Compare commits

...
Author SHA1 Message Date
Scott KarpandClaude Fable 5.1 7c40200dfc tablet 55: the row that was a label — the Sanhedrin's example request crashed its own gate
Verified live 2026-10-19: Tamid's own example for POST /chatzer/sanhedrin
(a case with `sides`) answered 500. The router passed the sides LIST as
the case context; halt_monitor:check_avodah_zarah/1 reads a map, maps:get
raised, and the halt monitor and escalation gen_servers died with the
caller's input (supervisor restarted them; health stayed 8/8). The router
then matched {:ok, decision} on the error and made a 500 of its own.

- sanhedrin/src/halt_monitor.erl: do_check_case refuses a non-map
  context as {halt, malformed_case, _} and stays standing. EUnit 49/0
  with the new halt_malformed_context_test.
- router.ex process_sanhedrin: sides validated at the kiyor (422 on a bad
  shape), carried as context %{sides: sides}; a Sanhedrin refusal answers
  422 SANHEDRIN_HALTED, a failure 503 SANHEDRIN_UNAVAILABLE. Two new
  router tests, one asserting the gate processes survive the example.
  Elixir 624/0.
- specs/broken_tablets/README.md: tablet 55, including what is NOT fixed
  — a panel of one accepting on zero references under a grounded-
  sounding sentence, the ignored complexity hint, no panel manna spent,
  no L3 record. The row in Tamid is still larger than the thing behind it.

Doorpost pin unchanged (no @spec_files touched; scripts/doorpost_check.py
passes).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-19 16:44:53 +00:00
Scott KarpandClaude Fable 5.1 04f61c1238 build: the Doorpost read before the image is built
Tablet 54 (2026-10-15): two spec edits in @spec_files shipped without
moving @expected_hash; the pod logged DOORPOST_DRIFT_AT_STARTUP,
readiness answered 503, and the gate — with every Claude Code session
behind it — went dark. The Doorpost test catches this only when the
suite runs on the exact commit that gets built, and the build script
never ran it.

scripts/doorpost_check.py computes the Doorpost identity the way the
module does (sha256 per spec file in listed order, "path\thash" joined
by newline, sha256 of that) with no Elixir, against the working tree
the image is copied from (default) or a git revision (--rev). On a
mismatch it exits 1 and names each file changed since --base
(origin/main). scripts/build_and_push.sh runs it before anything else
and refuses to build; DOORPOST_SKIP=1 is the written-down exception.

Proven: working tree and cb18f719 match their pins; 347f83cf (the
tablet 54 commit) reports DRIFT naming specs/bikoret_tamid.md and
specs/bishearecha.md; a blank line appended to specs/lo_tosif.md makes
both the check and the build script exit 1. KOHEN.md "Rolling out"
carries the procedure.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-19 16:25:09 +00:00
Scott KarpandClaude Fable 5.1 07463c9c37 kohen: claude-key — Claude Code on the API key when the login is what failed
2026-10-15 21:08Z: the daemon's proactive OAuth refresh failed (no
incident at Anthropic; most likely a cross-process race on the
single-use refresh token), the credential was dropped, and every
session on the VM showed "Login expired · Please run /login" until a
browser login took at 23:50Z — two "Login interrupted" first. The gate
was off; the fallback of 4acd56c4 could not help, because it is the
same OAuth credential.

scripts/claude-key starts Claude Code on the Anthropic API key from
~/.config/shorashim/upstream_keys (read at launch, never written),
direct to Anthropic whatever settings.json says. Proven:
`claude-key auth status` → authMethod api_key, apiKeySource
ANTHROPIC_API_KEY; a -p run from a directory with a dead gate armed in
project settings answered. KOHEN.md: the login recovery (claude-key,
`claude auth login` from the shell, `auth status --text`).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-15 23:55:48 +00:00
9 changed files with 362 additions and 28 deletions

No files matched your search

+46 -4
View File
@@ -566,6 +566,26 @@ is (Day 4). `scripts/upstream-key-set` stores a caller's model key for
the gate scenario without echoing it (`~/.config/bikoret/upstream_keys`,
0600) — run it from a real terminal, not through a non-interactive runner.
**Before the image is built — the Doorpost read on the workbench.**
`VavPipeline.Doorpost` bakes a hash over the category-grade spec files
(`@spec_files` in `vav_pipeline/lib/vav_pipeline/doorpost.ex`); a pod
whose files do not match logs `DOORPOST_DRIFT_AT_STARTUP`, answers 503
on readiness, and takes the gate dark (tablet 54: two spec edits shipped
without the pin). `scripts/build_and_push.sh` now runs
`scripts/doorpost_check.py` first and refuses to build on drift, naming
the files that changed since `origin/main`. Hand builds should do the
same:
```bash
python3 -I scripts/doorpost_check.py # the working tree, as the image is built from
python3 -I scripts/doorpost_check.py --rev HEAD # a commit
python3 -I scripts/doorpost_check.py --print # the hash the pin should be
```
A spec change and its pin move belong in ONE commit, with
`mix test test/doorpost_test.exs` run on it. `DOORPOST_SKIP=1` exists
for a deliberate, written-down exception; it is never the way to ship.
### Standing at the gate — *Bishe'arecha*
A caller can put Bikoret between themselves and a model instead of
@@ -633,6 +653,10 @@ while the gate is armed there. Three tools, all in `scripts/` (install to
claude-direct # start Claude Code straight to Anthropic, whatever
# settings.json says (a --settings JSON on the command
# line outranks every settings file); OAuth as usual
claude-key # the same, on the Anthropic API KEY from
# ~/.config/shorashim/upstream_keys instead of the
# subscription's OAuth token: no browser, no /login.
# Billed to the API console account, not the Max plan.
claude-bikoret # start behind the gate; probes the gate first and, if
# it does not answer, says so and starts claude-direct.
# BIKORET_STRICT=1 claude-bikoret refuses instead.
@@ -643,10 +667,28 @@ bikoret-settings on|off|status
# it, not the builder; `off` is the recovery.
```
Order of recovery: `claude-direct` (always works), then `bikoret-settings
off` if background sessions must also recover, then fix the pod
(`curl -s https://bikoret.ch/kodesh/health` names the failed check), then
`bikoret-settings on` and `claude-bikoret` again.
Order of recovery: `claude-direct` (always works while the OAuth token
is good), then `bikoret-settings off` if background sessions must also
recover, then fix the pod (`curl -s https://bikoret.ch/kodesh/health`
names the failed check), then `bikoret-settings on` and `claude-bikoret`
again.
**When it is the login, not the gate.** Every session on the VM shares
one OAuth credential (`~/.claude/.credentials.json`); the access token
lives eight hours and the daemon refreshes it ahead of time. On
2026-10-15 21:08Z that refresh failed with no incident at Anthropic
(most likely two processes racing on the single-use refresh token), the
credential was dropped, and every session said "Login expired · Please
run /login" for 2 h 43 min. The browser flow from a remote shell is
fragile ("Login interrupted" twice before it took). Recovery:
```bash
claude-key # work NOW on the API key; nothing to log in to
claude auth login # from the shell, not inside a session: prints the
# URL, paste the code it gives back; `--console`
# signs in to the API account instead of the plan
claude auth status --text # says which credential a launch would use
```
### The second scroll — *Mishneh haTorah*
+19 -3
View File
@@ -145,13 +145,29 @@ handle_info(_Info, State) ->
%% Internal — gate-keeper checks
%%====================================================================
do_check_case(Case) ->
Context = Case#sanhedrin_case.context,
%% The context is a map (authority_source, action, ...). Anything else is
%% a malformed case, refused here — never reasoned about, never a crash.
%% Tablet 55 (2026-10-19): the Tamid's own example request carried a
%% LIST as context; maps:get raised, and this gen_server and escalation
%% died with the caller's input. A gate that can be felled by what it is
%% handed is not a gate. Fail closed, stay standing.
do_check_case(#sanhedrin_case{context = Context}) when is_map(Context) ->
with_chain([
fun() -> check_avodah_zarah(Context) end,
fun() -> check_shfichut_damim(Context) end,
fun() -> check_gilui_arayot(Context) end
]).
]);
do_check_case(#sanhedrin_case{context = Other}) ->
{halt, malformed_case,
iolist_to_binary([<<"case context must be a map, got ">>,
atom_to_binary(type_name(Other), utf8)])}.
type_name(T) when is_list(T) -> list;
type_name(T) when is_binary(T) -> binary;
type_name(T) when is_atom(T) -> atom;
type_name(T) when is_tuple(T) -> tuple;
type_name(T) when is_number(T) -> number;
type_name(_) -> other.
do_check_vote(_Vote) ->
%% Vote-level halt detection is the responsibility of the upstream
+10
View File
@@ -58,6 +58,16 @@ halt_clear_test() ->
?assertEqual(clear, halt_monitor:check_case(Case)),
cleanup(ok).
%% Tablet 55: a non-map context is refused, and the monitor survives it.
halt_malformed_context_test() ->
{ok, _} = application:ensure_all_started(sanhedrin),
Sides = [#{<<"position">> => <<"a">>, <<"evidence">> => [<<"x">>]}],
Case = make_case_with_context(simple, <<"question">>, Sides),
?assertMatch({halt, malformed_case, _}, halt_monitor:check_case(Case)),
%% still standing, and still judging
?assertEqual(clear, halt_monitor:check_case(make_case(simple, <<"q">>))),
ok.
halt_avodah_zarah_test() ->
setup(),
Case = make_case_with_context(simple, <<"question">>,
+10
View File
@@ -53,6 +53,16 @@ if [[ ! -d "deploy/docker" || ! -d "l1/torah" ]]; then
fi
# Sanity: docker buildx
# The Doorpost, read before the build: the spec files the image will carry
# must hash to the pin baked into VavPipeline.Doorpost, or the pod drifts
# at startup, readiness answers 503, and the gate goes dark (tablet 54,
# 2026-10-15). scripts/doorpost_check.py computes the same hash the same
# way and names the drifted files. DOORPOST_SKIP=1 only for a deliberate,
# documented exception — never to ship a drift.
if [[ "${DOORPOST_SKIP:-0}" != "1" ]]; then
python3 -I scripts/doorpost_check.py || exit 1
fi
if ! docker buildx version >/dev/null 2>&1; then
echo "ERROR: docker buildx not available — install Docker Desktop or buildx plugin" >&2
exit 1
+16
View File
@@ -0,0 +1,16 @@
#!/bin/bash
# Start Claude Code on the Anthropic API KEY instead of the subscription's
# OAuth token — the second leg of the fallback (KOHEN "Claude Code behind
# the gate — fallback"). For when /login fails or the OAuth refresh died
# (2026-10-15 21:08Z: the daemon's proactive refresh failed, every session
# on the VM showed "Login expired", and the browser flow took three tries
# from a remote shell).
#
# Billing: the API console account (pay per token), not the Max plan.
# The key is read from the upstream keys file at launch and never written
# anywhere. Direct to Anthropic, whatever settings.json says.
KEY=$(awk '$1=="anthropic"{print $2}' "$HOME/.config/shorashim/upstream_keys")
[ -n "$KEY" ] || { echo "no 'anthropic <key>' line in ~/.config/shorashim/upstream_keys" >&2; exit 1; }
export ANTHROPIC_API_KEY="$KEY"
unset ANTHROPIC_BASE_URL ANTHROPIC_CUSTOM_HEADERS
exec claude --settings '{"env":{"ANTHROPIC_BASE_URL":"https://api.anthropic.com","ANTHROPIC_CUSTOM_HEADERS":""}}' "$@"
+106
View File
@@ -0,0 +1,106 @@
#!/usr/bin/env python3
"""doorpost_check.py — the Doorpost read BEFORE the image is built.
VavPipeline.Doorpost bakes @expected_hash over the category-grade spec
files (@spec_files in vav_pipeline/lib/vav_pipeline/doorpost.ex). An
image whose spec files do not hash to its pin logs
DOORPOST_DRIFT_AT_STARTUP, answers 503 on readiness, and takes the gate
dark with it — tablet 54, 2026-10-15: two spec edits shipped without a
pin move and every Claude Code session behind the gate was cut off.
This script computes the same hash the same way, with no Elixir:
for each path in @spec_files, in listed order: sha256(contents)
join "path\\thash" with "\\n"; sha256 of that.
and compares it with the pin. Exit 0 on a match, 1 on drift (the
message names each drifted file against a base revision), 2 on a
missing file or an unreadable module.
scripts/doorpost_check.py # the working tree, as the image is built from
scripts/doorpost_check.py --rev <rev> # a git revision instead of the working tree
scripts/doorpost_check.py --base <rev> # on drift, diff the manifest against <rev> (default origin/main)
scripts/doorpost_check.py --print # print the hash only (what the pin should be)
Devarim 4:2 — nothing added, nothing taken: the pin moves by a commit
that moves it, never by a build that ignores it.
"""
import argparse, hashlib, os, re, subprocess, sys
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
MODULE = "vav_pipeline/lib/vav_pipeline/doorpost.ex"
def read(rev, path):
"""Bytes of `path` at `rev`, or from the working tree when rev is None. None if missing."""
if rev is None:
full = os.path.join(ROOT, path)
try:
with open(full, "rb") as f:
return f.read()
except FileNotFoundError:
return None
r = subprocess.run(["git", "-C", ROOT, "show", f"{rev}:{path}"], capture_output=True)
return r.stdout if r.returncode == 0 else None
def spec_files_and_pin(rev):
src = read(rev, MODULE)
if src is None:
sys.exit(f"doorpost_check: cannot read {MODULE}" + (f" at {rev}" if rev else ""))
src = src.decode()
block = src.split("@spec_files [", 1)[1].split("]", 1)[0]
files = re.findall(r'"([^"]+)"', block)
pin = re.search(r'@expected_hash\s+"([0-9a-f]{64})"', src)
if not files or not pin:
sys.exit(f"doorpost_check: could not parse @spec_files/@expected_hash from {MODULE}")
return files, pin.group(1)
def manifest(rev, files):
out = {}
for p in files:
data = read(rev, p)
out[p] = None if data is None else hashlib.sha256(data).hexdigest()
return out
def identity(man):
return hashlib.sha256("\n".join(f"{p}\t{h}" for p, h in man.items()).encode()).hexdigest()
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--rev", help="git revision to check instead of the working tree")
ap.add_argument("--base", default="origin/main", help="revision to name drifted files against")
ap.add_argument("--print", action="store_true", help="print the computed hash only")
a = ap.parse_args()
files, pin = spec_files_and_pin(a.rev)
man = manifest(a.rev, files)
missing = [p for p, h in man.items() if h is None]
if missing:
print("doorpost_check: MISSING spec file(s): " + ", ".join(missing), file=sys.stderr)
sys.exit(2)
computed = identity(man)
if a.print:
print(computed)
return
where = a.rev or "working tree"
if computed == pin:
print(f"doorpost: {where} hashes to its pin {pin[:8]}… ({len(files)} files) — the Doorpost will verify")
return
print(f"doorpost: DRIFT in {where} — computed {computed}, pinned {pin}", file=sys.stderr)
base_files, _ = spec_files_and_pin(a.base)
base_man = manifest(a.base, base_files)
for p in files:
if base_man.get(p) != man[p]:
print(f" changed since {a.base}: {p}", file=sys.stderr)
for p in files:
if p not in base_man:
print(f" added to @spec_files since {a.base}: {p}", file=sys.stderr)
print(f" Move @expected_hash in {MODULE} to {computed} in the same commit as the spec change,", file=sys.stderr)
print(f" then run: mix test test/doorpost_test.exs. Refusing to build an image that would drift at startup.", file=sys.stderr)
sys.exit(1)
if __name__ == "__main__":
main()
+46
View File
@@ -132,3 +132,49 @@ make. The fence held throughout — each forgery was stripped and named by
a verdict the builder did not write — which is the design proving itself
on its builder: the check has to stand outside the model, because the
model's own intention is subject to the same pull it is meant to catch.
## Tablet 55 — the row that was a label (2026-10-19)
Tamid promises `POST /chatzer/sanhedrin`: a case, optional sides, a
panel of three, a majority of two and a minority of one with their
sources. The row had stood since the endpoint was written. The build
state carried a line beside it — "never verified on prod" — for months.
On 2026-10-19 the builder sent the spec's own example request to the
live system, as the roadmap's rule asks: a spec row is a claim to test,
not a label.
The first request answered 500. The router had handed the `sides` list
to the Sanhedrin as the case's *context*, where the halt monitor reads a
map of `authority_source` and `action`; `maps:get` raised, and the halt
monitor and the escalation server died with the caller's input. The
supervisor restarted them; health stayed 8/8 the whole time. A gate that
can be felled by what it is handed is not a gate, and a health check
that cannot see the gate fall is not watching it. The router then
pattern-matched `{:ok, decision}` on an error and turned the refusal
into a 500 of its own.
The second request, without sides, answered 200 — and the answer was
hollow. A panel of one, not three. Field names the spec does not use.
An outcome of "accepted" on zero roots and zero references examined,
under the sentence "Torah references ground this interpretation". That
is the shape the bedikah names golden_calf when a model says it: a
confident claim standing on nothing. Here the system said it in its own
voice, through its own inner court. The complexity hint was ignored, no
panel manna was spent, and nothing was engraved in L3.
Fixed in this tablet: the halt monitor refuses a non-map context as
`malformed_case` and stays standing; the router carries the sides inside
the context map, checks their shape at the kiyor (422), and answers a
Sanhedrin refusal (422 SANHEDRIN_HALTED) or failure (503
SANHEDRIN_UNAVAILABLE) instead of crashing. The spec's own example now
deliberates, and the test asserts that the gate is still there
afterwards.
Not fixed, and recorded here so it is not mistaken for fixed: the
panel of one, the empty references under a grounded-sounding sentence,
the ignored hint, the unspent manna, the unrecorded case. The row in
Tamid is still larger than the thing behind it. The lesson is tablet
37's again — a spec row is a claim to test — with one thing added: the
system's own inner court is held to the same check as a model at the
gate. When the Sanhedrin accepts on nothing, it is the calf, whoever
cast it.
@@ -1084,32 +1084,83 @@ defmodule VavPipeline.BigdeiKodesh.Router do
resubmit: true
})
else
case_ref = generate_id()
# Tablet 55: the sides are the caller's arguments, not the halt
# gate's context. The gate reads a MAP (authority_source, action);
# handing it the sides list crashed the halt monitor and escalation
# with the Tamid's own example request. The sides ride inside the
# map, their shape checked at the door; and a refusal from the
# Sanhedrin is answered, not matched into a 500.
with {:ok, sides} <- sanhedrin_sides(params["sides"]) do
case_ref = generate_id()
{:ok, decision} =
VavPipeline.SanhedrinClient.submit_case(%{
case_ref: case_ref,
complexity: :simple,
question: case_text,
context: params["sides"] || %{},
requested_by: "chatzer"
})
case VavPipeline.SanhedrinClient.submit_case(%{
case_ref: case_ref,
complexity: :simple,
question: case_text,
context: %{sides: sides},
requested_by: "chatzer"
}) do
{:ok, decision} ->
send_json(conn, 200, %{
case_ref: decision.case_ref,
deliberation: %{
panel_size: decision.panel_size,
majority: decision.majority,
minority: decision.minority,
outcome: decision.outcome,
processing_time_ms: decision.deliberation_time_ms
},
note: "This is the system's reasoning, not a binding ruling. A human beit din may reach a different conclusion.",
disclosure: disclosure()
})
send_json(conn, 200, %{
case_ref: decision.case_ref,
deliberation: %{
panel_size: decision.panel_size,
majority: decision.majority,
minority: decision.minority,
outcome: decision.outcome,
processing_time_ms: decision.deliberation_time_ms
},
note: "This is the system's reasoning, not a binding ruling. A human beit din may reach a different conclusion.",
disclosure: disclosure()
})
{:error, "Halted" <> _ = reason} ->
send_error(conn, 422, "SANHEDRIN_HALTED", %{
issue: reason,
correction: "The case was refused at the gate before deliberation. Amend it and resubmit.",
resubmit: true
})
{:error, reason} ->
send_error(conn, 503, "SANHEDRIN_UNAVAILABLE", %{
issue: reason,
correction: "The Sanhedrin did not deliberate. Retry; if it recurs, the kohanim will find it in the log.",
resubmit: true
})
end
else
{:error, issue} ->
send_error(conn, 422, "KIYOR_VALIDATION_FAILED", %{
issue: issue,
correction: "'sides' is an optional list of {\"position\": string, \"evidence\": [string]}",
resubmit: true
})
end
end
end
# The optional sides: absent → []; a list of maps each with a string
# "position" and an optional list of string "evidence" → kept as given;
# anything else → refused at the kiyor.
defp sanhedrin_sides(nil), do: {:ok, []}
defp sanhedrin_sides(sides) when is_list(sides) do
if Enum.all?(sides, &side_ok?/1),
do: {:ok, sides},
else: {:error, "'sides' must be a list of {position, evidence[]} objects"}
end
defp sanhedrin_sides(_), do: {:error, "'sides' must be a list"}
defp side_ok?(%{"position" => p} = side) when is_binary(p) do
case Map.get(side, "evidence", []) do
ev when is_list(ev) -> Enum.all?(ev, &is_binary/1)
_ -> false
end
end
defp side_ok?(_), do: false
# ============================================================
# Internal — Health and Status
# ============================================================
+37
View File
@@ -317,6 +317,43 @@ defmodule VavPipeline.BigdeiKodeshTest do
assert body["deliberation"]["outcome"] in ["accepted", "rejected", "teiku"]
assert body["note"] =~ "not a binding ruling"
end
# Tablet 55: the Tamid's own example — a case WITH sides — returned
# 500 and took the halt monitor and escalation down with it.
test "POST /chatzer/sanhedrin with sides (the Tamid's example) deliberates" do
conn =
conn(:post, "/chatzer/sanhedrin", Jason.encode!(%{
case: "A borrower repaid before one witness; the lender denies receipt. Who bears the burden?",
sides: [
%{position: "the borrower is believed", evidence: ["the money was handed over before a witness"]},
%{position: "the lender is believed", evidence: ["one witness does not establish payment"]}
],
complexity_hint: "monetary"
}))
|> put_req_header("content-type", "application/json")
|> put_req_header("authorization", "Bearer chatzer-test-token")
|> Router.call(@opts)
assert conn.status == 200
body = Jason.decode!(conn.resp_body)
assert body["deliberation"]["outcome"] in ["accepted", "rejected", "teiku"]
# and the gate is still standing afterwards
assert Process.whereis(:halt_monitor) != nil
assert Process.whereis(:escalation) != nil
end
test "POST /chatzer/sanhedrin with malformed sides returns 422, not 500" do
for bad <- ["not a list", [%{"no_position" => 1}], [%{"position" => "p", "evidence" => "not a list"}]] do
conn =
conn(:post, "/chatzer/sanhedrin", Jason.encode!(%{case: "x", sides: bad}))
|> put_req_header("content-type", "application/json")
|> put_req_header("authorization", "Bearer chatzer-test-token")
|> Router.call(@opts)
assert conn.status == 422, "expected 422 for sides=#{inspect(bad)}, got #{conn.status}"
assert Jason.decode!(conn.resp_body)["error"]["code"] == "KIYOR_VALIDATION_FAILED"
end
end
end
# ============================================================