core/evals/public_demo/runner.py
Shay 7e1951915a
fix(ci): re-pin public_demo lane SHA after showcase content drift (#807)
Gate A1 and related cognition-path changes shifted the showcase
determinism fingerprint (86db25b6…) while the lane pin still reflected
the pre-drift artifact (e323adb3…). Re-pin to the CI-verified SHA
2895df08… and refresh v1_dev.json + CLAIMS.md.

Hermetic demo hygiene: use no_load_state for register-tour, learning-
loop, and multi-hop showcase adapters, and isolate CORE_ENGINE_STATE_DIR
in the public_demo lane runner so local engine_state/ pollution cannot
affect eval output or wall-clock.
2026-06-17 15:42:40 -07:00

251 lines
8.1 KiB
Python

"""Runner for evals/public_demo/ (ADR-0099).
Verifies the four ADR-0099 invariants in one pass:
- All claims supported on a single fresh run.
- Two runs produce byte-identical JSON (excluding ``total_runtime_ms``).
- Total runtime ≤ 30 seconds.
- Showcase imports only from already-shipped modules (no new mechanism).
"""
from __future__ import annotations
import argparse
import hashlib
import json
import re
import shutil
import sys
import tempfile
from pathlib import Path
from typing import Any
REPO_ROOT = Path(__file__).resolve().parent.parent.parent
SHOWCASE_PATH = REPO_ROOT / "core" / "demos" / "showcase.py"
# ADR-0099 §"Hard constraints" grep gate: showcase must compose only
# from these top-level packages plus its own sub-modules.
ALLOWED_IMPORT_PREFIXES: tuple[str, ...] = (
"core.",
"chat.",
"generate.",
"language_packs.",
"teaching.",
"evals.",
)
ALLOWED_STDLIB: tuple[str, ...] = (
"__future__",
"subprocess",
"time",
"pathlib",
"typing",
"dataclasses",
"html",
"re",
"hashlib",
"json",
"os",
"sys",
)
def _case_all_claims_supported(payload: dict[str, Any]) -> dict[str, Any]:
if not payload.get("all_claims_supported", False):
return _fail("all_claims_supported", "showcase reports all_claims_supported=False")
bad_scenes = [
s["scene_id"] for s in payload["scenes"] if not s["all_claims_supported"]
]
if bad_scenes:
return _fail("all_claims_supported", f"failing scenes: {bad_scenes}")
return _pass(
"all_claims_supported",
{"scene_count": len(payload["scenes"])},
)
_VOLATILE_KEYS: frozenset[str] = frozenset(
{
"total_runtime_ms", # wall-clock
"json_path", # adapter output paths (per-run temp dirs)
"transient_corpus", # learning-loop embeds its temp corpus path
# ``generated_at_revision`` advances every commit; the lane's
# invariant is "same code → same SHA," not "same HEAD → same
# SHA." Stripping this here keeps the pinned lane SHA stable
# across commits unless the underlying demos' content changes.
"generated_at_revision",
}
)
def _strip_volatile(node: Any) -> Any:
"""Recursively drop volatile keys so two runs can be compared.
The deterministic content sits in: claims, evidence (sha256 prefixes),
trace_features (each adapter's report_sha256), and the per-scene
statement. Wall-clock and absolute paths are excluded.
"""
if isinstance(node, dict):
return {
k: _strip_volatile(v)
for k, v in node.items()
if k not in _VOLATILE_KEYS
}
if isinstance(node, list):
return [_strip_volatile(v) for v in node]
return node
def _case_byte_equality(payload_a: dict, payload_b: dict) -> dict[str, Any]:
a_stripped = _strip_volatile(payload_a)
b_stripped = _strip_volatile(payload_b)
a = json.dumps(a_stripped, sort_keys=True, indent=2).encode()
b = json.dumps(b_stripped, sort_keys=True, indent=2).encode()
if a != b:
return _fail(
"determinism_run_to_run_byte_equality",
"showcase JSON bytes differ across two runs (after volatile-key strip)",
)
return _pass(
"determinism_run_to_run_byte_equality",
{"sha256": hashlib.sha256(a).hexdigest()},
)
def _case_runtime_under_budget(payload: dict[str, Any]) -> dict[str, Any]:
runtime_ms = payload.get("total_runtime_ms")
budget_ms = payload.get("max_runtime_seconds", 30) * 1000
if runtime_ms is None:
return _fail("runtime_under_budget", "payload missing total_runtime_ms")
if runtime_ms > budget_ms:
return _fail(
"runtime_under_budget",
f"{runtime_ms} ms > budget {budget_ms} ms",
)
# Don't include the exact runtime_ms in details — it varies per
# run and would break the lane report's byte-equality even at
# one-second bucket granularity for near-boundary runs. The case
# passing already proves runtime ≤ budget; the exact ms is in the
# showcase's own JSON for callers who want it.
return _pass(
"runtime_under_budget",
{"budget_seconds": budget_ms // 1000},
)
_IMPORT_RE = re.compile(r"^\s*(?:from\s+([\w\.]+)\s+import|import\s+([\w\.]+))")
def _case_pure_composition() -> dict[str, Any]:
"""Grep gate: showcase imports only from shipped packages + stdlib."""
sources = (
SHOWCASE_PATH,
REPO_ROOT / "core" / "demos" / "showcase_adapters.py",
REPO_ROOT / "core" / "demos" / "learning_loop_adapter.py",
)
forbidden: list[str] = []
for source in sources:
if not source.exists():
continue
for line in source.read_text(encoding="utf-8").splitlines():
match = _IMPORT_RE.match(line)
if not match:
continue
mod = (match.group(1) or match.group(2)).strip()
head = mod.split(".", 1)[0]
if mod.startswith(ALLOWED_IMPORT_PREFIXES):
continue
if head in ALLOWED_STDLIB:
continue
if mod in ALLOWED_STDLIB:
continue
forbidden.append(f"{source.relative_to(REPO_ROOT)}: {mod}")
if forbidden:
return _fail(
"pure_composition_no_new_mechanism",
f"forbidden imports: {forbidden}",
)
return _pass(
"pure_composition_no_new_mechanism",
{"sources_checked": len(sources)},
)
def _pass(case_id: str, details: dict[str, Any]) -> dict[str, Any]:
return {"case_id": case_id, "passed": True, "details": details, "divergence": None}
def _fail(case_id: str, divergence: str) -> dict[str, Any]:
return {"case_id": case_id, "passed": False, "details": {}, "divergence": divergence}
def run() -> dict[str, Any]:
import os
from core.demos.showcase import run_showcase
# Hermetic lane: demos must not load the developer's lived engine_state/.
os.environ.setdefault(
"CORE_ENGINE_STATE_DIR",
tempfile.mkdtemp(prefix="public_demo_engine_state_"),
)
tmp_root = Path(tempfile.mkdtemp(prefix="public_demo_lane_"))
try:
run_a = run_showcase(output_dir=tmp_root / "run_a")
run_b = run_showcase(output_dir=tmp_root / "run_b")
finally:
shutil.rmtree(tmp_root, ignore_errors=True)
cases = [
_case_all_claims_supported(run_a),
_case_byte_equality(run_a, run_b),
_case_runtime_under_budget(run_a),
_case_pure_composition(),
]
return {
"lane": "public_demo",
"lane_version": "v1",
"split": "dev",
"adr": "ADR-0099",
"invariants": [
"public_showcase_pure_composition",
"public_showcase_all_claims_supported",
"public_showcase_json_byte_equality",
],
"total_cases": len(cases),
"passed_cases": sum(1 for c in cases if c["passed"]),
"failed_cases": sum(1 for c in cases if not c["passed"]),
"all_passed": all(c["passed"] for c in cases),
"cases": cases,
}
def _canonical_json(payload: dict[str, Any]) -> bytes:
return json.dumps(payload, sort_keys=True, indent=2).encode("utf-8") + b"\n"
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description="public_demo lane runner")
parser.add_argument("--report", type=Path, default=None)
args = parser.parse_args(argv)
summary = run()
lane_dir = Path(__file__).resolve().parent
report_path = args.report or (lane_dir / "results" / "v1_dev.json")
report_path.parent.mkdir(parents=True, exist_ok=True)
payload_bytes = _canonical_json(summary)
report_path.write_bytes(payload_bytes)
print(f"report: {report_path}")
print(f"sha256: {hashlib.sha256(payload_bytes).hexdigest()}")
print(f"passed: {summary['passed_cases']}/{summary['total_cases']}")
if not summary["all_passed"]:
for c in summary["cases"]:
if not c["passed"]:
print(f" FAIL {c['case_id']}: {c['divergence']}")
return 0 if summary["all_passed"] else 1
if __name__ == "__main__":
sys.exit(main())