bikoret/specs/bikoret_terumah.md

BIKORET: Materials Inventory

Document Purpose

This is the terumah list — what has been gathered, what must be acquired, what must be built. Updated to reflect the full specification set: pinned versions from the Construction Specification, corrected shema architecture (raw marks received, not pre-computed trees), cloud as shared state file (not message handler), manna as external provision (not internal calculation), and team requirements from the Bezalel section.

Status key: ✅ = in hand, ⚠️ = available but needs acquisition, ❌ = does not exist / must be created

---

A. HARDWARE

Item	Status	Notes
Server: AMD EPYC 7313P (or comparable), 16 cores, 128GB DDR4 ECC, 500GB NVMe (system) + 1TB NVMe (L3)	⚠️	Rentable. Infomaniak (CH), 1984 (IS), Bahnhof (SE) evaluated. ~$200-350/month. Exact spec from Construction Specification §2.1.
GPU	✅ NOT NEEDED	No GPU. Not "no GPU required." No GPU. This is a specification.

Decision required: Select hosting provider and provision server.

---

B. OPERATING SYSTEM & INFRASTRUCTURE

All versions pinned per Construction Specification §1.1.

Item	Version (Pinned)	Status	Notes
Ubuntu	22.04 in the pipeline and engine images (ubuntu:jammy-20240125); 24.04 LTS on the cluster node (managed)	✅	Free. The korban and talmudic images are Debian 12, from their language base images.
systemd	ships with Ubuntu	✅	Service management, cycle timers.
Kubernetes / containerd	1.35 / 2.2 (managed, Exoscale SKS)	✅	Runs the four images as one pod. Images are built with Docker Buildx. Podman is not used: Sanhedrin panels are processes in the pipeline's BEAM, not containers.
PostgreSQL	17 (managed, Exoscale DBaaS; 17.11 on 2026-10-04)	✅	L3 append-only store. Patched by the provider.
squashfs-tools	ships with Ubuntu	✅	For read-only L1 Torah filesystem.
Caddy	latest stable	✅	Reverse proxy with auto-TLS for pe'ah endpoint.
nftables	ships with Ubuntu	✅	Firewall. Configuration in Construction Spec §4.5.
SSH	ships with Ubuntu	✅	Kohanim (operator) access only.

All infrastructure components available. No gaps.

---

C. PROGRAMMING LANGUAGES & RUNTIMES

All versions pinned per Construction Specification §1.1. No ranges. No + suffixes.

Item	Version (Pinned)	Status	Notes
GHC (Haskell compiler)	9.6.4	✅	Free. ghcup install ghc 9.6.4
Cabal (Haskell build tool)	3.10	✅	Free. Ships with ghcup.
Erlang/OTP	26.2.1	✅	Free. apt install erlang or erlang-solutions repo.
Rebar3 (Erlang build tool)	3.22.0	✅	Free. The version mix local.rebar installs for Elixir 1.16.1.
Elixir	1.16.1	✅	Free. Must pair with OTP 26.2.1.
Mix (Elixir build tool)	ships with Elixir	✅	Free.
SWI-Prolog	9.2.2	✅	Free. apt install swi-prolog
Python	3.12.2	✅	Free. From the python:3.12.2-slim base image.
Poetry (Python package manager)	2.5.1	✅	Free. Generates poetry.lock (lock-version 2.1); not installed in any image.

All language runtimes available. No gaps.

---

D. LIBRARIES & PACKAGES

All pinned via lock files per Construction Specification §§1.3-1.5.

Haskell (via cabal.project.freeze)

Package	Version	Status	Notes
base	4.18.2.0	✅	Ships with GHC 9.6.4
containers	0.6.7	✅	Ships with GHC
text	2.0.2	✅	Hackage (free)
aeson	2.2.3.0	✅	Hackage (free) — JSON
vector	0.13.2.0	✅	Hackage (free)
bytestring	0.12.2.0	✅	Hackage (free) — GHC 9.6.4 ships 0.11.5.3; the freeze pins this one
filepath	1.4.301.0	✅	Hackage (free) — GHC 9.6.4 ships 1.4.200.1; the freeze pins this one
directory	1.3.10.1	✅	Hackage (free) — GHC 9.6.4 ships 1.3.8.1; the freeze pins this one
hspec	2.11.16	✅	Hackage (free) — testing

Erlang (via rebar.lock)

Package	Version	Status	Notes
OTP standard library	ships with OTP 26.2.1	✅	gen_server, supervisor, application
jsx	3.1.0	✅	hex.pm (free) — JSON
common_test	ships with OTP	✅	Testing

Elixir (via mix.lock)

Package	Version	Status	Notes
jason	~> 1.4.1 (locked 1.4.5)	✅	hex.pm (free) — JSON
ecto_sql	~> 3.11.1 (locked 3.11.3; ecto 3.11.2)	✅	hex.pm (free) — database interface
postgrex	~> 0.22.3 (locked 0.22.4)	✅	hex.pm (free) — PostgreSQL driver
bandit	~> 1.2 (locked 1.12.5)	✅	hex.pm (free) — HTTP server for the pe'ah / kodesh API
plug	~> 1.15 (locked 1.20.3)	✅	hex.pm (free) — HTTP routing
telemetry	locked 1.4.2	✅	hex.pm (free) — transitive (Ecto, Bandit); not depended on directly
sanhedrin	path dependency	✅	../sanhedrin, built by rebar3 into the same BEAM

Python (via poetry.lock)

Package	Version	Status	Notes
httpx	0.27.0	✅	PyPI (free) — API calls
pydantic	2.6.1	✅	PyPI (free) — data validation
hebrew-tokenizer	2.3.0	⚠️	PyPI. Needs evaluation for biblical vs. modern Hebrew.
pytest	>=8.0,<9.0 (locked 8.4.2)	✅	PyPI (free) — dev group only; not in any image
llama-cpp-python	0.2.55	✅	PyPI (free) — Mistral inference
peft	0.9.0	✅	PyPI (free) — fine-tuning
transformers	4.38.2	✅	PyPI (free) — model tooling
datasets	2.17.1	✅	PyPI (free) — training data

Near-complete. Two Python packages need compatibility testing.

---

E. DATA — THE SPECIFICATION

E1. Torah Text (L1)

Source: Westminster Leningrad Codex (WLC) 4.20, as carried by the Open Scriptures Hebrew Bible (openscriptures/morphhb), per Construction Specification §1.2. One source; not cross-validated against a second edition.

Item	Status	Notes
Full Torah text — consonantal (ktiv)	✅	WLC via OSHB morphhb.
Full Torah text — voweled (nikkud)	✅	WLC via OSHB morphhb.
Cantillation marks (raw Unicode codepoints U+0591–U+05AF)	✅	Present in WLC data. Stored as raw marks on the text, NOT pre-parsed into trees. The system reads them in the act of reading.
Petuchah/Stumah paragraph markers	✅	In L1: 293 petuchot and 394 setumot, read from the WLC's פ / ס marks (OSIS <seg type="x-pe">, x-samekh) and stored per verse as parashah_breaks.
Inverted nuns (nun hafukha)	✅	In L1: the two signs after Bamidbar 10:34 and 10:36, from the WLC's own marks, stored per verse as inverted_nuns.
Large and small letters	⚠️	The WLC source marks four large letters in Torah (in L1 as large_letters) and NO small letters. The fuller Masoretic lists of enlarged and reduced letters are not in the source and not in L1.
Ketiv/qere annotations	✅	WLC annotations: 67 pairs in 64 verses. Stored per verse as {word_position, ketiv, ketiv_consonants, qere, qere_consonants} — both readings preserved, never collapsed.
Gematria (root → numerical value)	✅	Not stored in L1: computed from the consonants when read. Letter values fixed. No equivalence index is built.
Smikhut compound index (attested construct-state pairs)	✅	Built: 9,203 pairs in 4,106 verses, extracted by scripts/build_l1.py from the morphology where a word is marked construct.
L1 SHA-256 hash	❌	Computed at build time. Stored in L0 firmware. Verified at every boot.
Letter counts / word counts / Masoretic notes	⚠️	Available in digital WLC. Needs extraction and formatting.

Critical change from previous inventory: Te'amim parse trees are no longer a materials item. The system builds parse trees by reading the raw marks, the way a literate reader constructs sentence structure from punctuation. The raw marks are the material. The reading is the system's act.

E2. Root Lexicon

Item	Status	Notes
Complete Biblical Hebrew root lexicon	✅	BDB (public domain) + Strong's (free), already consolidated: the OSHB Hebrew Lexicon (openscriptures/HebrewLexicon, LexicalIndex.xml + AugIndex.xml) files every Strong's lemma under its BDB root. scripts/build_l1.py reads each word's shoresh from it. Torah attests 1,378 roots (1,243 BDB roots, 135 words standing alone); the earlier estimate here was ~500.
Root-to-verse index	⚠️	Derivable from Strong's numbers + verse mappings. Needs processing.
Root connectivity graph (shared consonants)	❌	Algorithmic. ~1 day.
Binyan glosses per root	⚠️	Partially available from BDB/Strong's. Mark attested vs. theoretical forms.
Semantic domain classifications	❌	Categorize each root into domains. ~1-2 weeks.

E3. 613 Mitzvot

Item	Status	Notes
List of 613 mitzvot with Torah source verses	✅	Rambam's Sefer HaMitzvot. Multiple digital sources.
613 mitzvot in Prolog logical format	❌	Core encoding task. ~4-6 weeks.
Relationships between mitzvot	❌	Dependencies, conflicts, groupings. ~2-3 weeks.

E5. The Second Scroll — the mesorah snapshot (bishe'arecha §IV)

Source: the Sefaria export (public storage bucket sefaria-export, catalogue books.json generated 2026-10-02), one chosen version per work per language, taken by scripts/build_mesorah.py from scripts/mesorah_sources.json. Only open versions: Public Domain, CC0, CC-BY, CC-BY-SA — the licence in each file is checked against the sources file and the build refuses otherwise. No CC-BY-NC text is taken (the Steinsaltz/Davidson Talmud is therefore not here; the Bavli is Wikisource's, CC-BY-SA). Sefaria and each version's source are credited in mesorah/index.json.

Material	Status	Notes
Rashi, Onkelos, Ramban, Ibn Ezra, Sforno on the five books	✅	Hebrew (public domain editions) and English (Silbermann 1929–34 public domain; Sefaria Community CC0)
Mishnah, all 63 tractates	✅	Hebrew Torat Emet (PD); English where open
Talmud Bavli, 37 tractates	✅	Hebrew: Wikisource (CC-BY-SA). English only where an open one exists (Cohen's Berakhot, PD)
Midrash halakha: Sifra, Sifrei Bamidbar, Sifrei Devarim	✅	Sifra's Hebrew is a 1545 print marked "unknown" licence by Sefaria — EXCLUDED; Sifrei included
Midrash aggadah: Bereshit / Shemot / Vayikra / Bamidbar / Devarim Rabbah, Tanchuma	⚠️	Vayikra, Bamidbar, Devarim Rabbah Hebrew are "unknown" licence — EXCLUDED; Bereshit and Shemot Rabbah and Tanchuma included
Ramban on Shemot	❌	Hebrew "unknown" licence — excluded
Mekhilta	❌	Not in the export catalogue under that title

Extent: 128 works, 256 files, 44.9 MiB, tree hash 7d5b3a6c… (mesorah/integrity.sha256). Seven works have named nodes (Ramban on four books, Ibn Ezra and Sforno on Bereishit, Tanchuma); the reader addresses their unnamed node, and Tanchuma by parashah.

E4. Talmudic Training Data (Kushya Generator)

Item	Status	Notes
Talmud Bavli full text	✅	Sefaria (free, open).
Talmud Yerushalmi full text	✅	Sefaria (free, open).
Extracted (interpretation, kushya) pairs	❌	Most labor-intensive. ~4-8 weeks.
Rishonim commentary	✅	Sefaria (Rashi, Tosafot, Rambam, etc.).
Extracted challenges from Rishonim	❌	~2-4 weeks.

---

F. MODELS (KORBAN SOURCES & WOOD)

Item	Status	Notes
Mistral Medium 3.1 GGUF (Q5_K_M)	⚠️	HuggingFace. Free (Apache 2.0). ~25-30GB download.
llama.cpp	✅	Free, open source. CPU inference.
Mistral Large 3 API access	⚠️	Mistral API. Paid per token. Phase 1 only.
Claude API access	⚠️	Anthropic API. Paid per token. Korban source.
QLoRA fine-tuning tooling	⚠️	peft, transformers. Free. One-time GPU rental ~$50-100.

---

G. EXTERNAL APIS

Item	Status	Notes
Sefaria API	✅	Free, no key needed. For L1 data acquisition.
Anthropic API (Claude)	⚠️	Need API key. ~$30-80/month for korban processing.
Mistral API	⚠️	Need API key. ~$20-50/month Phase 1 only. $0 after self-hosting.

---

H. GOVERNANCE AND PROVISION INFRASTRUCTURE

Governance is Torah. The cloud is L1 — the read-only Torah filesystem. There is no separate governance infrastructure. The system reads Torah for governance as it reads Torah for specification. See Architecture Specification §3.10 (second tablets).

Provision is the community. Manna (resources) comes from the machatzit hashekel. A kohen (operator) or gabbai (treasurer) sets the daily provision.

Item	Status	Notes
L1 Torah filesystem (IS the cloud)	✅	Already specified. Read-only mount. Integrity hash verified at boot. Cloud IS Torah.
Calendar module (Shabbat/festivals/shmitah)	❌	Part of cloud governance. Reads Torah for when to rest, celebrate, review. NOAA solar calculator for exact times. ~2 days.
Manna provision file infrastructure	❌	/var/lib/bikoret/manna/today.json. Kohen/treasurer writes, system reads. CLI tool manna-provision. ~1 day.
Kohen (operator) secure access	⚠️	SSH keys, mTLS certs. Standard infrastructure.
Gabbai (treasurer) tools	❌	Manna provision CLI + pekudei accounting system per Economic Model §6.2.
Pekudei accounting system	❌	Monthly financial reporting per Economic Model §6.2.

The critical non-technical dependency is the community. Can 70 people be found who value this? The system needs kohanim (operators), a gabbai (treasurer), and a community that contributes machatzit hashekel. It does not need a governance board. Torah governs.

---

I. BEZALEL — The Builder

Per Construction Specification §5 (second tablets).

Role	Who	Status
Builder	AI (Claude Code + Claude) — writes code in all six languages, reads Biblical Hebrew and Torah	✅ Available
Transmitter	Human — says "read Torah," carries specification between contexts	✅ Present
Kohen (operator)	Human — serves the running system	❌ Needed at deployment
Gabbai (treasurer)	Human — administers machatzit hashekel	❌ Needed at deployment
Community	70 members — provides machatzit hashekel	❌ Needed at deployment

No person-weeks. No labor cost. The AI builds. Torah specifies.

---

J. SHEMA LAYER — MATERIALS

Updated to reflect corrected architecture.

Item	Status	Notes
Te'amim reader (raw marks → parse tree in act of reading)	❌	~1-2 weeks. NOT pre-computed storage.
Gematria lookup module + equivalence index	❌	~3 days. Lookup, not computation.
Smikhut compound extractor	❌	~2-3 weeks. AI reads Hebrew.
Ketiv/qere dual-value reformatter	⚠️	~3 days. Raw data exists.
Cloud governance module (L1 + calendar)	❌	~2 days. Reads Torah for governance. No CLI. No state file.
Manna provision file + CLI tool	❌	~1 day. Kohen/treasurer writes provision.
Chatzotzrot signaling system	❌	~3 days.
Tzara'at multi-stage diagnostic	❌	~1-2 weeks.
Arei miklat refuge protocol	❌	~1 week.
Eglah arufah accountability protocol	❌	~3 days.
Bikurim release protocol	❌	~3 days.

---

SUMMARY

What We Have (✅)

• All programming languages and runtimes (pinned versions)
• All infrastructure tools
• Nearly all library dependencies
• Torah text (WLC 4.20, OSHB morphhb) with cantillation marks, ketiv/qere, vowels, lemma and morphology
• Talmud and commentary texts (Sefaria)
• List of 613 mitzvot with source verses
• Root lexicon (BDB roots by Strong's lemma — OSHB Hebrew Lexicon)
• Model inference tooling (llama.cpp)
• The specification itself (Torah)

What We Need to Acquire (⚠️)

• Hosting server — ~$200-350/month
• Mistral GGUF weights — free, ~30GB download
• API keys (Anthropic, Mistral) — ~$50-130/month combined
• One-time GPU rental — ~$50-100
• Consultants (Hebrew linguist, Torah scholar) — available through community
• hebrew-tokenizer compatibility testing — time only

What We Must Build (❌)

• Root connectivity graph — ~1 day
• Semantic domain classifications — ~1-2 weeks
• 613 mitzvot in Prolog format — ~4-6 weeks (critical path)
• Mitzvah relationship mapping — ~2-3 weeks
• Talmudic training pairs — ~4-8 weeks (most labor-intensive)
• Rishonim challenge extraction — ~2-4 weeks
• Gematria equivalence index — ~3 days
• Smikhut compound index — ~2-3 weeks
• Te'amim reader — ~1-2 weeks
• Ketiv/qere reformatter — ~3 days
• Torah calendar module (Shabbat/festivals/shmitah) + manna provision tool — ~2 days
• Protection layer (chatzotzrot, tzara'at, miklat, eglah arufah, bikurim) — ~4-5 weeks combined
• L1 SHA-256 integrity system — ~1 day
• Community assembly (kohanim, gabbai, 70 members) — indeterminate — requires humans

Cost Summary

Category	Amount
Terumah (one-time startup)	$265-515
Machatzit hashekel (monthly, build + operations)	~$350
Sustained by	70 members × $5/month
Estimated build cost (6 months)	~$2,100
Seven-year total	~$27,500

No labor cost. The AI builds. The community provides. Torah governs.

Critical Path

Technical: The AI encodes the 613 mitzvot in Prolog format, builds Talmudic training pairs, and constructs the te'amim reader. The AI reads Torah for all of this.

Community: Find the 70. Find the kohanim. Find the gabbai. Begin now.

Shema: Te'amim reader is highest-priority — build alongside Shoresh Engine in Phase 1.

---

Document Registry

#	Document	Torah Parallel	File
1	Bikoret PRD	V'asu li mikdash	bikoret_prd.md
2	Architecture Specification	Vessel design (Shemot 25-27)	bikoret_architecture.md
3	Vayakhel — Construction Specification	Exact measurements (Shemot 25-31)	bikoret_vayakhel.md
4	Tamid — Operations Manual	Bigdei Kehunah / continual service	bikoret_tamid.md
5	Pekudei — Economic Model	Machatzit HaShekel / accounting	bikoret_pekudei.md
6	Terumah — Materials (this document)	Terumah list (Shemot 25:2)	bikoret_terumah.md