold): per[cat]["partial"] += 1
	nm = num_match(p, gold)
	if nm is not None:
	per[cat]["num_n"] += 1
	if nm: per[cat]["num_ok"] += 1
	sm = slot_match(p, gold, r.get("module_name", ""))
	if sm is not None:
	per[cat]["slot_n"] += 1
	if sm: per[cat]["slot_ok"] += 1
	rc = recall(p, gold)
	if rc is not None:
	per[cat]["rec_n"] += 1
	per[cat]["rec_sum"] += rc
	if rc >= 0.8: per[cat]["rec80"] += 1
	agg["bleu"] += bleu4(p, gold); agg["rouge"] += rouge_l(p, gold); agg["f1"] += f1(p, gold)

	print("PRIMARY slot = every gold number + the module name present in the prediction")
	print(" rec80 = >=80% of the gold answer's content words present")
	print("SECONDARY EM = byte-identical after normalisation; scores template wording,")
	print(" not content, and reads ~0 for any model not trained on the generator")
	print()
	print(f"{'category':22s} {'n':>4} {'slot':>14} {'rec80':>14} {'mean rec':>9} {'EM':>14}")
	macro_s = macro_n = 0.0; micro_s = micro_n = 0; present = 0
	for c in CATS:
	n = per[c]["n"]
	if not n: continue
	present += 1
	ks, kn = per[c]["em_strict"], per[c]["em_norm"]
	lo, hi = wilson(ks, n)
	sn, sk = per[c]["slot_n"], per[c]["slot_ok"]
	rn, r8 = per[c]["rec_n"], per[c]["rec80"]
	mrec = per[c]["rec_sum"]/rn if rn else 0.0
	slo, shi = wilson(sk, sn) if sn else (0.0, 0.0)
	print(f"{c:22s} {n:>4} "
	f"{(f'{sk:>3}/{sn:<3} {100*sk/sn:5.1f}%' if sn else ' n/a '):>14} "
	f"{(f'{r8:>3}/{rn:<3} {100*r8/rn:5.1f}%' if rn else ' n/a '):>14} "
	f"{100*mrec:8.1f}% "
	f"{ks:>3}/{n:<3} {100*ks/n:5.1f}%"
	f" slot95[{slo:.0f}, {shi:.0f}]")
	macro_s += 100ks/n; macro_n += 100kn/n
	micro_s += ks; micro_n += kn
	agg["slot_n"] += sn; agg["slot_ok"] += sk
	agg["rec_n"] += rn; agg["rec80"] += r8; agg["rec_sum"] += per[c]["rec_sum"]

	lo, hi = wilson(micro_s, nall)
	print()
	print(f"MACRO avg EM strict {macro_s/present:5.2f}% EM normalised {macro_n/present:5.2f}%")
	print(f"MICRO avg EM strict {micro_s}/{nall} = {100*micro_s/nall:5.2f}% [95% CI {lo:.1f}, {hi:.1f}]")
	print(f" EM normalised {micro_n}/{nall} = {100*micro_n/nall:5.2f}%")
	SN, SK = int(agg["slot_n"]), int(agg["slot_ok"])
	RN, R8 = int(agg["rec_n"]), int(agg["rec80"])
	if SN:
	lo2, hi2 = wilson(SK, SN)
	print(f"SLOT micro {SK}/{SN} = {100*SK/SN:5.2f}% [95% CI {lo2:.1f}, {hi2:.1f}]"
	f" ({nall-SN} samples have no numeric slot)")
	if RN:
	print(f"REC80 micro {R8}/{RN} = {100*R8/RN:5.2f}% mean content recall "
	f"{100*agg['rec_sum']/RN:5.2f}%")
	print(f"BLEU-4 {agg['bleu']/nall:5.2f} ROUGE-L {agg['rouge']/nall:5.2f} F1 {100*agg['f1']/nall:5.2f}")
	print()
	print("paste into results_template_per_category.csv:")
	cells = ",".join(f"{100*per[c]['em_strict']/per[c]['n']:.1f}" if per[c]['n'] else "" for c in CATS)
	print(f" {model},{condition},{cells},{macro_s/present:.2f},{100*micro_s/nall:.2f}")

	if bitvec:
	# Per-sample correctness only. Contains no model output, no Verilog and
	# no question text -- one bit per evaluation sample, in sorted
	# question_hash order. Enough to recompute every number above and to run
	# McNemar between systems; not enough to reconstruct a prediction.
	order = sorted(ref)
	bits = "".join("1" if correct.get(h) else "0" for h in order)
	osha = hashlib.sha256("\n".join(order).encode()).hexdigest()
	with open(bitvec, "w") as f:
	f.write(f"# model {model}\n")
	f.write(f"# condition {condition}\n")
	f.write(f"# ref_sha256 {sha}\n")
	f.write(f"# order_sha256 {osha}\n")
	f.write(f"# n {len(order)} correct {bits.count('1')}\n")
	f.write(bits + "\n")
	print(f"wrote {bitvec} ({bits.count('1')}/{len(order)} correct)")

	if out_json:
	json.dump({"model":model,"condition":condition,"ref_sha256":sha,
	"n":nall,"missing":len(missing),"duplicates":dup,
	"per_category":{c:dict(per[c]) for c in CATS},
	"macro_em_strict":macro_s/present,"micro_em_strict":100*micro_s/nall,
	"macro_em_norm":macro_n/present,"micro_em_norm":100*micro_n/nall,
	"bleu4":agg['bleu']/nall,"rouge_l":agg['rouge']/nall,"f1":100*agg['f1']/nall},
	open(out_json,"w"), indent=2)
	print(f"\nwrote {out_json}")

	def verify_csv(path):
	"""Flag any cell that cannot be a count over the stated n."""
	import csv
	rows = list(csv.DictReader(open(path)))
	if not rows: sys.exit("empty csv")
	hdr = {}
	for c in CATS:
	k = next((k for k in rows[0] if k.startswith(c)), None)
	if k:
	m = re.search(r'n=(\d+)', k); hdr[c] = (k, int(m.group(1)) if m else None)
	bad = 0; checked = 0
	for r in rows:
	for c, (k, n) in hdr.items():
	v = (r.get(k) or "").strip()
	if not v or n is None: continue
	checked += 1
	kk = float(v)*n/100.0
	if abs(kk-round(kk)) > 1e-6:
	bad += 1
	print(f" IMPOSSIBLE row {r.get('row','?'):>3} {r.get('model','?')[:26]:26s}"
	f" {r.get('condition','?')} {c:22s} {v}% of n={n} -> {kk:.3f} correct")
	print(f"\nchecked {checked} cells, {bad} impossible.")
	print("A clean table reports 0. Anything else was not produced by score.py." if bad
	else "All cells are realisable as integer counts.")
	return 1 if bad else 0

	if __name__ == "__main__":
	ap = argparse.ArgumentParser()
	ap.add_argument("--pred"); ap.add_argument("--ref", default="corpus/eval_sample.jsonl")
	ap.add_argument("--model", default="?"); ap.add_argument("--condition", default="?")
	ap.add_argument("--out-json"); ap.add_argument("--verify-csv")
	ap.add_argument("--bitvector", help="write per-sample 0/1 correctness (no model text)")
	a = ap.parse_args()
	if a.verify_csv: sys.exit(verify_csv(a.verify_csv))
	if not a.pred: ap.error("--pred is required (or use --verify-csv)")
	score(a.pred, a.ref, a.model, a.condition, a.out_json, a.bitvector)