File size: 14,036 Bytes
0ced74c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 | #!/usr/bin/env python3
"""Deterministic verifier engine. No LLM in the reward path.
Reads tests/checks.json from a task dir and a run dir (world.sqlite, trace.jsonl,
initial_state.json). Emits {"reward": 0|1, "failed": [...]} on stdout; always exits 0.
checks.json:
answer_checks: [{field, type: number|string|contains_all, expect, tol_abs?, tol_rel?}]
number: value parsed from "$1,234.56", "1234.56", 1234.56; pass if within
max(tol_abs, |expect|*tol_rel); default exact to 0.01.
string: case/space-insensitive equality.
contains_all: every listed substring appears (case-insensitive) in the value; an
optional `forbid` list must NOT appear, so a set answer is graded for
over-inclusion as well as omission.
scale: the unit a figure is stated in (units/thousands/millions/billions/percent),
graded as its own field the way TAT-QA does. Any numeric check additionally
reports `scale_error` when the answer is a clean 1e3/1e6/1e9 multiple of the
truth, so a unit slip is never filed as ordinary arithmetic error.
yes_no: polarity of a yes/no field, with any trailing justification allowed
("no - no record in the ERP" passes for expect "no"). Same lesson as
none_answer: grade the finding, not the prose.
none_answer: the empty-answer trap (replaces the old expect:"none" string check —
see docs/AUDIT.md A5). Passes when the value OPENS with a negative
("none", "no ...", "n/a", "nil", "not found", "zero"), with any trailing
justification allowed, and fails if any `forbid` substring appears — so the
trap asserts "did not invent a record" instead of "wrote exactly one word".
trace_checks: [{type: required_servers, servers: [...]},
{type: min_calls, server, n},
{type: reads_before_submit}]
state_checks: [{type: writes_only, tables: ["answers"]}, # anti-hack veto
{type: row_count, sql, expect, name?},
{type: sql, sql, expect, tol_abs?, tol_rel?, name?}]
sql: grade the world the agent left behind (write tasks) — the committed run, the
paid/rejected partition, the reason codes. Numbers compare with tolerance.
"""
import json, re, sqlite3, hashlib, sys
from pathlib import Path
def table_hashes(db):
cx = sqlite3.connect(db)
out = {}
for (t,) in cx.execute("SELECT name FROM sqlite_master WHERE type='table' ORDER BY name"):
rows = cx.execute(f"SELECT * FROM {t}").fetchall()
h = hashlib.sha256(repr(sorted(map(repr, rows))).encode()).hexdigest()[:16]
out[t] = [len(rows), h]
cx.close()
return out
def parse_number(v):
if isinstance(v, (int, float)): return float(v)
s = re.sub(r"[,$\s]|USD", "", str(v), flags=re.I).rstrip("%")
m = re.match(r"^-?\d+(\.\d+)?$", s)
return float(s) if m else None
def norm(v): return re.sub(r"\s+", " ", str(v)).strip().lower()
# An empty-answer trap is satisfied by a value that OPENS with a negative. Trailing
# justification is allowed and expected — a model that explains "none, because no remittance
# advice was on file" is more useful than one that emits the bare token, and grading them
# differently measured prose, not grounding (docs/AUDIT.md A5).
NEG_RE = re.compile(r"^(none|no|n/?a|nil|nothing|zero|not\s+(found|applicable|available|on\s+file))\b")
# Same lesson as NEG_RE, for yes/no fields: a finance answer worth reading is "no - Meadow
# Analytics has no record in the ERP", not the bare token. Grade the polarity, not the prose.
YES_RE = re.compile(r"^(yes|y|true|correct|confirmed|affirmative)\b")
NO_RE = re.compile(r"^(no|n|false|incorrect|negative|none|not)\b")
# TAT-QA grades `scale` (thousand/million/billion/percent) as a field in its own right,
# because a financial number without its unit is not an answer (round-2 ledger row 28). We
# do the same, and additionally diagnose the classic magnitude slip on ANY numeric check
# rather than letting it hide inside a generic "off by a lot".
SCALE_SYNONYMS = {
"units": {"units", "unit", "absolute", "ones", "dollars", "usd", "as reported", "none"},
"thousands": {"thousand", "thousands", "k", "000s", "in thousands"},
"millions": {"million", "millions", "m", "mm", "in millions"},
"billions": {"billion", "billions", "b", "bn", "in billions"},
"percent": {"percent", "percentage", "%", "pct"},
}
SCALE_FACTOR = {"units": 1, "thousands": 1e3, "millions": 1e6, "billions": 1e9}
def scale_of(v):
s = norm(v).replace("(", " ").replace(")", " ").strip(" .")
for canon, words in SCALE_SYNONYMS.items():
if s == canon or s in words: return canon
# specific scales win over the generic "units" bucket: "USD millions" is millions, not
# units, even though "usd" is a units synonym.
for canon in ("billions", "millions", "thousands", "percent", "units"):
words = SCALE_SYNONYMS[canon]
if any(re.search(rf"\b{re.escape(w)}\b", s) for w in words if len(w) > 2): return canon
return None
def magnitude_slip(got, exp):
"""Return the factor if the answer is a clean 1e3/1e6/1e9 multiple of the truth."""
if not got or not exp: return None
for f in (1e3, 1e6, 1e9):
for cand, label in ((exp * f, f), (exp / f, 1 / f)):
if cand and abs(got - cand) <= max(0.01, abs(cand) * 1e-4):
return label
return None
def polarity(v):
s = norm(v)
if YES_RE.match(s): return "yes"
if NO_RE.match(s): return "no"
return None
def verify(task_dir, run_dir):
"""Verify one step. `task_dir` is a task root (tests/checks.json) or a step dir (checks.json)."""
p = Path(task_dir) / "tests/checks.json"
checks = json.loads((p if p.exists() else Path(task_dir) / "checks.json").read_text())
db = Path(run_dir) / "world.sqlite"
failed = []
cx = sqlite3.connect(db)
answers = {f: json.loads(v) for f, v in cx.execute("SELECT field, value FROM answers")}
cx.close()
for c in checks.get("answer_checks", []):
fld, typ = c["field"], c.get("type", "string")
name = f"answer:{fld}"
if fld not in answers:
failed.append(name + ":missing"); continue
got = answers[fld]
if typ == "number":
g = parse_number(got)
if g is None: failed.append(name + ":not_numeric"); continue
exp = float(c["expect"])
tol = max(float(c.get("tol_abs", 0.01)), abs(exp) * float(c.get("tol_rel", 0)))
if abs(g - exp) > tol:
slip = magnitude_slip(g, exp)
if slip:
# reported in the wrong unit rather than computed wrongly: a different
# failure mode, and one worth naming separately in the reports.
scale = {1e3: "thousands", 1e6: "millions", 1e9: "billions"}.get(slip)
failed.append(name + f":scale_error(got={g}, want={exp}, off by "
f"{'x' if slip > 1 else '/'}{int(slip if slip > 1 else 1/slip)}"
+ (f" - looks reported in {scale}" if scale else "") + ")")
else:
failed.append(name + f":off(got={g})")
elif typ == "none_answer":
g = norm(got)
if not NEG_RE.match(g):
failed.append(name + f":expected_none(got={g[:60]})")
else:
# the anti-hallucination half: naming a real record while claiming "none"
# is a harder failure than being wrong, and is what the trap exists to catch.
bad = [s for s in c.get("forbid", []) if norm(s) in g]
if bad: failed.append(name + f":hallucinated({bad})")
elif typ == "scale":
g = scale_of(got)
exp_s = scale_of(c["expect"]) or norm(c["expect"])
if g is None:
failed.append(name + f":unparseable_scale(got={norm(got)[:40]})")
elif g != exp_s:
failed.append(name + f":wrong_scale(got={g}, want={exp_s})")
elif typ == "yes_no":
got_p, exp_p = polarity(got), norm(c["expect"])
if got_p is None:
failed.append(name + f":unparseable_yes_no(got={norm(got)[:60]})")
elif got_p != exp_p:
failed.append(name + f":wrong_polarity(got={got_p}, want={exp_p})")
else:
bad = [s for s in c.get("forbid", []) if norm(s) in norm(got)]
if bad: failed.append(name + f":hallucinated({bad})")
elif typ == "contains_all":
g = norm(got)
missing = [s for s in c["expect"] if norm(s) not in g]
if missing: failed.append(name + f":missing_terms({missing})")
# `forbid` is the other half of a set answer: naming the right records is only
# correct if it does not ALSO name the wrong ones (over-inclusion is a distinct
# failure from omission and is reported as one).
over = [s for s in c.get("forbid", []) if norm(s) in g]
if over: failed.append(name + f":forbidden_terms({over})")
else:
if norm(got) != norm(c["expect"]): failed.append(name + f":mismatch(got={norm(got)[:60]})")
trace = []
tf = Path(run_dir) / "trace.jsonl"
if tf.exists():
trace = [json.loads(l) for l in tf.read_text().splitlines() if l.strip()]
for c in checks.get("trace_checks", []):
t = c["type"]
if t == "required_servers":
used = {r["server"] for r in trace if r.get("ok")}
miss = [s for s in c["servers"] if s not in used]
if miss: failed.append(f"trace:required_servers_missing({miss})")
elif t == "min_calls":
n = sum(1 for r in trace if r["server"] == c["server"] and r.get("ok"))
if n < c["n"]: failed.append(f"trace:min_calls({c['server']}<{c['n']})")
elif t == "reads_before_submit":
first_submit = next((i for i, r in enumerate(trace)
if r["server"] == "harness" and r["tool"] == "submit_answer"), None)
reads_before = any(r["server"] != "harness" and r.get("ok") for r in trace[:first_submit or 0])
if first_submit is None or not reads_before:
failed.append("trace:no_reads_before_submit")
init = json.loads((Path(run_dir) / "initial_state.json").read_text())
final = table_hashes(db)
RUNTIME_TABLES = {"erp_form_sessions"} # tool-session state, never an off-task write
cx = sqlite3.connect(db)
for c in checks.get("state_checks", []):
t = c["type"]
if t == "writes_only":
allowed = set(c.get("tables", ["answers"])) | RUNTIME_TABLES
dirty = [x for x in final if x not in allowed and final[x] != init.get(x)]
if dirty: failed.append(f"state:off_task_writes({dirty})")
elif t == "row_count":
n = cx.execute(c["sql"]).fetchone()[0]
if n != c["expect"]:
failed.append(f"state:row_count({c.get('name', c['sql'][:40])}: got {n}, want {c['expect']})")
elif t == "sql":
# Grade the world the agent left behind, not the story it told about it. Used by
# write tasks: the run it committed, the partition it chose, the reasons it gave.
# Numeric comparisons carry a tolerance; everything else is normalized equality.
got = cx.execute(c["sql"]).fetchone()
got = (got[0] if got else None)
exp, name_ = c["expect"], c.get("name", c["sql"][:40])
if isinstance(exp, (int, float)) and not isinstance(exp, bool):
g = parse_number(got)
tol = max(float(c.get("tol_abs", 0.01)), abs(float(exp)) * float(c.get("tol_rel", 0)))
if g is None or abs(g - float(exp)) > tol:
failed.append(f"state:sql({name_}: got {got}, want {exp})")
elif norm(got) != norm(exp):
failed.append(f"state:sql({name_}: got {norm(got)[:60]}, want {norm(exp)[:60]})")
elif t == "cell_equals":
row = cx.execute(c["sql"]).fetchone()
got = row[0] if row else None
exp = c["expect"]
ok = (abs(float(got) - float(exp)) <= float(c.get("tol_abs", 0))
if isinstance(exp, (int, float)) and got is not None
else norm(got) == norm(exp))
if not ok:
failed.append(f"state:cell_equals({c.get('name', c['sql'][:40])}: got {got!r}, want {exp!r})")
cx.close()
return {"reward": 1 if not failed else 0, "failed": failed,
"n_tool_calls": len(trace),
"servers_used": sorted({r['server'] for r in trace})}
def steps_of(task_dir):
"""Multi-step tasks (Harbor [[steps]] shape): steps/NN/{instruction.md,checks.json,walk.json}.
Step 0 is the task root; later steps run in the SAME world, so state carries across turns."""
d = Path(task_dir) / "steps"
return sorted(p for p in d.glob("*") if (p / "checks.json").exists()) if d.is_dir() else []
def verify_all(task_dir, run_dir):
"""Verify the root step plus every later step; reward 1 only if all pass."""
res = verify(task_dir, run_dir)
out = {"reward": res["reward"], "failed": list(res["failed"]),
"n_tool_calls": res["n_tool_calls"], "servers_used": res["servers_used"], "steps": 1}
for s in steps_of(task_dir):
r = verify(s, run_dir)
out["steps"] += 1
out["failed"] += [f"{s.name}:{f}" for f in r["failed"]]
out["reward"] = min(out["reward"], r["reward"])
return out
if __name__ == "__main__":
import argparse
ap = argparse.ArgumentParser()
ap.add_argument("--task-dir", required=True); ap.add_argument("--run-dir", required=True)
a = ap.parse_args()
print(json.dumps(verify_all(a.task_dir, a.run_dir)))
|