mirror of
https://github.com/norandom/OpenUltraSAST.git
synced 2026-10-16 19:59:26 +00:00
feat(model-grounded-detection): group 3, the evidence ladder and the judge
3.1 adds StaticFinding.rung beside evidence_level rather than renaming it. The existing evidence_level carries `static_corroboration`, emitted by the flat IR overlay -- a path Req 4.4 demotes to the suspicion band. Restating that as model_corroborated would claim a CPG arbitrated something it never saw, and rewriting every legacy finding to suspicion would be a product-visible regression nobody asked for. So `rung` defaults to suspicion and is raised only by ladder.at_rung(finding, verdict), which is the structural form of Req 5.4: nothing rises above suspicion without a Verdict behind it. The witness travels with the rung, because a rung without its evidence is just a louder assertion. It reaches markdown, JSON and SARIF. 3.2 is the judge, and it is defined as much by what it lacks. Three exhaustive outcomes: the model entails (report it, never call the model), the model contradicts (drop the claim, record why), the model cannot decide (report suspicion). The LLM gets one bounded typed question with tools=[] so it cannot search for the site, asked only about the residual the graph cannot settle -- whether a sanitizer on a REAL flow suffices. Its answer can only confirm a flow the CPG already found; it can never conjure one. Req 8.5 is asserted structurally: a test parses judge.py and fails on any AST name like k_runs/majority/vote/average/acceptance/floor/budget, any for/while loop, or an import from the deleted round machinery. That test first matched the word "votes" in the module's own docstring, which is why it now inspects code rather than text. Verified live against a real CPG, not only scripted rows: vuln rung=model_entailed request.args["name"] -> db.execute(...) fixed rung=suspicion no source reaches this sink model calls made: 0 Both sides arbitrated by the graph with the LLM never consulted -- the thing the prior architecture could not do at any price. 678 passing, mypy/ruff clean, gates byte-identical, zero orphaned modules. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
This commit is contained in:
Marius Ciepluch
co-authored by
Claude Opus 5
Claude Opus 5
parent
7387d2f190
commit
b8f8f727c6
@@ -0,0 +1,159 @@
|
||||
"""model-grounded-detection task 3.2: the LLM proposes, the model disposes (Req 8).
|
||||
|
||||
The judge is where the prior architecture's whole premise is inverted. There is no K-run averaging, no noise
|
||||
floor and no acceptance rule; there is one bounded, typed question per candidate, and the answer is CHECKED
|
||||
against the CPG rather than voted on. The three outcomes are exhaustive:
|
||||
|
||||
the model entails it -> report it, and never call the model at all
|
||||
the model contradicts it -> drop the claim, and record the contradiction
|
||||
the model cannot decide -> report `suspicion`, honestly
|
||||
|
||||
Every test here runs with no network and a scripted client.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
def _spec(): # type: ignore[no-untyped-def]
|
||||
from openultrasast.model.specs import TaintSpec
|
||||
|
||||
return TaintSpec(family="injection", language="python", sources=("request.args",),
|
||||
sinks=("os.system",), sanitizers=("escape",))
|
||||
|
||||
|
||||
def _cpg(rows): # type: ignore[no-untyped-def]
|
||||
from openultrasast.cpg.backend import CpgResult
|
||||
|
||||
return CpgResult(cpg_path=Path("cpg.bin"), run=lambda q, p: rows)
|
||||
|
||||
|
||||
ENTAILING_ROW = [{"sink": "os.system(cmd)", "sinkLine": "4", "sinkMethod": "run",
|
||||
"source": "request.args['c']", "sanitized": False, "length": 3}]
|
||||
|
||||
|
||||
class _Client:
|
||||
"""A scripted chat client that records whether it was called at all."""
|
||||
|
||||
def __init__(self, answer: str = '{"vulnerable": true, "family": "injection"}') -> None:
|
||||
self.calls: list[object] = []
|
||||
self._answer = answer
|
||||
|
||||
def complete(self, **kwargs: object): # type: ignore[no-untyped-def]
|
||||
from openultrasast.tool_hunter import ChatResponse
|
||||
|
||||
self.calls.append(kwargs)
|
||||
return ChatResponse(content=self._answer, tool_calls=[], reasoning=None)
|
||||
|
||||
|
||||
def test_an_entailed_candidate_costs_no_model_call() -> None:
|
||||
"""The point of the arbiter: where the graph decides, the LLM is not consulted at all."""
|
||||
from openultrasast.model.judge import judge
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
client = _Client()
|
||||
answer = judge(_cpg(ENTAILING_ROW), _spec(), function="run", client=client, model="m")
|
||||
assert answer.rung is Rung.ENTAILED
|
||||
assert client.calls == [], "an entailed candidate must not reach the model"
|
||||
|
||||
|
||||
def test_a_claim_the_model_contradicts_is_dropped_with_its_reason() -> None:
|
||||
"""Req 8.2. The LLM says vulnerable; the CPG shows no flow reaches the sink. The graph wins."""
|
||||
from openultrasast.model.judge import judge
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
client = _Client('{"vulnerable": true, "family": "injection"}')
|
||||
answer = judge(_cpg([]), _spec(), function="run", client=client, model="m")
|
||||
assert answer.rung is Rung.SUSPICION
|
||||
assert answer.contradiction, "a dropped claim must say why it was dropped"
|
||||
assert "no flow" in answer.contradiction.lower() or "not reach" in answer.contradiction.lower()
|
||||
|
||||
|
||||
def test_a_claim_the_model_supports_advances_to_corroborated() -> None:
|
||||
"""Req 8.3: a sanitized path is a real flow whose sufficiency only the LLM can judge."""
|
||||
from openultrasast.model.judge import judge
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
rows = [{"sink": "os.system(cmd)", "sinkLine": "4", "sinkMethod": "run",
|
||||
"source": "request.args['c']", "sanitized": True, "length": 3}]
|
||||
client = _Client('{"vulnerable": true, "family": "injection"}')
|
||||
answer = judge(_cpg(rows), _spec(), function="run", client=client, model="m")
|
||||
assert answer.rung is Rung.CORROBORATED
|
||||
assert len(client.calls) == 1, "exactly one bounded question, never a loop"
|
||||
|
||||
|
||||
def test_the_llm_declining_leaves_a_corroborated_flow_at_suspicion() -> None:
|
||||
"""The model saw a flow but the LLM says the sanitizer suffices; the claim does not advance."""
|
||||
from openultrasast.model.judge import judge
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
rows = [{"sink": "os.system(cmd)", "sinkLine": "4", "sinkMethod": "run",
|
||||
"source": "request.args['c']", "sanitized": True, "length": 3}]
|
||||
client = _Client('{"vulnerable": false, "family": "injection"}')
|
||||
answer = judge(_cpg(rows), _spec(), function="run", client=client, model="m")
|
||||
assert answer.rung is Rung.SUSPICION
|
||||
|
||||
|
||||
def test_without_a_client_the_verdict_still_stands_on_its_own() -> None:
|
||||
"""Req 11.3: the endpoint is optional. The model layer does not need it to entail."""
|
||||
from openultrasast.model.judge import judge
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
assert judge(_cpg(ENTAILING_ROW), _spec(), function="run", client=None, model="").rung is Rung.ENTAILED
|
||||
sanitized = [{"sink": "os.system(cmd)", "sinkLine": "4", "sinkMethod": "run",
|
||||
"source": "request.args['c']", "sanitized": True, "length": 3}]
|
||||
answer = judge(_cpg(sanitized), _spec(), function="run", client=None, model="")
|
||||
assert answer.rung is Rung.SUSPICION and "endpoint" in answer.contradiction.lower()
|
||||
|
||||
|
||||
def test_the_question_is_bounded_typed_and_offers_no_tools() -> None:
|
||||
"""Req 8.1: the LLM answers one question about a site it was handed; it never searches for the site."""
|
||||
from openultrasast.model.judge import judge
|
||||
|
||||
rows = [{"sink": "os.system(cmd)", "sinkLine": "4", "sinkMethod": "run",
|
||||
"source": "request.args['c']", "sanitized": True, "length": 3}]
|
||||
client = _Client()
|
||||
judge(_cpg(rows), _spec(), function="run", client=client, model="m")
|
||||
call = client.calls[0]
|
||||
assert call["tools"] == [], "no tools: the candidate is given, not searched for"
|
||||
assert call["json_object"] is True, "a typed answer, not prose"
|
||||
text = " ".join(str(m.get("content", "")) for m in call["messages"]) # type: ignore[union-attr]
|
||||
assert "os.system(cmd)" in text and "request.args['c']" in text, "the question carries the model's own evidence"
|
||||
|
||||
|
||||
def test_the_prompt_is_redacted() -> None:
|
||||
"""Req 11.4: every prompt passes through redaction."""
|
||||
from openultrasast.model.judge import judge
|
||||
|
||||
rows = [{"sink": "connect(pw)", "sinkLine": "4", "sinkMethod": "run",
|
||||
"source": "AKIA" + "IOSFODNN7" + "EXAMPLE", "sanitized": True, "length": 2}]
|
||||
client = _Client()
|
||||
judge(_cpg(rows), _spec(), function="run", client=client, model="m")
|
||||
text = " ".join(str(m.get("content", "")) for m in client.calls[0]["messages"]) # type: ignore[index,union-attr]
|
||||
assert "AKIAIOSFODNN7EXAMPLE" not in text
|
||||
|
||||
|
||||
def test_there_is_no_k_run_averaging_anywhere_in_the_judge() -> None:
|
||||
"""Req 8.5, asserted structurally: the noise architecture must not grow back."""
|
||||
import ast
|
||||
|
||||
source = Path("src/openultrasast/model/judge.py").read_text()
|
||||
tree = ast.parse(source)
|
||||
|
||||
# Check CODE, not prose: the docstring is allowed to name the machinery it replaced.
|
||||
names = {
|
||||
node.id for node in ast.walk(tree) if isinstance(node, ast.Name)
|
||||
} | {
|
||||
node.attr for node in ast.walk(tree) if isinstance(node, ast.Attribute)
|
||||
} | {
|
||||
node.name for node in ast.walk(tree) if isinstance(node, ast.FunctionDef)
|
||||
}
|
||||
for banned in ("k_runs", "majority", "vote", "votes", "average", "mean", "acceptance", "floor", "budget"):
|
||||
assert banned not in names, f"{banned} has no place in the judge"
|
||||
|
||||
# No loop over runs anywhere in the module.
|
||||
assert not [n for n in ast.walk(tree) if isinstance(n, ast.For | ast.While)], "the judge asks once"
|
||||
|
||||
imported = {n.module for n in ast.walk(tree) if isinstance(n, ast.ImportFrom) and n.module}
|
||||
assert not any("rounds" in m or "acceptance" in m for m in imported)
|
||||
@@ -0,0 +1,95 @@
|
||||
"""model-grounded-detection task 3.1: the evidence ladder is the verdict type (Req 5).
|
||||
|
||||
The rung says what ESTABLISHED a finding, so a deterministic model result is never confused with an LLM
|
||||
suspicion. The integrity constraint is Req 5.4: nothing reaches a rung above `suspicion` without a `Verdict`,
|
||||
and a `Verdict` above `suspicion` is only ever produced by the CPG with no model call in the establishing step.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
def _finding(**kw: object): # type: ignore[no-untyped-def]
|
||||
from openultrasast.findings import StaticFinding
|
||||
|
||||
base = dict(
|
||||
finding_id="r1:app.py:1", path="app.py", title="t", severity="high", confidence="medium",
|
||||
evidence_level="static_corroboration", rationale="why", line=1, function_name="run",
|
||||
reachability_status="unknown", reachability_evidence=[], reachability_conditions=[], tags=[],
|
||||
ranking_priority=1.0,
|
||||
)
|
||||
base.update(kw)
|
||||
return StaticFinding(**base) # type: ignore[arg-type]
|
||||
|
||||
|
||||
def test_the_rungs_are_ordered_and_comparable() -> None:
|
||||
from openultrasast.model.ladder import Rung, at_or_above
|
||||
|
||||
assert at_or_above(Rung.ENTAILED, Rung.CORROBORATED)
|
||||
assert at_or_above(Rung.CORROBORATED, Rung.SUSPICION)
|
||||
assert not at_or_above(Rung.SUSPICION, Rung.CORROBORATED)
|
||||
assert at_or_above(Rung.EXECUTION_CONFIRMED, Rung.ENTAILED)
|
||||
|
||||
|
||||
def test_a_finding_defaults_to_suspicion() -> None:
|
||||
"""Req 5.4: a claim nothing has arbitrated is a suspicion, whatever produced it."""
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
assert _finding().rung == Rung.SUSPICION
|
||||
|
||||
|
||||
def test_only_a_verdict_raises_a_finding_above_suspicion() -> None:
|
||||
from openultrasast.model.ladder import Rung, Verdict, at_rung
|
||||
|
||||
finding = _finding()
|
||||
verdict = Verdict(rung=Rung.ENTAILED, family="injection", witness="src -> sink")
|
||||
raised = at_rung(finding, verdict)
|
||||
assert raised.rung == Rung.ENTAILED
|
||||
assert "src -> sink" in raised.rationale, "the witness travels with the rung it justifies"
|
||||
assert raised.finding_id == finding.finding_id and raised.line == finding.line
|
||||
|
||||
|
||||
def test_a_none_verdict_leaves_the_finding_at_suspicion() -> None:
|
||||
from openultrasast.model.ladder import Rung, at_rung
|
||||
|
||||
assert at_rung(_finding(), None).rung == Rung.SUSPICION
|
||||
|
||||
|
||||
def test_the_legacy_evidence_level_is_left_alone() -> None:
|
||||
"""`static_corroboration` is the *flat IR overlay's* word and Req 4.4 demotes that path to the suspicion
|
||||
band. Restating it as `model_corroborated` would claim a CPG arbitrated something it never saw."""
|
||||
from openultrasast.model.ladder import Rung, Verdict, at_rung
|
||||
|
||||
raised = at_rung(_finding(), Verdict(rung=Rung.CORROBORATED, family="injection", witness="w"))
|
||||
assert raised.evidence_level == "static_corroboration", "the legacy field keeps its own meaning"
|
||||
assert raised.rung == Rung.CORROBORATED
|
||||
|
||||
|
||||
def test_the_rung_reaches_markdown_json_and_sarif(tmp_path: Path) -> None:
|
||||
"""Req 5.3: every output format carries the rung."""
|
||||
from openultrasast.findings import write_findings
|
||||
from openultrasast.model.ladder import Rung, Verdict, at_rung
|
||||
from openultrasast.reports import write_markdown_report, write_sarif_report
|
||||
|
||||
finding = at_rung(_finding(), Verdict(rung=Rung.ENTAILED, family="injection", witness="src -> sink"))
|
||||
md, js, sarif = tmp_path / "r.md", tmp_path / "r.json", tmp_path / "r.sarif"
|
||||
write_markdown_report([finding], md)
|
||||
write_findings([finding], js)
|
||||
write_sarif_report([finding], [], sarif)
|
||||
|
||||
assert "model_entailed" in md.read_text()
|
||||
assert json.loads(js.read_text())["findings"][0]["rung"] == "model_entailed"
|
||||
props = json.loads(sarif.read_text())["runs"][0]["results"][0]["properties"]
|
||||
assert props["rung"] == "model_entailed"
|
||||
|
||||
|
||||
def test_every_rung_round_trips_through_json(tmp_path: Path) -> None:
|
||||
from openultrasast.findings import write_findings
|
||||
from openultrasast.model.ladder import Rung
|
||||
|
||||
for rung in Rung:
|
||||
out = tmp_path / f"{rung.value}.json"
|
||||
write_findings([_finding(rung=rung)], out)
|
||||
assert json.loads(out.read_text())["findings"][0]["rung"] == rung.value
|
||||
Reference in New Issue
Block a user