The word "expert" in the previous status name implied raw-capability parity
with frontier LLMs on the same benchmark — which the gate does NOT verify.
What the gate actually verifies is CORE *claim-shape compliance*:
* signed digest (replay-reproducible from on-disk lane results)
* replay determinism (same inputs → byte-equal trace_hash)
* typed refusal (fabrication refused, not paraphrased)
* exact recall (no ANN, no cosine, no attention bottleneck)
* grounding-source provenance
These are claim shapes a transformer LLM cannot structurally produce
regardless of raw accuracy. A frontier LLM might score higher on the
same benchmark but cannot pass this contract.
Rename scope (semantics only, per ADR-0113):
status string "expert-demo" → "audit-passed"
predicate key predicates.expert_demo → predicates.audit_passed
reason key expert_demo_reason → audit_passed_reason
YAML key expert_demo_claims → audit_passed_claims
CLI command core demo expert → core demo audit-passed
output dir evals/expert_demos/ → evals/audit_passed/
artifact filenames expert_demo.{json,html} → audit_passed.{json,html}
HTML title CORE Expert-Demo: X → CORE Audit-Passed: X
Internal Python identifiers (module/file/function/class names like
`expert_demo.py`, `evaluate_expert_demo`, `ExpertDemoClaim`,
`expert_demo_claim_for`) are deliberately kept to minimize churn. ADR
file titles (ADR-0106..0112) preserved as historical record.
`expert` namespace reserved for ADR-0114+: an actual capability tier
above `audit-passed` backed by a public benchmark with a stated
threshold. ADR-0114 proposes the first such target — GSM8K-math —
laying out a falsifiable 7-phase arc (parser → solver → verifier →
stepped-realizer → eval lane → first `expert` ledger tier promotion).
Tests: 184 directly-affected tests green (140 capability/expert-demo
suite + 34 demo/audit-tour + 10 correction-cue). Smoke suite 67/67.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
151 lines
6 KiB
Python
151 lines
6 KiB
Python
"""ADR-0112 — Runnable Expert-Demo Showcase invariants.
|
|
|
|
Pins six load-bearing invariants:
|
|
|
|
1. Building an expert-demo for a promoted domain (`mathematics_logic`,
|
|
`physics`) succeeds and reports ``all_claims_supported=True``.
|
|
|
|
2. The recomputed digest matches the signed ``claim_digest`` byte-for-byte.
|
|
|
|
3. The composer refuses an unpromoted domain (no signed claim).
|
|
|
|
4. Output JSON is byte-identical across two consecutive runs (same
|
|
on-disk lane result files → same showcase bytes).
|
|
|
|
5. Per-lane shape-check verdicts are present for every attached lane on
|
|
both public and holdout splits.
|
|
|
|
6. The showcase does not mutate any source artifact (lane result files,
|
|
reviewers.yaml).
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import hashlib
|
|
import tempfile
|
|
from pathlib import Path
|
|
|
|
import pytest
|
|
|
|
from core.demos.expert_demo import (
|
|
SAMPLE_CASES_PER_SPLIT,
|
|
build_expert_demo,
|
|
run_expert_demo,
|
|
)
|
|
|
|
|
|
_REPO_ROOT = Path(__file__).resolve().parent.parent
|
|
|
|
|
|
_PROMOTED_DOMAINS = ("mathematics_logic", "physics")
|
|
|
|
|
|
@pytest.mark.parametrize("domain_id", _PROMOTED_DOMAINS)
|
|
class TestPromotedDomainsBuildSuccessfully:
|
|
def test_build_returns_all_claims_supported(self, domain_id: str) -> None:
|
|
payload = build_expert_demo(domain_id)
|
|
assert payload["all_claims_supported"] is True
|
|
assert payload["all_lanes_pass"] is True
|
|
assert payload["all_digests_match"] is True
|
|
|
|
def test_derived_digest_matches_signed(self, domain_id: str) -> None:
|
|
payload = build_expert_demo(domain_id)
|
|
dv = payload["digest_verification"]
|
|
assert dv["matches"] is True
|
|
assert dv["derived"] == dv["signed"]
|
|
# 64-char lowercase hex (SHA-256)
|
|
assert len(dv["signed"]) == 64
|
|
assert all(c in "0123456789abcdef" for c in dv["signed"])
|
|
|
|
def test_every_lane_split_has_shape_check(self, domain_id: str) -> None:
|
|
payload = build_expert_demo(domain_id)
|
|
assert len(payload["lanes"]) == 3, (
|
|
"every promoted domain in this slate attaches exactly three lanes"
|
|
)
|
|
for lane in payload["lanes"]:
|
|
assert set(lane["splits"].keys()) == {"public", "holdout"}
|
|
for split_name, split in lane["splits"].items():
|
|
assert split["shape_check"]["passed"] is True, (
|
|
f"{lane['lane_id']}/{split_name} failed shape check: "
|
|
f"{split['shape_check']}"
|
|
)
|
|
assert split["shape_check"]["shape"] is not None
|
|
|
|
def test_sample_cases_are_capped(self, domain_id: str) -> None:
|
|
payload = build_expert_demo(domain_id)
|
|
for lane in payload["lanes"]:
|
|
for split_name, split in lane["splits"].items():
|
|
assert len(split["sample_cases"]) <= SAMPLE_CASES_PER_SPLIT
|
|
assert len(split["sample_cases"]) >= 1, (
|
|
f"{lane['lane_id']}/{split_name} has zero sample cases"
|
|
)
|
|
|
|
|
|
class TestUnpromotedDomainRefused:
|
|
def test_unpromoted_domain_raises_value_error(self) -> None:
|
|
with pytest.raises(ValueError, match="No audit_passed_claims entry"):
|
|
build_expert_demo("systems_software")
|
|
|
|
def test_unknown_domain_raises_value_error(self) -> None:
|
|
with pytest.raises(ValueError, match="No audit_passed_claims entry"):
|
|
build_expert_demo("not_a_real_domain")
|
|
|
|
|
|
class TestByteDeterminism:
|
|
@pytest.mark.parametrize("domain_id", _PROMOTED_DOMAINS)
|
|
def test_two_runs_produce_byte_identical_json(self, domain_id: str) -> None:
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
out_a = Path(tmp) / "a"
|
|
out_b = Path(tmp) / "b"
|
|
run_expert_demo(domain_id=domain_id, output_dir=out_a)
|
|
run_expert_demo(domain_id=domain_id, output_dir=out_b)
|
|
bytes_a = (out_a / "audit_passed.json").read_bytes()
|
|
bytes_b = (out_b / "audit_passed.json").read_bytes()
|
|
assert bytes_a == bytes_b
|
|
sha_a = hashlib.sha256(bytes_a).hexdigest()
|
|
sha_b = hashlib.sha256(bytes_b).hexdigest()
|
|
assert sha_a == sha_b
|
|
|
|
|
|
class TestComposerIsReadOnly:
|
|
def test_run_does_not_mutate_reviewers_yaml(self) -> None:
|
|
path = _REPO_ROOT / "docs" / "reviewers.yaml"
|
|
before = path.read_bytes()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
run_expert_demo(domain_id="physics", output_dir=Path(tmp))
|
|
after = path.read_bytes()
|
|
assert before == after
|
|
|
|
def test_run_does_not_mutate_lane_result_files(self) -> None:
|
|
results_dir = _REPO_ROOT / "evals" / "foundational_physics_ood" / "results"
|
|
before = {p.name: p.read_bytes() for p in results_dir.glob("v1_*.json")}
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
run_expert_demo(domain_id="physics", output_dir=Path(tmp))
|
|
after = {p.name: p.read_bytes() for p in results_dir.glob("v1_*.json")}
|
|
assert before == after
|
|
|
|
|
|
class TestOutputArtifacts:
|
|
def test_json_and_html_both_written(self) -> None:
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
out = Path(tmp)
|
|
run_expert_demo(domain_id="physics", output_dir=out)
|
|
assert (out / "audit_passed.json").is_file()
|
|
assert (out / "audit_passed.html").is_file()
|
|
html = (out / "audit_passed.html").read_text(encoding="utf-8")
|
|
assert "<title>CORE Audit-Passed: physics</title>" in html
|
|
assert "a104cad136f3219df05dc7ce6a78437c02f7b5827cd3cdce568db3acda6a43ed" in html
|
|
|
|
def test_html_contains_per_lane_sections(self) -> None:
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
out = Path(tmp)
|
|
run_expert_demo(domain_id="physics", output_dir=out)
|
|
html = (out / "audit_passed.html").read_text(encoding="utf-8")
|
|
for lane_id in (
|
|
"foundational_physics_ood",
|
|
"inference_closure",
|
|
"fabrication_control",
|
|
):
|
|
assert lane_id in html
|
|
assert "public split" in html
|
|
assert "holdout split" in html
|