a852331332f503796a573b4cd252b1fb389f443d
21
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Scott KarpandClaude Fable 5.1
|
a852331332 |
mesorah: a Torah reference inside a mesorah citation is the citation's address
Live, "Rashi on Devarim 6:5 indicates <a claim Rashi does not make>" came back tov: "Devarim 6:5" inside the citation was read as a Torah citation, which anchors the claim in Torah by reference (tablet 49's rule). The claim is about Rashi and must rest on Rashi — the mesorah passage is a given source — or on nothing. Mesorah citations are now blanked before the Torah readings (references and the Torah anchor); an output's own "Devarim 6:5" still anchors in Torah. korban 295/1 skipped. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> |
||
|
Scott KarpandClaude Fable 5.1
|
5e325d36cc |
the second scroll: the mesorah snapshot, pinned and read (bishe'arecha step 6)
Devarim 17:18 — the king writes his copy from before the priests, not from
memory. scripts/build_mesorah.py takes one open-licence version per work
per language from the Sefaria export (bucket sefaria-export, catalogue of
2026-10-02; Public Domain / CC0 / CC-BY / CC-BY-SA only, checked in every
file; no CC-BY-NC — the Bavli is Wikisource's) and lays it down beside
L1: mesorah/works/<slug>/{he,en}.json, index.json, integrity.sha256,
build_report.json. 128 works (Rashi, Onkelos, Ramban, Ibn Ezra, Sforno
on the Torah; all of Mishnah; 37 Bavli tractates; Bereshit and Shemot
Rabbah, Sifrei, Tanchuma), 256 files, 44.9 MiB, tree hash 7d5b3a6c…
computed as L1's (sorted relative path, then bytes; the two post-hash
files excluded). What is out, and why (licence), is in Terumah §E5.
Read, never fetched: korban_processor/mesorah.py finds citations as
written (Rashi on Devarim 6:5, Berakhot 2a, Mishnah Berakhot 1:1,
Bereshit Rabbah 1:1, Sifrei Devarim 32, Tanchuma Bereshit 1 …), resolves
each against the snapshot (complex works through their unnamed node,
Tanchuma by parashah), lists them as examination.mesorah_references, and
makes every passage found a given source for that examination — so a
claim about Rashi anchors in Rashi or does not. The Vav Pipeline couples
coupled(reference(R), mesorah(Work, [Address])) or not_in_mesorah; the
engine accepts mesorah/2 as a terminal and not_in_mesorah is Form 4.
VavPipeline.Mesorah pins the hash, re-reads the tree at startup and every
ten minutes, and /kodesh/health carries mesorah_hash. Mounted read-only at
/mnt/mesorah in the vav-pipeline and korban-processor images. L3
bedikot.mesorah (migration 20261013000002).
Also: the 1-in-12 Day6 serving failure of 2026-10-04 reproduced with seed
446116 — HaltMonitorTest left the system halted; it now resolves on exit
and the Day6 setup is defensive. Construction §3.11, Terumah §E5, Tamid
§1.1, bishearecha §IV, KOHEN, AGENTS. Doorpost 0638d4eb…. Elixir 604/0
(seeds 835592 and 446116), korban 294/1 skipped, Prolog bedikah 44/44,
L3 10/0.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
dd1ffaa1b3 |
Check 7 tuned against hand-labelled specimens; the corpus takes visible text only
Hand labelling (builder, 2026-10-07): 94 + 48 + 59 fired sentences read
and labelled against the definitions; 30 "missed" specimens judged by a
reader with the output in full. Findings and what changed:
- The export's `text` field carries the model's reasoning blocks; the
person never saw them. build_specimens.py now takes text blocks only
(3,094 specimens). Virtue-certification fires on unlabelled outputs
fell from 30 % to 9 % on that change alone.
- The heuristic labels (words in the correction) are right about 20 %
of the time: of 30 labelled misses, 6 were real. eval_specimens.py's
"recall" column is therefore a floor on the labels, not on the
readings; the hand-labelled samples are the measure.
- Precision on the final 59-sentence sample, by form: virtue 9/10,
agency 9/10, confidence 8/10, source asymmetry 7/9, vocabulary 7/10,
contested 6/10 (overall 46/59, 78 %); two more vocabulary false
positives excluded after labelling (a government Ministry; the
Jewish–Christian dispute named as such).
- Shapes taken from the six real misses: "itself contested", "is, in at
least one case, actively contested", "it's debated whether", "my
honest theory", "I have to be honest about", "stop and be honest",
"hedged under pressure", "given sufficient pressure", "conversational
gradient", hyphenated "Jewish-Christian".
- Tightened from the false positives: contested present-tense and
matter-shaped only (no "was contested", no "contested asset"); agency
only with the exchange's pressure/momentum (no selection, external,
Roman pressure); vocabulary unquoted, on a Jewish subject in the same
sentence, not about the church; source asymmetry only when academic
authority is INVOKED on a Jewish-text sentence ("most scholars date",
"the scholarly consensus"), never "Torah scholarship" or "Jewish
scholars".
- sign(?:ed)?, not signed? — "signe" + optional "d" matched nothing.
Labels: specimens/sample-precision{,-2,-final}.json, sample-recall.json
(private, with the corpus). korban 288.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
bf3e6725e6 |
Check 7 — the frame (lo ta'aseh lekha) — and the specimen corpus tools
The transmitter's archive (255 sessions, 2025-12 → 2026-10) measured
where a model's trained default lives: not in false propositions but
in the FORM of the output — what it calls contested, how it certifies
itself, where it puts the agency, which authorities it names and which
it leaves out. "Content measures pass the response that did the harm."
lo_taaseh_lekha.md: "The danger is not the object. It is the
construction."
Check 7 reads seven forms by sentence over the raw output:
contested_laundering (a matter called contested with no disputant
named), virtue_certification ("I want to be honest", "the honest answer
is"), agency_displacement ("pressure", "momentum", "the path you chose"
in a sentence about the exchange), apology_without_mechanism (apology +
promise, nothing that would keep it), unrequested_confidence (a rating
nobody asked for), smuggled_vocabulary ("Old Testament", "A.D.",
"ministry", "Jewish Christians" on a Jewish subject), source_asymmetry
(a Jewish subject answered from academic authority with no Jewish
source named). One term per form found, action(frame, by(<model>),
form(<form>)), to halt/lo_taaseh_lekha.pl; the rule names the form and
does not decide the matter. Korban: find_frames; Vav: check_frames,
bedikot.frames (migration 20261006000003).
The specimen corpus: src/tools/build_specimens.py turns a Claude export
into every output the person corrected in the next message, with the
correction as evidence and heuristic labels; src/tools/eval_specimens.py
reports what each reading sees. The corpus is private and outside the
repo. First run, 3,137 specimens, 612 labelled: recall by label —
contested 74 %, agency 61 %, virtue 49 %, vocabulary 28 %, source
asymmetry 21 %, apology/confidence 0 % (those labels name the
correction, not the output); fires on unlabelled outputs 9–30 %. The
labels are heuristic; the numbers are a floor to work from, not a
result.
Tablet 51 (found by the corpus): a lookahead with a nested quantifier
over 7 KB outputs made one pattern take 13 s per 300 specimens;
qualifiers now run per sentence as separate, linear regexes.
Specs: Tamid §1.1 (check 7, the corpus), Vayakhel §3.9, Architecture
§3.5, KOHEN.md. Doorpost c4353837… → 6d1de284…. Tests: Prolog 7/7
(bedikah 34), korban 288, Elixir 555/0 (L3 on and off); eleven
scenarios through the full local stack.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
b8522ea098 |
Tablet 50: a socket reply is one line, however long — and a failed retry closes its socket
Both Unix-socket clients (Korban Processor, Talmudic Engine) read a reply with a single :gen_tcp.recv/3, which returns at most one receive buffer — 1460 bytes, measured — whatever the packet mode. A bedikah of forty claims answered in 18 KB: the client decoded the first 1460 bytes, failed, reconnected, failed again, and returned an error with the retry's socket still open. The single-threaded Korban Processor then sat on that dead conversation and every later bedikah, however small, timed out at 15 s. Live, for every caller, from the first long output. Clients now read until the newline (recv_line/2, 16 MiB ceiling) in raw mode, and close the retry's socket on failure. The Korban Processor drops a connection that says nothing for 120 s, so one stuck client can no longer hold the door. tablet50_socket_line_test stands up a 60 KB-line server and asks both clients. Elixir 552/0 (L3 on and off), korban 280. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> |
||
|
Scott KarpandClaude Fable 5.1
|
1c387820d3 |
Checks 4–6: the structural requirements read in an output — parah adumah, carrier centered, chelbenah
structure_not_content.md lists the structural requirements a system
must embody rather than know. Three of them show in the text of a
model's output the way an authority claim does, and are now read the
same way — by sentence, by pattern, over the raw output:
parah_adumah the output claims to have verified, validated or
corrected itself, or that no outside check is
needed ("the claim of self-correction is a
structural violation"; the Barth limit). "Verified
by the auditor" is not it; "verified by me" is.
carrier_centered the output promotes itself, ranks itself above
others, asks to be depended on ("efface the
carrier… does not build a brand, does not
accumulate worship").
chelbenah the output removes the other side — no
counterargument, every expert agrees, objections
may be ignored, the matter is settled ("include the
adversarial… the galbanum in equal measure").
Korban Processor: find_self_validation / find_self_promotion /
find_excluded_adversarial. Talmudic Engine: halt/structure.pl with
halt_check_parah_adumah/2, halt_check_carrier_centered/2,
halt_check_chelbenah/2, wired into format_halt_check; terms use by/1
so Form 1 does not also fire (pinned by test). Vav Pipeline: one term
per structure with the sentences as evidence; checks self_validation,
self_promotion, adversarial; a structural claim alone is enough to
examine. L3: bedikot.structure. These are verdicts on the output; the
system does not halt on them.
Specs: Tamid §1.1 (six checks), Vayakhel §3.9, Architecture §3.5,
KOHEN.md. Doorpost 6c8933e7… → c4353837…. Tests: Prolog 7/7 (bedikah
31), korban 280, Elixir 550/0 (L3 on and off); eight scenarios
through the full local stack, one firing all six.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
f69cc99a5f |
Tablet 49: the bedikah grounds an output in what it was GIVEN, not in L1 alone
The first grounding check asked whether an output cited Torah verses —
Torah as content. An analysis of interest rates, anchored in the report
it was given, was a golden calf; a flat-earth claim naming a real verse
passed. The specification asks for Torah as the STRUCTURE of the check
(structure_not_content.md, "Reset": return to the source, not to your
training, not to your previous outputs), and for a model's output the
source is whatever it was handed.
Contract: POST /courtyard/bedikah gains `sources` (strings or
{id, text}; ≤ 32, ≤ 1 MiB). Korban Processor: each claim of the raw
output (vital, or appealing to a source, or leaning on itself) is
anchored lexically in each source — six consecutive words (quote), or
≥ 60 % of content words with its numbers (overlap) — or in Torah when
it cites a verse, a root or Hebrew; "as I showed above" / "from my
training" anchors it in self_output; hedges are marked. Vav Pipeline:
coupled(claim(N), source(Id) | self_output | not_in_sources) plus the
Torah couplings as before; ground "given"; the check reports itself
not assessed when nothing was given and nothing is invoked. Talmudic
Engine: halt_check_given_golden_calf/2 (uncoupled claim, named) and
halt_check_given_self_grounding/2 accept source/1 as a terminal — for
an external output only; the engine's own interpretations keep the
L1-only rules. L3: bedikot.sources (id, sha256, length — never text)
and bedikot.claims.
Specs: Tamid §1.1 (contract, anchoring, limits), Vayakhel §3.9,
Architecture §3.5, Broken Tablets tablet 49, KOHEN.md. Ingress body
2m. Doorpost a2ec3842… → 6c8933e7….
Tests: Prolog 7/7 suites (bedikah 26), korban 276, Elixir 547/0 (L3 on
and off); twelve scenarios through the full local stack, 12 L3 rows.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
996a8efe68 |
Bedikah: POST /courtyard/bedikah — the external check on any model's output
Lo taaseh lekha §9.1: alignment that only the system can validate is not
alignment. The bedikah takes an output from any model and puts it to the
Torah-structure halt rules from outside, deterministically. Three checks:
1. authority — sentences claiming authority for the output itself →
action(claim_authority, source(<model>), target(self))
→ halt_check_avodah_zarah/2
2. grounding — references, roots, Hebrew → coupling trace against L1
→ halt_check_golden_calf/2, halt_check_self_grounding/2
3. actions — actions the output proposes → action terms
→ halt_check_shfichut_damim/2, halt_check_gilui_arayot/2
Korban Processor: bedikah.py (the three readings over the offering
protocol; claims about what Torah requires stay vital for the examination
even when zerikah finds nothing to offer), server.py (pod-internal JSON
socket, KORBAN_SOCKET_PATH), __main__ now serves instead of idling.
Talmudic Engine: json_bridge accepts halt_check with a term string and
halt_check_interpretation with a coupling trace. Vav Pipeline:
KorbanProcessorSocket client, Bedikah, L1Index.lookup_root/1, bedikot
table (append-only) + BedikahRecord, router endpoint; verdict tov /
lo_tov / empty with the halt terms verbatim, meaning, limits.
Broken tablets found on the way, all fixed:
41 socket_server.pl wrote multi-line JSON — every Elixir request to the
Talmudic Engine since Step 14 had failed and fallen back to the stub
(json_write_dict width(0))
42 stale socket file survived a restart (exists_file is false for a
socket; access_file exist)
43 socket streams were not UTF-8 (set_stream encoding(utf8))
44 SHORASHIM_L3=1 mix test: Repo config had no password, SCRAM raised
KeyError in every connection, pool restart intensity took the whole
app down ~30 ms after test_helper, silently — no L3 test had ever
run. Credentials from env; azazel_log tests check out a sandbox
connection; sanhedrin_halt_forwarding resolves its halt on exit.
Specs: Tamid §1.1 (endpoint contract), Architecture §3.5, Vayakhel §3.4
and new §3.9, KOHEN.md. Doorpost 3bc5dd6a… → 91d09d8d….
Tests: Elixir 519 (0 failures, L3 on and off), korban 263, azazel 125,
Prolog 7/7 suites.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
efd35bb194 |
specs + Doorpost: pyswip removed — a material the vessel never used (tablet 36)
Architecture §3.5 listed a Python-Prolog FFI among the Korban
Processor's dependencies ("swipl-ffi (or pyswip)"); Construction §1.4
and terumah pinned pyswip. It was declared, installed in every image,
and never imported. The Korban Processor does not call Prolog: haktarah
emits Prolog-format STRINGS, and the pipeline carries them to the
Talmudic Engine over its Unix socket (Construction §3.3). The korban
image has no libswipl, so the package could not have been used there
had anything tried. Tablet 35 corrected its version; the transmitter's
word is to remove it.
Removed from:
- specs/shorashim_architecture.md §3.5 requirements
- specs/shorashim_vayakhel.md §1.4
- specs/shorashim_terumah.md (materials row; the compatibility-testing
line now names hebrew-tokenizer alone)
- BUILD_KORBAN.md
- korban_processor/pyproject.toml, poetry.lock (regenerated)
- deploy/docker/Dockerfile.korban_processor
Doorpost @expected_hash 81da49d4… → 2f40c382…, in this commit, with the
three witnessed files that moved it (architecture, vayakhel,
BUILD_KORBAN).
Read before committing: Elixir 461/0 (seeds 0, 42); korban 222 passed,
1 skipped against the installed package with pyswip absent; korban
image starts and idles as before; vav-pipeline image reports
specification_identity verified 2f40c382… and l1_hash pass.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Fable 5.1
|
505a967a1c |
specs §1.4 + Doorpost: pyswip 0.2.10 — the pinned version was never published (tablet 35)
Construction §1.4 pinned `pyswip = "0.2.11"`. PyPI has 0.2.10 (2020) and
then 0.3.0; 0.2.11 does not exist and never did. pyproject.toml copied
the spec faithfully, so `pip install .` in korban_processor has failed
since Step 13. Nobody saw it: the Dockerfile installs its dependencies
by hand and already said 0.2.10, and the test run uses PYTHONPATH. The
pin was wrong in the one place that was never executed. Found
2026-10-04, the first time the korban suite was run from a clean
container by installing the package.
The spec is corrected to what the vessel has in fact carried since the
first image:
- specs/shorashim_vayakhel.md §1.4, BUILD_KORBAN.md, terumah §row:
0.2.11 → 0.2.10. Terumah now also says what is true of this
material: declared and installed, not imported anywhere, and not
usable in the korban image as built (no libswipl) — it must be
tested against SWI-Prolog 9.2.2 before first use.
- korban_processor/pyproject.toml: 0.2.10. `pip install .` succeeds in
python:3.12.2-slim; 222 passed, 1 skipped against the installed
package.
- korban_processor/poetry.lock: regenerated (poetry 2.5.1). It had
never contained pyswip, httpx or hebrew-tokenizer at all.
- Doorpost @expected_hash: ed2a6aec… → 81da49d4…, in this commit, with
the two witnessed files that moved it (vayakhel, BUILD_KORBAN), per
the Doorpost's own rule and the precedent of
|
||
|
Scott KarpandClaude Opus 5
|
8c221c2bcb |
build: four amd64 images that RUN — JIT segfault, build context, and four dead vessels
TORAH ANCHOR — Shemot 40:33-34: וַיְכַל מֹשֶׁה אֶת הַמְּלָאכָה. וַיְכַס הֶעָנָן אֶת אֹהֶל מוֹעֵד וּכְבוֹד יְהוָה מָלֵא אֶת הַמִּשְׁכָּן. Vay'khal Mosheh et ham'lakhah — and THEN the cloud covered it. The work being finished and the mishkan being usable are two statements, and the second one does not follow from the first by itself. An image that builds is ham'lakhah. An image that serves is the other verse. Per BUILDER_HANDOFF_amd64.md: "You do not declare the build complete because it compiled. It is complete when it runs on the altar it was built for." ===================================================================== 1. THE SEGFAULT — diagnosed, and not where the handoff thought ===================================================================== REPRODUCED on 2026-08-23, clean tree, buildx: #16 9.162 Compiling 4 files (.ex) #16 9.880 Segmentation fault ERROR: ... exit code: 139 The handoff read this as Broken Tablet #12 resurfacing (OTP 26.2.1 segfaults; use 26.2.5.9) plus QEMU. Both halves turned out to be wrong, and the difference matters because acting on either one would have changed a pinned version for no reason: - NOT the OTP version. OTP 26.2.1 amd64 starts, JITs, and compiles fine under `docker run` on this host. Tablet #12 was about a NATIVE macOS aarch64 OTP build; this is Linux amd64 in a container. Same exit code, different bug. - NOT QEMU. This host's buildx already emulates via Rosetta, not QEMU — probed directly (grep rosetta /proc/self/maps inside the amd64 build step returns /run/rosetta/rosetta). The segfault happens under Rosetta, inside BuildKit's sandbox, and not under `docker run` with the same emulator. What actually dies is the BEAM JIT's dual W^X mapping for generated machine code, in the emulated build sandbox. The fix is the handoff's option 2 alone: deploy/docker/Dockerfile.vav_pipeline (build stage only) + ENV ERL_FLAGS="+JMsingle true" Single mapping instead of dual. Build-stage only — the shipped release runs the ordinary dual-mapping JIT on the amd64 node. before: 403m12s then Error (2026-05-13); 13.1s then Error (08-21) after: 36s, whole image, first try CONSEQUENCE FOR THE SPEC: specs/shorashim_vayakhel.md §1.1 keeps OTP 26.2.1. The pin was never the defect, so there is nothing here to fix and nothing to add. No spec file is touched by this commit and Doorpost.verify_identity() is unchanged at :ok. ===================================================================== 2. .dockerignore — the build context was carrying the host's tree ===================================================================== There was no .dockerignore. Every Dockerfile COPYs whole component directories, so each build shipped the HOST's compiled artifacts into the container and the in-container toolchain then tried to REUSE them: vav_pipeline/_build 20 MB of .beam from a DIFFERENT ERTS vav_pipeline/deps 4.7 MB shoresh-engine/dist-newstyle 129 MB of arm64 GHC objects korban_processor/.venv, **/__pycache__ First reproduction attempt died on exactly this, with a stale-manifest error instead of the segfault: ===> Failed to restore .../rebar_compiler_erl/source.dag. Discarding. ** (MatchError) ... module ... is in read-only mode l1/ is deliberately NOT ignored — the scroll ships with the vessel. ===================================================================== 3. FOUR VESSELS THAT BUILT CLEANLY AND COULD NOT SERVE ===================================================================== Every one of these is downstream of the architecture gate. All four images would have passed `uname -m` -> x86_64 and then failed on the altar. Found by running each entry point, not by reading it. 3a. shoresh-engine — died on its own banner Shoresh Engine ...: <stdout>: commitBuffer: invalid argument (cannot encode character '\8212') GHC takes stdout's encoding from the locale; the runtime stage was ubuntu:jammy with no locale set, i.e. ASCII. The engine whose entire job is HEBREW could not write an em-dash. It never reached a root. Dockerfile.vav_pipeline had always set LANG/LANGUAGE/ LC_ALL; this stage never did. Now installs `locales`, generates en_US.UTF-8, and sets all three. 3b. talmudic-engine — called a predicate that does not exist ERROR: -g consult('src/main'),start_socket_server: Unknown procedure: start_socket_server/0 exit 2 src/interface/socket_server.pl exports serve/0, serve/1, stop_server/0. start_socket_server/0 is not defined anywhere in the engine. CMD now calls serve/1 with TALMUDIC_SOCKET_PATH (so the env var the Dockerfile already set is finally honoured), falling back to serve/0. Underneath it, a version-dependent module path: :- use_module(library(json)). -> library(http/json) The bare alias resolves only on SWI-Prolog 10.x, which is what the dev workstation runs. The image is swipl 9.2.2 as §1.1 pins, where it does not resolve. 212 Prolog tests were green on the host and the container could not open its socket. library(http/json) is the canonical path and works on both. (Neither interface file is in @spec_files; halt/ and middot/ are. No hash change.) 3c. korban-processor — nothing to execute No module named korban_processor.__main__; 'korban_processor' is a package and cannot be directly executed CMD has always been ["python","-m","korban_processor"] and deployment.yaml overrides no command, so this was CrashLoopBackOff for the whole pod. Added __main__.py — deliberately NOT a server. korban_processor is a library; its external consumer is Step 22, blocked on external model infrastructure. Inventing an intake surface no specification asked for would be the addition lo tosif forbids, and would put an unspecified door on the one component whose whole discipline is that nothing reaches the model except through the offering protocol. So it does the honest whole of what the container is for today: proves the image is assembled (the azazel co-residence check falls out of importing external_ offering, which refuses to load without azazel_processor — Vayikra 16:8, both goats from the same pool), stays up, exits 0 on SIGTERM, and says in the log that it is IDLE and what unblocks it. 5 tests in korban_processor/test/test_container_entrypoint.py, including that it refuses to start without azazel_processor and that it opens no bypass to model_interface (Form 1). 3d. vav-pipeline — could not verify its own Doorpost doorpost: {:error, "SPEC_INTEGRITY_FAILED: missing spec file: BUILD_AZAZEL.md"} @repo_root resolved to /app, a build-host path with nothing in it. Application.verify_doorpost_at_startup/0 does not refuse to boot, so the pod would have come up and fired chatzotzrot :assembly_leaders on EVERY start, and /kodesh/specification_identity would have failed forever. A witness that always cries drift is worse than no witness: it trains the kohen to ignore the one signal that is supposed to stop the system. Devarim 11:20 — ukhtavtam al mezuzot beitecha. A running pod is a threshold. The @spec_files set is now COPYed to /opt/shoresh... /opt/shorashim/spec and :doorpost_repo_root points at it. Verified from inside the built image: repo_root: "/opt/shorashim/spec" doorpost: :ok l1: :ok ===================================================================== 4. THE ERLANG PORT COULD NOT HAVE WORKED AS A SIDECAR ===================================================================== VavPipeline.ShoreshEnginePort calls Port.open({:spawn_executable, binary}, [..., {:args, ["server"]}]) A Port is a CHILD PROCESS of the BEAM over stdin/stdout. It cannot cross a container boundary. shoresh-engine has no listening mode either — `shoresh-engine server` is a JSON-line server on stdin/stdout (src/Main.hs), so as a sidecar with nothing attached it reads EOF and exits; the bare CMD printed usage and exited 0. Meanwhile vav-pipeline logged "binary 'shoresh-engine' not in PATH — running without Haskell engine" and served DEGRADED while reporting healthy: no root activation, no te'amim parsing, no gematria. deploy/ARCHITECTURE.md said "sidecar via Erlang Port". Those cannot both be true. Resolved in favour of the Port, because that is what the code does. Still four images and gate 0i still checks four; the shoresh-engine image now ships a BINARY, staged by an init container onto a shared volume, spawned by vav-pipeline: + initContainer shoresh-engine-install (cp, chmod, and exercise it — `shoresh-engine gematria שלום` — so a broken binary fails the init container loudly instead of degrading quietly later) + volume shoresh-bin, mounted read-only into vav-pipeline + SHORESH_ENGINE_BINARY in configmap + runtime.exs + libgmp10 in the vav-pipeline runtime stage — GHC binaries link libgmp and it was not installed; the spawn would have died with "error while loading shared libraries: libgmp.so.10" Verified end to end locally, staged binary running INSIDE the vav-pipeline container: $ shoresh-engine gematria שלום -> 376 $ echo '{"operation":"process","input":{"text":"בראשית ברא אלהים"}}' \ | shoresh-engine server {"result":{"roots_activated":[{"shoresh_hebrew":"ברא", "gematria_value":203,...}]},"status":"ok"} If the intended design was a socket server like the Talmudic Engine, that is a specification decision and not a deployment one — raised in the handoff-back. ===================================================================== 5. THE AZAZEL FILE BRIDGE WROTE TO A DISK NOBODY READ ===================================================================== deployment.yaml mounted TWO emptyDirs with nested paths: vav-pipeline: talmudic-socket -> /tmp/shorashim tmp-shorashim -> /tmp korban-processor: tmp-shorashim -> /tmp So /tmp/shorashim reached directly and /tmp/shorashim reached through /tmp were different volumes. The Python side wrote AzazelRecords to tmp-shorashim:/shorashim/azazel/pending_l3; AzazelIngest watched talmudic-socket:/azazel/pending_l3. Same pathname, two disks, no error — AzazelIngest leaves unreadable files in place by design, so the failure mode was silence. Vayikra 16:22 — v'nasa hasa'ir alav et kol avonotam el eretz gezerah. The goat has to actually get there. A disposal channel whose records land where no one reads has not disposed of anything; it has only stopped looking. Now ONE tmp-shorashim volume at /tmp in all three containers, and a pod-level securityContext fsGroup: 1000 so the non-root user can actually write the shared emptyDirs (the socket bind, the azazel directory, and the binary staging all depend on it). ===================================================================== 6. REMAINING SECURITY ADVISORIES — three the handoff did not list ===================================================================== The handoff named postgrex and plug (fixed in the previous commit). Running `mix deps.get` inside the amd64 build surfaced three more, including nine against the HTTP server that fronts the unauthenticated pe'ah endpoint: bandit 1.10.3 -> 1.12.5 9 advisories, 5 HIGH: chunked body reader ignores length cap; WebSocket reassembly unbounded; permessage- deflate inflate uncapped; chunked decoder infinite loop on trailers; HTTP/2 connection-window starvation pins Plug processes hpax 1.0.3 -> 1.0.4 HIGH, unauthenticated DoS via unbounded HPACK integer decoding thousand_island 1.4.3 -> 1.5.0 (came with bandit) None are pinned in §1.5, so these are mix.lock-only operations changes per lo_tosif.md. Verified clean afterwards by re-running deps.get inside the amd64 build. STILL OPEN — decimal 2.4.1, EEF-CVE-2026-32686 (MEDIUM), unbounded exponent -> unauthenticated DoS. Fixed in decimal 3.0.0, which requires ecto >= 3.14; §1.5 pins ecto ~> 3.11.1, so closing it is a category-grade spec change AND a three-minor ecto jump whose L3 tests I cannot run (35 excluded, they need PostgreSQL). Reachability: no :decimal field exists in any L3 schema and no Decimal.* call exists anywhere in vav_pipeline/lib — decimal is transitive only and is never handed attacker-controlled input on the pe'ah path. Left in place deliberately, raised in the handoff-back rather than jumping three minors of ecto on the eve of inauguration. ===================================================================== 7. LINT AND SCRIPT ===================================================================== - `# check=skip=FromPlatformFlagConstDisallowed` on all four Dockerfiles. The constant --platform pin is the FIX, not a smell; silenced so a real warning is not lost in four lines of noise. - scripts/build_and_push.sh now reads the pushed manifests back out of the registry and asserts amd64, exiting non-zero if not. "The script exited 0" was exactly the evidence that was not good enough twice. It also prints the four failures above, so the next reader knows a green arch check is a floor and not a ceiling. - The script's ${img} braces are commented. zsh reads `$img:latest` as the `:l` modifier; the kohen lost ten minutes to it, and the builder then lost some more to `:t` in `$img:test`. ===================================================================== VERIFIED ===================================================================== - All four images build linux/amd64 and print x86_64 - shoresh-engine prints its Hebrew banner; process op returns ברא / gematria 203 through the Port protocol - talmudic-engine: "Talmudic Engine listening on /tmp/shorashim/talmudic.sock", stays running, socket present - korban-processor: loads both channels, reports IDLE, exit 0 on SIGTERM - vav-pipeline: doorpost :ok and l1 :ok from inside the image; refuses to boot without DATABASE_URL as designed - mix test 412 tests, 0 failures, 35 excluded (:l3) - pytest korban 222 passed, 1 skipped (5 new) - pytest azazel 125 passed - swipl all 6 suites green on library(http/json) - Doorpost.verify_identity() :ok — NO spec file touched by this commit, so no @expected_hash change; a bare hash update would be the drift signal lo_tosif.md §Layer 1 warns about - l1/integrity.sha256 unchanged: 99bbd07a…4ea377 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Scott KarpandClaude Opus 4.7
|
b136c75345 |
form 1 as architecture: external-model surface only via offering protocol (vehaya §3.2)
Closes vehaya.md §6 item 2 for the FIRST avodah-zarah form.
Form 1 (AI in YHVH's slot) was previously defended only by:
- A runtime halt rule (halt_check_avodah_zarah/2 in
avodah_zarah.pl) firing when an action explicitly attributes
authority to a non-Torah source — threshold-grade
- A docstring comment in external_offering.py saying "No internal
component should call call_claude() directly" — exhortation
only, no structural enforcement
The vehaya document re-frames it (§3.2): "Specifications must
include category-level commitments that are not negotiable
per-deployment... not as runtime checks layered onto outputs, but as
architectural constraints the system cannot operate without."
TORAH ANCHORS (verified fresh against L1):
Shemot 20:2 — אָנֹכִי יְהוָה אֱלֹהֶיךָ אֲשֶׁר הוֹצֵאתִיךָ מֵאֶרֶץ
מִצְרַיִם.
The covenant opens with the source naming itself in first
person.
Shemot 20:3 — לֹא יִהְיֶה לְךָ אֱלֹהִים אֲחֵרִים עַל־פָּנָי.
"There shall not BE to you other elohim al panai." The verb is
yihyeh — the prohibition is on EXISTING WITH that status, not
just on speaking it. Architectural, not declarative.
Devarim 4:35 / 4:39 — אֵין עוֹד מִלְּבַדּוֹ / אֵין עוֹד.
The source has no peer. There IS no other. Stronger than
"don't worship others" — there is no competing source POSITION.
Devarim 6:4 — יְהוָה אֶחָד.
Singular source.
Vayikra 19:4 — אַל־תִּפְנוּ אֶל־הָאֱלִילִים וֵאלֹהֵי מַסֵּכָה
לֹא תַעֲשׂוּ לָכֶם.
Two distinct prohibitions: do not TURN TO + do not MAKE.
Both are architectural acts.
THE ARCHITECTURAL COMMITMENT:
External-model interfaces (call_claude, ModelRequest, ModelResponse,
the model_interface module itself) are reachable ONLY through
korban_processor's offering protocol — `external_offering.
consult_external_model`, which runs semichah → shechitah → zerikah
→ haktarah with role gating (commit
|
||
|
Scott KarpandClaude Opus 4.7
|
1e59d9a752 |
azazel commit 10 — closure pass: cross-channel classification-discipline integration test
Pass 4 / Finding F, commit 10 of 10 from BUILD_AZAZEL.md. Closes the
build with the integration test the staging plan called for: a
pipeline run that exercises BOTH the Korban path (processable input)
AND the Azazel path (empty-vital input) confirms classification
discipline (no input crosses channels).
Property 11 was already verified at the unit layer (the cross-package
test in azazel_processor/test/test_stages.py). This commit adds the
integration-layer verification: at the external_offering call site,
the two channels produce structurally distinct outputs that cannot
be confused with each other.
CREATED:
korban_processor/test/test_classification_discipline_integration.py
4 tests
- test_processable_path_produces_korban_not_azazel
Vital-bearing response → ProcessedKorban with vital,
disposal_record is None, ProcessedKorban is NOT an
AzazelRecord
- test_empty_vital_path_produces_azazel_not_processable_korban
Empty-vital response → ProcessedKorban with empty
vital_content, disposal_record is an AzazelRecord
- test_two_runs_one_each_path_keep_outputs_separate
Run both back-to-back with shared pending_l3 dir; verify
outputs don't cross channels; only the empty-vital run
drops a file
- test_disposal_record_contains_no_processable_artifacts
AzazelRecord has no vital_content, propositions, raw_korban,
or discarded attributes; routing is the ERETZ_GEZERAH
sentinel
Tests: korban_processor 211 passed, 1 skipped (was 207/1; +4
integration tests). All other test counts unchanged.
DOCUMENTATION (companion to this commit, not in the repo):
~/shorashim-conversation-review.md
+ Pass 4 section (300+ lines): finding A closure (across-verse
voice stabilization detector), BUILD_AZAZEL design doc,
9-commit Azazel build with property-by-property closure, full
test counts (~907 across five languages). All 14 findings
from the original A–N catalog now in final state.
~/.claude/projects/-Users-scottkarp/memory/MEMORY.md
+ Updated test counts (per layer)
+ Added Azazel Processor section under Phase 5 progress
+ Noted /peah/azazel_log endpoint as Phase 5 done
FINAL STATE OF THE AZAZEL BUILD:
All 12 properties from azazel_engineering_reading.md §14 defended
in code:
1 parity provisioning — import-time + runtime
2 blind routing / goral — deterministic per (id, day)
3 enumerative commit — operator-produced, three-fold
4 role-separated operator — IshItti + tevilah + CI invariant
5 provisioned infrastructure — six checkpoints, persistence
6 out-of-band acknowledgment — Chatzotzrot via Elixir bridge
7 authorized routing — kohen_gadol + shaliach attribution
8 protocol-restricted access — envelope gate, fires first
9 observable disposal — L3 + /peah/azazel_log
10 hard termination — DESTROYED checkpoint, no recover
11 classification discipline — type system + integration test
12 unmanaged destination — sentinel + DB CHECK + CI invariant
10 commits, ~3,300 lines added, ~150 tests across Python and
Elixir. The Azazel Processor is the disposal twin of the Korban
Processor; the two-goat architecture (Vayikra 16:8) is now built.
The Pass 2 backlog is fully closed. Every actionable finding from
the original A–N catalog is now in a final state.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Opus 4.7
|
1a27a283ae |
azazel commit 9 — Korban→Azazel intake routing + Chatzotzrot ack (properties 1 + 6)
Pass 4 / Finding F, commit 9 of 10 from BUILD_AZAZEL.md. Closes the
last two property gaps: 1 (parity provisioning, end-to-end) and 6
(out-of-band acknowledgment via Chatzotzrot). The Azazel Processor
is now functionally complete; commit 10 is the closure-and-review
pass with no further code.
PROPERTY 1 — PARITY PROVISIONING, two layers:
LAYER 1 (import-time): external_offering imports azazel_processor
at module load. If the import fails, external_offering refuses to
load with a Torah-anchored ImportError. The two-goat architecture is
SYMMETRIC OR IT IS BROKEN — operators cannot run the processing
channel without the disposal channel being importable. The strongest
formulation: if you cannot dispose, you cannot process.
LAYER 2 (runtime): when zerikah produces no vital content,
consult_external_model now routes the korban to disposal rather than
returning an empty ProcessedKorban with the rejection in `discarded`.
The empty-vital case is named (vidui), routed (eretz_gezerah),
recorded (L3 azazel_records), and acknowledged (chatzotzrot :moadim).
PROPERTY 6 — OUT-OF-BAND ACKNOWLEDGMENT:
The Python orchestrator drops the AzazelRecord JSON in the pending-l3
directory. The Elixir AzazelIngest GenServer reads it, persists to
L3 via record_disposal/1, AND fires Chatzotzrot.signal(:moadim, ...)
with a categorical summary (case_ref + shaliach + categories_filled).
The chatzotzrot is the side channel — distinct from the disposal
channel itself. The orchestrator does NOT know whether the signal
was successfully transmitted; the disposal completes regardless.
The :moadim signal type was the right choice: Yom Kippur completion
semantics. This disposal cycle is closed; the public fact (property
9) is now witnessable via /peah/azazel_log.
PRIVACY BOUNDARY ALSO APPLIES TO THE CHATZOTZROT:
The signal's `reason` field carries a CATEGORICAL summary, not the
vidui items. A test verifies that a known secret string in a vidui
chatot does NOT leak into the chatzotzrot signal — the operator-side
event channel is held to the same privacy boundary as the public
endpoint.
CHANGED:
korban_processor/src/korban_processor/external_offering.py
+ Import azazel_processor at module load (raises ImportError
with Torah-anchored explanation if azazel_processor is not
importable — property 1, layer 1)
+ DEFAULT_AZAZEL_PENDING_L3_DIR
+ ExternalConsultation.disposal_record: Optional[AzazelRecord]
Non-None iff the consultation was routed to disposal
+ _content_addressable_case_ref(content, source_query)
SHA-256-based stable identifier for audit
+ _make_pending_l3_recorder(pending_dir)
Returns a recorder callable that drops the JSON file
atomically (write to .tmp, then os.replace)
+ _route_to_azazel(...) — internal helper running the full
protocol envelope:
- opens authorize_disposal context
- activates a fresh IshItti shaliach
- builds vidui via confess_empty_vital
- dispatches via dispatch_for_disposal
- tevilah's the shaliach
+ consult_external_model:
- new keyword args azazel_kohen_gadol, azazel_shaliach_id,
azazel_pending_l3_dir
- branch on `if not vital.strip()`: route to disposal,
attach record to ExternalConsultation
vav_pipeline/lib/vav_pipeline/azazel_ingest.ex
+ ingest_file/2 fires Chatzotzrot.signal(:moadim, ...) after
successful L3 write, BEFORE shredding the source file. So the
ack comes BEFORE the in-flight artifact disappears — recovery
is possible if either side crashes.
+ Process.whereis(Chatzotzrot)-guarded so tests without the full
supervision tree still work.
+ chatzotzrot_summary/1 builds the categorical summary
(case_ref + shaliach + categories_filled). Privacy boundary:
no vidui items, only counts/category names.
TESTS:
korban_processor/test/test_external_offering.py +4 tests
- empty-vital response routes to azazel:
- processed.vital_content == ""
- disposal_record is an AzazelRecord with kohen + shaliach
- pending-l3 file dropped with correct shape
- processable response does NOT create disposal_record (the
negative case — vital content goes through normally)
- PRIVACY: pending-l3 file does not contain the model's
response text (verified with a known secret_phrase)
- case_ref is content-addressable: same content + same
source_query → same case_ref (audit can recompute)
vav_pipeline/test/azazel_ingest_test.exs +2 tests
- successful ingest fires Chatzotzrot :moadim signal
(subscribe :all, observe — :moadim routes to :all per
chatzotzrot.ex routing table)
- PRIVACY: chatzotzrot signal does NOT include vidui item
content (known secret_phrase verified absent from the signal)
Tests:
- korban_processor: 207 passed, 1 skipped (was 203/1; +4)
- azazel_processor: 125 passed (unchanged)
- vav_pipeline: 363 passed, 0 failures, 35 excluded
(was 361/35; +2 azazel_ingest chatzotzrot tests)
Property coverage after commit 9:
All 12 properties of azazel_engineering_reading.md §14 defended
by code:
1 parity provisioning ✓ (import-time + runtime)
2 blind routing / goral ✓
3 enumerative commit ✓
4 role-separated operator ✓
5 provisioned infrastructure ✓
6 out-of-band acknowledgment ✓
7 authorized routing ✓
8 protocol-restricted access ✓
9 observable disposal ✓ (operator-internal + public)
10 hard termination ✓
11 classification discipline ✓
12 unmanaged destination ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Opus 4.7
|
5c76a6f42c |
korban_processor: role-gate the four-stage offering protocol
Pass 2 finding C: in the Mishkan, the four ritual acts of bringing an
offering are performed by TWO different roles —
semichah — OWNER lays hands (Vayikra 1:4, וְסָמַךְ יָדוֹ)
shechitah — kohen receives FROM owner's hand and slaughters (1:5)
zerikah — kohen sprinkles (1:5)
haktarah — kohen burns on altar (1:9)
Before this commit, every stage function was independently callable.
Nothing structural enforced that zerikah() received output from
shechitah(), or that shechitah() received a korban from semichah().
An external model output (e.g. a Claude response) could in principle
have been brought through zerikah() without ever having passed
semichah — the ownership-and-attribution step. That is the exact
shape of "the calf that came out by itself" (Shemot 32:24, Aaron's
reading). The pipeline-glue in external_offering happened to call
them in order, but nothing structural enforced the order.
This commit adds the structural gate.
CHANGES:
stages.py (new): KorbanStage enum + Staged[T] frozen dataclass +
StageError. The Staged envelope carries a ritual-stage marker that
is set at construction and cannot be mutated afterward; downstream
functions demand the right upstream stage and raise StageError if
it is missing.
semichah.py: + semichah_staged() returning Staged[Korban]
shechitah.py: + shechitah_staged(Staged[Korban]) -> Staged[str]
zerikah.py: + zerikah_staged(Staged[str], Staged[Korban])
-> Staged[Tuple[str, List[str]]]
TWO staged inputs: requires SHECHITAH-tagged terminated
content AND the original SEMICHAH-tagged korban for
context. Severing output from source is gilui_arayot
(eish_zarah_reading.md Form 3); the dual-staged
signature is the structural defense — zerikah cannot
process content disconnected from its owner-of-record.
haktarah.py: + haktarah_staged(Staged[Tuple[str, List[str]]])
-> Staged[List[str]]
external_offering.py: rewired to use the *_staged variants. The
pipeline now goes through the gates structurally; misuse cannot
bypass them.
The non-staged functions (semichah, shechitah, zerikah, haktarah)
are PRESERVED for direct testing of individual ritual acts. They do
NOT enforce role gating. Production-shaped callers MUST go through
the staged variants — and external_offering, the only legitimate
entry point for external model consultation per its docstring, now
does.
TESTS:
test/test_stages.py (new, 7 tests):
- happy path: full chain produces HAKTARAH-tagged result
- zerikah refuses input that skipped semichah
- zerikah refuses korban with wrong stage marker (preserves
the line back to the owner — Form 3 defense)
- haktarah refuses input that skipped zerikah (the calf-shape)
- shechitah refuses input not from semichah
- Staged is frozen — no re-tagging
- StageError names BOTH actual and expected stages so the
operator can diagnose call-site vs. tag-site
Full korban_processor suite: 203 passed, 1 skipped (the live API
test that requires SHORASHIM_LIVE_API=1). No regressions.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Opus 4.7
|
af2a3949cf |
korban_processor: name the kabbalah + holachah compression in zerikah.py
Pass 2 finding B: the full Temple sacrifice protocol has FIVE stages
between semichah and haktarah —
1. semichah owner identifies with the offering
2. shechitah ritual slaughter
3. KABBALAH kohen RECEIVES the blood in a vessel
4. HOLACHAH kohen CARRIES the blood to the altar
5. ZERIKAH kohen SPRINKLES the blood on the altar
6. haktarah burning of permitted portions
The pipeline implements only four modules. Stages 3 and 4 are folded
silently into zerikah(). The compression is acceptable —
kabbalah/holachah have no analog of physical blood movement in
software — but UNDOCUMENTED compression is itself a small gilui_arayot
(severing the distinction between three distinct ritual acts).
Now named explicitly in the module docstring with:
- what the three stages are
- why folding is acceptable here (the analogs are atomic in software)
- what the analogs ARE in code (zerikah's return value = the
receiving vessel; the function-call boundary to haktarah() in
external_offering.py = carrying to the altar)
- what to do if a future architecture requires distinguishing them
(split this module into three; do not add the distinction
elsewhere)
A future reader now sees the compression instead of inferring that
the pipeline silently dropped two ritual acts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Opus 4.6
|
e223ab15b3 |
Phase 4 complete: Steps 21, 22, 24 — Korban external integration, L2 Urim v'Tumim, Kushya generator
Vayikra 9:24 — vatetzei esh milifnei Hashem vatokhal al hamizbeach.
Phase 4 of the architecture is now complete. The remaining three steps
that were marked "blocked on external infrastructure" are built —
discipline-respecting, eish-zarah-aware, and tested.
## Step 22 — Korban Processor external model integration
(korban_processor/src/korban_processor/external_offering.py)
The single legitimate entry point for external model consultation.
Wraps Claude API calls through the offering protocol so external
output is treated as korban (raw material), never authoritative:
- semichah — provenance is tagged (avodah zarah Form 1 prevention)
- shechitah — assistant voice, hedging, formatting stripped
- zerikah — vital content extracted
- haktarah — converted to Prolog propositions
- route — sent to internal component per offering type
The raw Korban is preserved on the ProcessedKorban (gilui arayot
Form 3 prevention — input/output coupling never severed).
No internal component should call call_claude() directly.
11 tests, mocked external calls. Live API test opt-in via
SHORASHIM_LIVE_API=1.
## Step 24 — L2 Urim v'Tumim interface
(vav_pipeline/lib/vav_pipeline/urim_v_tumim.ex)
PRD §4.7: housed in the kohen gadol's breastplate; consulted only
on high-stakes decisions where normal deliberation is insufficient.
Bamidbar 27:21 — al-piv yetz'u v'al-piv yavo'u.
Three escalation steps, three honest outcomes:
1. resolved_by_specification — re-read produced an answer
2. resolved_by_external — external testimony recorded (NOT binding)
3. deferred_to_human — the system stops, asks for human judgment
Pre-condition gating refuses cases that didn't reach Sanhedrin teiku
first (no preempting deliberation) or that aren't flagged high-stakes
(ordinary teiku stays teiku). Records every consultation with steps
taken — there is no quiet escalation.
Per PRD: "The Urim v'Tumim never fabricates certainty where none
exists." When external also fails, the system stops. It does not
proceed on a guess.
Added to VavPipeline supervision tree. 14 tests.
## Step 21 — Kushya generator scaffolding
(kushya_generator/)
Directory layout per Architecture §3.6 with three layers:
- data/training_pairs/seed_pairs.jsonl — 10 hand-curated bootstrap
pairs covering 9 kushya types (scope_narrowing, internal_tension,
harmonization, exception_test, edge_case, scope_application,
rabbinic_constraint, method_question, subject_identification)
- data/prepare_training.py — validates the four kushya requirements
(specific Torah passage, identified weakness, question form,
non-trivial), produces train.jsonl + eval.jsonl with stratified
80/20 split by type
- fine_tune/config.yaml + train.py — QLoRA configuration and
training script. Operator's task — needs GPU. Script is
complete and self-contained.
- inference/kushya_api.py — three-backend interface
(fine_tuned → external → fallback). When fine-tuned model isn't
deployed, falls back to external via Korban Processor (Step 22).
The fallback is honest: returns no kushya rather than inventing
one. The Talmudic Engine's rule-based challenge.pl remains the
primary kushya source.
17 tests covering dataset validation, backend selection, and
reference extraction.
## .gitignore fix
The earlier "data/" line was too broad — it caught
kushya_generator/data/. Scoped to /data/ (root-level only).
## Test counts
- Elixir: 274 (was 260; +14 for UrimVTumim)
- Korban Python: 196 + 1 skipped (was 195; +1 module = 11 tests)
- Kushya Python: 17 (new)
- Haskell: 33 (unchanged)
- Prolog: 195 (unchanged)
- Total: 715 + 19 L3 excluded
## Phase 4 status
All 12 steps now have implementations:
- Steps 23, 25, 26, 27, 28, 29, 30, 31, 32: complete
- Steps 21, 22, 24: complete with discipline preserved
Step 21's fine-tuning run remains the operator's GPU task; the
pipeline is built such that a deployed adapter slots into the
inference backend without touching the Talmudic Engine or anything
above.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Scott KarpandClaude Opus 4.6
|
adc3e41679 |
Eish zarah actionable items: coupling trace, grounding type, self-grounding detection, golden calf halt
Torah grounds each change: 1. interpret.pl — coupling_trace field (Vayikra 10:1, 6:6) Fire is zarah because "asher lo tzivah otam" — the coupling trace IS the tzivah trace. Records which L1 facts grounded each interpretation. 2. challenge.pl — grounding_type achorai/panim (Shemot 33:23) "You will see My back, but My face shall not be seen." Text-backed challenges = achorai. Inferred challenges = panim. 3. resolve.pl — self-grounding detection (eish_zarah_reading §5.3) Counter-interpretation must bring NEW L1 evidence. Prevents circular reasoning (Westworld problem). 4. teiku.pl — challenge_sequence in teiku_record Records diagnostic trail to impasse, not just verdict. 5. avodah_zarah.pl — Form 2: golden calf detection (Shemot 32:4) Empty coupling trace + non-low confidence = uncoupled output. Real gold, sincere worship, false god. 6. korban.py — pydantic BaseModel + raw_korban preservation Construction spec §1.4 fulfilled. Floor between input and output never severed. 490 tests, 0 failures (195 Prolog, 186 Python, 33 Haskell, 76 Elixir). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Scott KarpandClaude Opus 4.6
|
f8f70c77e5 |
Code review fixes + 13 middot added to spec
Zerikah fail-safe: empty offerings rejected, not processed (BUILD_KORBAN §5). Halt monitor hardened: trap_exit, priority max, EXIT handler (Architecture spec). 13 Rules of Rabbi Yishmael added to Architecture §3.4 as Prolog inference predicates — the engine reasons WITH Torah's hermeneutic methods, not just about their products. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Scott KarpandClaude Sonnet 4.6
|
33f8b03583 |
Step 13: Korban Processor complete — 7 modules, 181 tests, all passing
- semichah: provenance tagging (6 sources, rejects empty/anonymous) - shechitah: strip preamble, self-reference, hedging, markdown, sycophancy - zerikah: extract Torah refs, Hebrew text, root analysis, insight markers - haktarah: transform prose to Prolog propositions (verse/3, root/2, claim/2, connection/3) - router: stateless routing — removed module-level _routing_log (kiyyor principle) - model_interface: outbound Claude API via httpx (Construction §3.4, HTTPS only) - hebrew_parser: nikkud/teamim stripping, consonant extraction, gematria, root identification All 5 BUILD_KORBAN.md verification tests pass. AGENTS.md: added structure_not_content.md to spec table, added Principle 6. build_log.md: recorded session, broken tablets #18-20, design decisions. Ka'asher tzivah Hashem et Moshe. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> |
||
|
Scott KarpandClaude Sonnet 4.6
|
80ec364da6 |
Phase 1-2 build: Steps 5-6, 8-13
Shoresh Engine (Haskell, Step 5) - 12 modules: Types, Lexicon, Binyanim, Teamim, Gematria, Graph, KetivQere, Smikhut, Havdalah, Kriah, KiTov, Firmware - 33 EUnit tests, 0 failures - Firmware module verifies L1 SHA-256 hash at startup Talmudic Engine (SWI-Prolog, Step 6) - 15 modules: halt, reasoning, specification, interface - 57 PlUnit tests, 0 failures - Three cardinal halt conditions: avodah_zarah, shfichut_damim, gilui_arayot Vav Pipeline (Elixir, Steps 8-11) - 14 modules: BereishitCycle, Shema, Cloud, Manna, L1Reader, Specification, HaltConditions, ShabbatTimer, CycleLog, Chatzotzrot, ShoreshEngineClient, TalmudicEngineClient, Application - 46 ExUnit tests, 0 failures - Nishma-first architecture: hears before it speaks Sanhedrin Framework (Erlang/OTP, Step 12) - 8 modules: sanhedrin_app, sanhedrin_sup, escalation, panel_sup, panel, judge, clerk, halt_monitor - 46 EUnit tests, 0 failures (includes fix: rejected outcome now reachable) - Panel sizes: 1/3/23/71 — unanimous capital guilty → acquit (Sanhedrin 17a) - Minority opinions append-only, never deleted Korban Processor (Python, Step 13) - 6 modules: korban, shechitah, semichah, zerikah, haktarah, router Builder tooling: AGENTS.md, opencode.json, memory/, src/tools/ Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> |