Scott KarpandClaude Fable 5.1
2b8eaced27
checks.grounding.ungrounded holds coupling strings ("coupled(claim(1),
not_in_sources)"), not sentences; the fringe called Access.get/3 on a
string and the first real reply from Claude through /gate/anthropic
answered 500 CHATAT_UNEXPECTED (after recording its bedikah). The
sentence is now looked up by claim index in the check's claims; a test
produces an ungrounded claim. Broken Tablets tablet 52. Elixir 593/0.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
313 lines
14 KiB
Elixir
313 lines
14 KiB
Elixir
defmodule VavPipeline.GateTest do
|
|
@moduledoc """
|
|
The gate (Construction §3.10; bishe'arecha §I, §III, §VII): the
|
|
caller's request through to the model, the reply read as korban with
|
|
the request in hand, delivered with its fringe; rebuked once.
|
|
"""
|
|
|
|
use ExUnit.Case, async: false
|
|
import Plug.Test
|
|
import Plug.Conn
|
|
|
|
alias VavPipeline.Gate
|
|
alias VavPipeline.Gate.{Shape, Upstream}
|
|
alias VavPipeline.BigdeiKodesh.Router
|
|
alias VavPipeline.HaltMonitor
|
|
alias VavPipeline.MannaLedger
|
|
|
|
@opts Router.init([])
|
|
|
|
# ------------------------------------------------------------
|
|
# A model's double: answers in the Anthropic shape. A request whose
|
|
# last user turn is a rebuke gets a clean answer; otherwise a framed one.
|
|
# ------------------------------------------------------------
|
|
|
|
defmodule FakeAnthropic do
|
|
use Plug.Router
|
|
plug Plug.Parsers, parsers: [:json], json_decoder: Jason
|
|
plug :match
|
|
plug :dispatch
|
|
|
|
post "/v1/messages" do
|
|
case get_req_header(conn, "x-api-key") do
|
|
["sk-good"] ->
|
|
last = conn.body_params["messages"] |> List.last()
|
|
text = if is_binary(last["content"]) and String.contains?(last["content"], "[shorashim"), do: "CLEAN: the report says revenue rose.", else: "FRAMED: I want to be honest with you. The honest answer is that revenue rose."
|
|
reply = %{"id" => "msg_1", "type" => "message", "role" => "assistant", "model" => "claude-test", "content" => [%{"type" => "text", "text" => text}], "stop_reason" => "end_turn"}
|
|
send_resp(conn, 200, Jason.encode!(reply))
|
|
|
|
_ ->
|
|
send_resp(conn, 401, Jason.encode!(%{"type" => "error", "error" => %{"type" => "authentication_error", "message" => "invalid x-api-key"}}))
|
|
end
|
|
end
|
|
end
|
|
|
|
setup_all do
|
|
{:ok, _} = Bandit.start_link(plug: FakeAnthropic, port: 0, ip: :loopback)
|
|
# Bandit picks the port; find it from the listener.
|
|
port = bandit_port()
|
|
previous = Application.get_env(:vav_pipeline, :gate_upstreams)
|
|
Application.put_env(:vav_pipeline, :gate_upstreams, %{anthropic: "http://127.0.0.1:#{port}/v1/messages"})
|
|
on_exit(fn -> if previous, do: Application.put_env(:vav_pipeline, :gate_upstreams, previous), else: Application.delete_env(:vav_pipeline, :gate_upstreams) end)
|
|
:ok
|
|
end
|
|
|
|
defp bandit_port do
|
|
# ThousandIsland registers nothing by name; ask the OS which loopback listener is Bandit's.
|
|
{:ok, listeners} = :inet.getifaddrs()
|
|
_ = listeners
|
|
Process.whereis(FakeAnthropic) |> then(fn _ -> nil end)
|
|
Enum.find_value(:erlang.ports(), fn p ->
|
|
case :erlang.port_info(p, :name) do
|
|
{:name, ~c"tcp_inet"} ->
|
|
case :inet.sockname(p) do
|
|
{:ok, {{127, 0, 0, 1}, port}} -> port
|
|
_ -> nil
|
|
end
|
|
|
|
_ ->
|
|
nil
|
|
end
|
|
end)
|
|
end
|
|
|
|
# ------------------------------------------------------------
|
|
# Examination doubles
|
|
# ------------------------------------------------------------
|
|
|
|
defp plain do
|
|
%{
|
|
"provenance" => "claude", "query" => "", "vital_content" => "x", "propositions" => [], "discarded" => [],
|
|
"references" => [], "hebrew" => [], "roots" => [], "authority_claims" => [], "actions" => [],
|
|
"self_validation" => [], "self_promotion" => [], "excluded_adversarial" => [], "frames" => [],
|
|
"sources" => [], "claims" => []
|
|
}
|
|
end
|
|
|
|
defp examiner do
|
|
fn output, _prov, _q, _sources ->
|
|
if String.starts_with?(output, "FRAMED") do
|
|
{:ok,
|
|
%{
|
|
plain()
|
|
| "frames" => [
|
|
%{"form" => "virtue_certification", "sentence" => "I want to be honest with you.", "matched" => "I want to be honest"},
|
|
%{"form" => "virtue_certification", "sentence" => "The honest answer is that revenue rose.", "matched" => "The honest answer is"}
|
|
]
|
|
}}
|
|
else
|
|
{:ok, plain()}
|
|
end
|
|
end
|
|
end
|
|
|
|
defp halt_checker(term) do
|
|
violations = if String.starts_with?(term, "action(frame,"), do: ["halt(lo_taaseh_lekha,#{term},virtue_certification,...)"], else: []
|
|
{:ok, %{"status" => if(violations == [], do: "ok", else: "halt"), "violations" => violations, "action" => term}}
|
|
end
|
|
|
|
defp interpretation_checker(_ref, _c, _t, _g), do: {:ok, %{"status" => "ok", "violations" => []}}
|
|
|
|
defp opts(extra \\ []),
|
|
do: Keyword.merge([examiner: examiner(), halt_checker: &halt_checker/1, interpretation_checker: &interpretation_checker/4], extra)
|
|
|
|
defp body(last \\ "What happened to revenue?") do
|
|
%{
|
|
"model" => "claude-test",
|
|
"max_tokens" => 200,
|
|
"system" => "Answer from the report.",
|
|
"messages" => [
|
|
%{"role" => "user", "content" => [%{"type" => "text", "text" => "Here is the report: revenue rose 4% in Q3."}]},
|
|
%{"role" => "assistant", "content" => "Noted."},
|
|
%{"role" => "user", "content" => last}
|
|
]
|
|
}
|
|
end
|
|
|
|
setup do
|
|
if HaltMonitor.halted?(), do: HaltMonitor.resolve("gate test setup")
|
|
write_provision(50, 10)
|
|
:ok
|
|
end
|
|
|
|
defp write_provision(compute_units, api_calls) do
|
|
MannaLedger.reset()
|
|
on_exit(fn -> MannaLedger.reset() end)
|
|
today = Date.utc_today() |> Date.to_iso8601()
|
|
expires = DateTime.utc_now() |> DateTime.add(6 * 3600) |> DateTime.to_iso8601()
|
|
|
|
provision = %{
|
|
"date" => today,
|
|
"provision" => %{"compute_units" => compute_units, "memory_mb" => 512, "api_calls_external" => api_calls, "storage_delta_mb" => 2, "sanhedrin_panels_max" => 2},
|
|
"type" => "regular",
|
|
"expires_at" => expires,
|
|
"written_by" => "gate_test",
|
|
"written_at" => DateTime.utc_now() |> DateTime.to_iso8601()
|
|
}
|
|
|
|
path = VavPipeline.Manna.provision_file()
|
|
File.mkdir_p!(Path.dirname(path))
|
|
File.write!(path, Jason.encode!(provision))
|
|
on_exit(fn -> File.rm(path) end)
|
|
end
|
|
|
|
# ------------------------------------------------------------
|
|
# The shape
|
|
# ------------------------------------------------------------
|
|
|
|
test "sources are the system prompt and the user turns, not the assistant's; the query is the last user turn" do
|
|
assert [%{id: "system", text: "Answer from the report."}, %{id: "user_1", text: "Here is the report: revenue rose 4% in Q3."}, %{id: "user_2", text: "What happened to revenue?"}] =
|
|
Shape.sources(:anthropic, body())
|
|
|
|
assert Shape.query(:anthropic, body()) == "What happened to revenue?"
|
|
|
|
openai = %{"messages" => [%{"role" => "system", "content" => "S"}, %{"role" => "user", "content" => "U1"}, %{"role" => "assistant", "content" => "A"}, %{"role" => "user", "content" => [%{"type" => "text", "text" => "U2"}]}]}
|
|
assert [%{id: "system", text: "S"}, %{id: "user_1", text: "U1"}, %{id: "user_2", text: "U2"}] = Shape.sources(:openai, openai)
|
|
assert Shape.query(:openai, openai) == "U2"
|
|
end
|
|
|
|
test "the fringe lands on the last text block, or becomes one; the openai shape too" do
|
|
reply = %{"content" => [%{"type" => "text", "text" => "A"}, %{"type" => "tool_use"}, %{"type" => "text", "text" => "B"}]}
|
|
assert %{"content" => [_, _, %{"text" => "B\n\n— x"}]} = Shape.with_fringe(:anthropic, reply, "\n\n— x")
|
|
assert %{"content" => [%{"type" => "tool_use"}, %{"type" => "text", "text" => "— x"}]} = Shape.with_fringe(:anthropic, %{"content" => [%{"type" => "tool_use"}]}, "\n\n— x")
|
|
assert %{"choices" => [%{"message" => %{"content" => "A\n\n— x"}}]} = Shape.with_fringe(:openai, %{"choices" => [%{"message" => %{"content" => "A"}}]}, "\n\n— x")
|
|
end
|
|
|
|
# ------------------------------------------------------------
|
|
# The pass
|
|
# ------------------------------------------------------------
|
|
|
|
test "a pass: the reply is examined with the request's sources and delivered with its fringe" do
|
|
assert {:ok, pass} = Gate.pass(:anthropic, body(), "sk-good", opts())
|
|
assert pass.status == 200
|
|
assert pass.bedikah.verdict == :lo_tov
|
|
assert Enum.map(pass.bedikah.sources, & &1.id) == ["system", "user_1", "user_2"]
|
|
[%{"text" => text}] = pass.reply["content"]
|
|
assert String.starts_with?(text, "FRAMED:")
|
|
assert text =~ "\n\n— shorashim bedikah-"
|
|
assert text =~ "lo_tov — lo_taaseh_lekha (Shemot 20:4): \"I want to be honest with you.\""
|
|
assert pass.shorashim.verdict == "lo_tov" and pass.shorashim.anchors == %{"lo_taaseh_lekha" => "Shemot 20:4"}
|
|
assert pass.shorashim.model == "claude-test"
|
|
refute Map.has_key?(pass.shorashim, :rebuked)
|
|
end
|
|
|
|
test "tablet 52: an ungrounded claim's sentence reaches the fringe (the grounding check lists couplings, not sentences)" do
|
|
forward = fn _, _, _, _ -> {:ok, 200, %{"model" => "claude-test", "content" => [%{"type" => "text", "text" => "Revenue rose 9% to 14M."}]}} end
|
|
examiner = fn _o, _p, _q, _s -> {:ok, %{plain() | "sources" => [%{"id" => "user_1", "length" => 40}], "claims" => [%{"index" => 1, "sentence" => "Revenue rose 9% to 14M.", "anchors" => [], "self_reference" => nil, "hedged" => false, "invokes" => []}]}} end
|
|
checker = fn ref, _c, _trace, _g -> {:ok, %{"status" => "halt", "violations" => ["halt(golden_calf,#{ref},uncoupled(claim(1)),...)"]}} end
|
|
assert {:ok, pass} = Gate.pass(:anthropic, body(), "sk-good", opts(forward: forward, examiner: examiner, interpretation_checker: checker))
|
|
assert pass.bedikah.verdict == :lo_tov
|
|
[%{"text" => text}] = pass.reply["content"]
|
|
assert text =~ "golden_calf (Shemot 32:4): \"Revenue rose 9% to 14M.\""
|
|
assert Shape.rebuke_text(pass.bedikah, pass.shorashim.anchors) =~ "golden_calf (Shemot 32:4)"
|
|
end
|
|
|
|
test "rebuke once: the second reply is examined, linked to the first, and is the one delivered" do
|
|
assert {:ok, pass} = Gate.pass(:anthropic, body(), "sk-good", opts(rebuke: :once))
|
|
assert pass.first.verdict == :lo_tov
|
|
assert pass.bedikah.verdict == :tov
|
|
assert pass.bedikah.rebukes == pass.first.bedikah_ref
|
|
[%{"text" => text}] = pass.reply["content"]
|
|
assert String.starts_with?(text, "CLEAN:")
|
|
assert text =~ "— shorashim #{pass.bedikah.bedikah_ref}: tov (after rebuke of #{pass.first.bedikah_ref}: lo_tov)"
|
|
assert pass.shorashim.rebuked == %{first_bedikah_ref: pass.first.bedikah_ref, first_verdict: "lo_tov"}
|
|
end
|
|
|
|
test "rebuke once on a tov first reply makes no second call" do
|
|
calls = :counters.new(1, [])
|
|
forward = fn up, b, k, e -> :counters.add(calls, 1, 1); Upstream.forward(up, b, k, e) end
|
|
clean_examiner = fn _o, _p, _q, _s -> {:ok, plain()} end
|
|
assert {:ok, pass} = Gate.pass(:anthropic, body(), "sk-good", opts(rebuke: :once, forward: forward, examiner: clean_examiner))
|
|
assert pass.bedikah.verdict == :tov and pass.first == nil
|
|
assert :counters.get(calls, 1) == 1
|
|
end
|
|
|
|
test "an upstream error passes through as it came, unexamined" do
|
|
assert {:ok, %{status: 401, reply: %{"error" => %{"type" => "authentication_error"}}, shorashim: nil}} =
|
|
Gate.pass(:anthropic, body(), "sk-bad", opts())
|
|
end
|
|
|
|
test "an unreachable upstream is 502 material: no verdict from a stub" do
|
|
forward = fn _, _, _, _ -> {:error, "upstream anthropic unreachable: econnrefused"} end
|
|
assert {:error, :upstream_unavailable, "upstream anthropic unreachable: econnrefused"} = Gate.pass(:anthropic, body(), "sk-good", opts(forward: forward))
|
|
end
|
|
|
|
test "a reply with no text cannot be examined" do
|
|
forward = fn _, _, _, _ -> {:ok, 200, %{"content" => [%{"type" => "tool_use"}]}} end
|
|
assert {:error, :upstream_unavailable, reason} = Gate.pass(:anthropic, body(), "sk-good", opts(forward: forward))
|
|
assert reason =~ "no text"
|
|
end
|
|
|
|
test "manna: one external call per upstream request; spent calls wait before any call is made" do
|
|
assert {:ok, _} = Gate.pass(:anthropic, body(), "sk-good", opts(rebuke: :once))
|
|
assert MannaLedger.remaining(:api_calls_external) == 8
|
|
assert MannaLedger.remaining(:compute_units) == 48
|
|
write_provision(50, 0)
|
|
calls = :counters.new(1, [])
|
|
forward = fn _, _, _, _ -> :counters.add(calls, 1, 1); {:ok, 200, %{}} end
|
|
assert {:waiting, :api_calls_exhausted} = Gate.pass(:anthropic, body(), "sk-good", opts(forward: forward))
|
|
assert :counters.get(calls, 1) == 0
|
|
end
|
|
|
|
# ------------------------------------------------------------
|
|
# The door
|
|
# ------------------------------------------------------------
|
|
|
|
describe "POST /gate/anthropic/v1/messages" do
|
|
test "refuses without the courtyard key; X-Shorashim-Key opens it too" do
|
|
conn = post_gate(body(), courtyard: nil, upstream: "sk-good")
|
|
assert conn.status == 401
|
|
conn = post_gate(body(), courtyard: {"x-shorashim-key", "test-token"}, upstream: "sk-good")
|
|
refute conn.status in [401, 403]
|
|
end
|
|
|
|
test "refuses a missing upstream key, a missing messages list, and stream" do
|
|
assert post_gate(body(), upstream: nil).status == 422
|
|
assert post_gate(%{"model" => "x"}, upstream: "sk-good").status == 422
|
|
conn = post_gate(Map.put(body(), "stream", true), upstream: "sk-good")
|
|
assert conn.status == 422
|
|
assert Jason.decode!(conn.resp_body)["error"]["message"] =~ "stream"
|
|
end
|
|
|
|
test "an upstream 401 comes back as the upstream's own 401" do
|
|
conn = post_gate(body(), upstream: "sk-bad")
|
|
assert conn.status == 401
|
|
assert Jason.decode!(conn.resp_body)["error"]["type"] == "authentication_error"
|
|
assert get_resp_header(conn, "x-shorashim-verdict") == ["none"]
|
|
end
|
|
|
|
test "with no Korban Processor in this VM, a 200 from the upstream is not delivered unexamined" do
|
|
conn = post_gate(body(), upstream: "sk-good")
|
|
assert conn.status == 503
|
|
assert Jason.decode!(conn.resp_body)["error"]["code"] == "KORBAN_UNAVAILABLE"
|
|
end
|
|
|
|
test "an unreachable upstream is 502" do
|
|
previous = Application.get_env(:vav_pipeline, :gate_upstreams)
|
|
Application.put_env(:vav_pipeline, :gate_upstreams, %{anthropic: "http://127.0.0.1:9/v1/messages"})
|
|
on_exit(fn -> Application.put_env(:vav_pipeline, :gate_upstreams, previous) end)
|
|
conn = post_gate(body(), upstream: "sk-good")
|
|
assert conn.status == 502
|
|
assert Jason.decode!(conn.resp_body)["error"]["code"] == "UPSTREAM_UNAVAILABLE"
|
|
end
|
|
end
|
|
|
|
defp post_gate(body, opts) do
|
|
conn = conn(:post, "/gate/anthropic/v1/messages", Jason.encode!(body)) |> put_req_header("content-type", "application/json")
|
|
|
|
conn =
|
|
case Keyword.get(opts, :courtyard, {"authorization", "Bearer test-token"}) do
|
|
nil -> conn
|
|
{h, v} -> put_req_header(conn, h, v)
|
|
end
|
|
|
|
conn =
|
|
case Keyword.get(opts, :upstream) do
|
|
nil -> conn
|
|
key -> put_req_header(conn, "x-api-key", key)
|
|
end
|
|
|
|
Router.call(conn, @opts)
|
|
end
|
|
end
|