arborist

An arborist for trees and forests of cross-linked information.

Arborist ingests documents into a content-addressed, Merkle-committed SQLite store, distills them into recursive cores, and answers questions over the resulting corpus via an OpenAI-compatible LLM. Every cached answer carries a verifiable Merkle proof tying it back to its source documents — Merkle-AGI v9.8 / Merkle Providence Reverse RAG, runnable end-to-end.

Quickstart

Two end-to-end paths. Pick whichever corpus you want first; both share the same query, verify, falsify, and inspect surfaces.

Windows / no make? Every make <target> below works verbatim — a bundled make.bat shim forwards to a pure-stdlib runner (tasks.py). The quickstart needs only Python 3.10+ (no make, no bzip2, no curl). In cmd type make query Q="…"; in PowerShell use .\make.bat query Q="…" (or py -3 tasks.py query Q="…" directly). See Setup → Windows. The Makefile remains the canonical path on Linux/macOS.

A. Wikipedia 2003 (the canonical bootstrap dataset)

make bootstrap                          # one-time: venv + dev extras
make fetch-cur                          # download the 2003-05-16 snapshot (~82 MB)
make ingest-cur-attached                # ~3 min: 128k articles, 4 parallel shards
make distill-shards-parallel            # surface → core (first-sentence)
make distill-shards-tfidf-parallel      # core → keyword sets for retrieval
make query Q="What is anarcho-capitalism?"

A Hermes-3 inference runs against the local corpus, picks 4–8 source articles by Merkle root, and returns an answer plus a verifier label that names what the lexical verifier could confirm. The current four-rung ladder for claim-lattice modes is POINTER-LINKED → ANCHOR-WARRANTED → EVIDENCE-WARRANTED → UNGROUNDED (with -PARTIAL suffix on HYBRID). Repeat the same question and a cache hit replays in ~100 ms.

B. Crawl any live website and query it

make bootstrap-crawler                                            # one-time: install [crawler] extras
make crawl-ingest URL=https://russell.ballestrini.net DEPTH=2     # BFS + ingest
make query Q="who is Russell Ballestrini?"                        # cross-shard query — picks up the new shard automatically

The crawl shard is named after the seed hostname (crawl_russell_ballestrini_net.db) under ~/.arborist/shards/. Add FAST=1 for aggressive crawling against your own sites; MAX=N to cap discovery; DEPTH=N to bound BFS. Robots Disallow is always honored. After ingest, make recrawl-check DOMAIN=russell.ballestrini.net does a conditional-HEAD freshness probe per page.

After the answer

make inspect KEY=<cache_key>           # sidecar: classify each unverified span
make falsify KEY=<cache_key> REASON='…' # mark wrong, keep history
make burn    KEY=<cache_key> REASON='…' # delete (kindergarten only — refuses if children exist)

make help lists every target.

Setup

Get the source

git clone ssh://git@git.unturf.com:2222/engineering/unturf/arborist.git
cd arborist

HTTPS variant if SSH isn't set up:

git clone https://git.unturf.com/engineering/unturf/arborist.git

Install prerequisites

The quickstart needs only Python 3.10+ (SQLite ≥ 3.35 ships with CPython; the 2003 dump is decompressed in-process via Python's bz2). GNU make, curl, and bzip2 are conveniences for the Makefile path on Linux/macOS — on Windows the bundled make.bat / tasks.py need none of them (see Windows).

macOS (Homebrew)

xcode-select --install        # if you don't already have CLT
brew install python@3.12 git make

Apple's make is GNU make, no extra step needed. bzip2 and curl are bundled.

Ubuntu / Debian

sudo apt update
sudo apt install -y git python3 python3-venv python3-dev build-essential curl bzip2

22.04 ships Python 3.10; 24.04 ships 3.12 — both work.

Windows (native — no make, no WSL)

Install Python 3.10+ from https://www.python.org/downloads/ and tick "Add python.exe to PATH". That is the only prerequisite — the runner uses stdlib urllib for the download and stdlib bz2 for decompression, so there is no make, curl, or bzip2 to install.

The bundled make.bat makes every documented make <target> work as-is:

:: in cmd.exe (run from the repo root)
make bootstrap
make fetch-cur
make ingest-cur-attached
make distill-shards-parallel
make distill-shards-tfidf-parallel
make query Q="What is anarcho-capitalism?"

# in PowerShell, prefix with .\  (PowerShell doesn't search the current dir)
.\make.bat bootstrap
.\make.bat query Q="What is anarcho-capitalism?"

Or call the runner directly with the py launcher: py -3 tasks.py <target>. py -3 tasks.py help lists every supported target. The venv lands at .venv\Scripts\ (vs .venv/bin/ on POSIX); tasks.py resolves that automatically. Override the interpreter used to build the venv with the ARBORIST_PYTHON env var (e.g. set ARBORIST_PYTHON=py -3.12).

tasks.py covers the quickstart subset — bootstrap, fetch/ingest/distill, query/inspect/falsify/burn, the crawl path, stats/verify/search, clean. The bench / π* / NLI / textbook / docs targets stay Makefile-only; for those (or to use the canonical Makefile) install WSL2 and follow the Ubuntu instructions:

wsl --install -d Ubuntu-24.04

OpenBSD

pkg_add git python-3.12 gmake curl

OpenBSD's default make is BSD make. Arborist's Makefile uses GNU-make features (?=, conditional functions). Substitute gmake for make in every command, e.g. gmake bootstrap, gmake query Q='…'.

Bootstrap

make bootstrap

Creates .venv/, installs the package in editable mode with the [dev] extras, and exposes arborist at .venv/bin/arborist (.venv\Scripts\arborist.exe on Windows). No system-wide install. Re-running make bootstrap is a no-op if the venv is up to date.

After bootstrap, every workflow lives behind a make target. Run make help to list them (Windows: py -3 tasks.py help for the quickstart subset).

Data: Wikipedia 2003-05-16 (Phase III SQL dump)

The 2003 dataset lives at https://dumps.wikimedia.org/archive/2003/2003-05-16/en/ as three files. This is a MySQL extended-INSERT format dump; for XML-format dumps from 2006 onward see the Phase IV section below.

file	size	what
20030516_cur_tablesql.bz2	82 MB	one row per article: snapshot of every Wikipedia page on 2003-05-16
old_tablesqlbz2.1	640 MiB	first half of the revision-history table (split bzip2 stream)
old_tablesqlbz2.2	252 MiB	second half — cat them together to decompress

make fetch-cur                # snapshot only (~82 MB on disk)
make fetch-old                # full history (~1.4 GB on disk after concat)
make fetch                    # both

Files land in data/. Re-running is idempotent (curl skips if already present).

Sharded ingest (the canonical path)

Per-shard SQLite files, no write-lock contention. Each shard process owns its own DB; cross-shard reads attach all shards as UNION ALL views.

make ingest-cur-attached SHARDS=4    # cur snapshot, 4 parallel shards (~3 min)
make ingest-old-attached SHARDS=4    # full history, ~30–40 min

Shards land in ~/.arborist/shards/. Override with SHARDS_DIR=/path/to/somewhere.

Resumable

Add --resume (or just re-run the make target — --resume is the default for attached ingests). Each shard tracks its own high-water mark in meta; an interrupted ingest picks up where it left off without re-hashing already-cached docs.

Single-DB ingest (simpler, smaller)

For experiments under a few thousand docs, a single SQLite file is fine:

make ingest-cur INGEST_LIMIT=1000       # one DB at $(DB), default ~/.arborist/arborist.db
make ingest-cur-parallel SHARDS=4       # 4 processes, one shared DB (WAL serialized)

Distillation (cores feed retrieval)

Two distillers ship: first-sentence-v1 (one sentence per chunk) and tfidf-keywords-v1 (top-K distinctive terms per doc). Cores are Merkle-signed back to their source docs and serve as enriched titles for retrieval.

make distill-shards-parallel              # first-sentence cores, per-shard
make distill-shards-tfidf-parallel        # TF-IDF cores, per-shard

Run both — they generate independent cores per source. TF-IDF cores let neologisms (like a personal term that never appears in any title) match retrievals via keyword overlap.

Data: Wikipedia 2010-11 (and other Phase IV snapshots)

In 2006 MediaWiki swapped its dumps from MySQL INSERT INTO cur syntax to XML. Arborist reads both — the SQL path above for 2003-2005 cur dumps, and a streaming XML path for any dated snapshot in https://dumps.wikimedia.org/archive/. The largest single snapshot in that archive is enwiki 2010-11-08:

file	size	what
enwiki-20101011-pages-articles.xml.bz2	6.2 GB	latest revision of every main-namespace article on 2010-11-08 (~3.4M pages, ~1.9M after redirects)
enwiki-20101011-abstract.xml	2.9 GB	first-paragraph abstracts only — pre-distilled summaries at ~1/100th the chunk volume

Other useful dated snapshots in the archive: 2006-07 (1.8 GB), 2006-12 (1.9 GB), 2010-03 (varies by language). All work via the same source class.

# defaults target enwiki 20101011 (2010-11)
make fetch-xml                # ~6.2 GB compressed download
make ingest-xml-attached SHARDS=4    # ~2 hours sharded; ~95 GB on disk after

# pick any other snapshot by overriding the date variables
make fetch-xml WP_XML_YEAR=2006 WP_XML_MONTH=2006-07 WP_XML_DATE=20061104

# abstract feed (one-paragraph summaries, full coverage at ~5-10 GB total)
make fetch-abstract
make ingest-abstract

The XML source streams .xml.bz2 directly via iterparse with bounded memory (each <page> is processed and cleared). Same shard / resume / Merkle contract as the SQL source. Title-prefix namespace filtering kicks in for older export schemas that omit per-page <ns>.

To ingest historical revisions instead of just the current snapshot, point WP_XML at a pages-meta-history.xml.bz2 file and use make ingest-xml-history — the source emits one Document per revision and arborist's prior-doc detection chains them with supersedes edges.

Data: personal Grok export

If you have an xAI data-export bundle, point GROK_EXPORT at its root directory (the one containing ttl/30d/export_data/<user-id>/):

export GROK_EXPORT=$HOME/Downloads/<your-user-uuid>
make ingest-grok-attached            # conversations  -> $(SHARDS_DIR)/grok.db
make ingest-grok-media-attached      # media prompts  -> $(SHARDS_DIR)/grok.db

Both walk the export tree, find prod-grok-backend.json, and yield one Document per conversation (or media-generation post). Conversation titles, full message text, and turn ordering are preserved. Each becomes a normal queryable doc in the cluster — your prior chats become memory the corpus can consult.

--resume is the default for these targets. Re-run any time to pick up new exports.

Data: git and Mercurial repos (self-play)

Arborist can consult itself. Point a source at any local clone and every text file at HEAD becomes a queryable Document; re-ingesting after new commits chains old → new via supersedes edges, so the audit trail grows alongside the repo:

make ingest-self                         # this arborist tree, into ~/.arborist/shards/arborist-self.db
make ingest-git GIT_REPO=/path/to/repo   # any other git clone
make ingest-hg  HG_REPO=/path/to/repo    # mercurial flavor

URI shape: git://<repo-name>/file/<relative-path> (no commit hash — that's what enables the supersedes chain on re-ingest). Binary files are skipped (NUL-byte heuristic + UTF-8 decode probe). extra carries the current commit hash, timestamp, and subject for informational purposes; the cryptographic identity is the content-derived document_root as for every other Document.

Data: live websites (the crawler)

Arborist can BFS-discover and ingest a website starting from a seed URL, respecting robots.txt and crawl delays. The crawler is off by default — heavy deps (aiohttp, bs4, lxml, mwparserfromhell, etc.) ship as the [crawler] extras and aren't pulled into the default test suite.

make bootstrap-crawler                                              # one-time, install [crawler] extras
make crawl-ingest URL=https://russell.ballestrini.net DEPTH=2       # BFS + ingest the discovered pages

The shard filename is derived from the seed URL's hostname so cross-shard query picks it up automatically:

URL=https://russell.ballestrini.net  →  $(SHARDS_DIR)/crawl_russell_ballestrini_net.db

After ingest, the same make query searches across crawl shards alongside Wikipedia, Grok, and self-play sources:

make query Q="who is Russell Ballestrini?"

Knobs:

variable	default	what
URL=	(required)	seed URL; BFS stays on its hostname (no subdomain crossover)
DEPTH=	2	max BFS depth from seed
MAX=	0	cap discovery at N URLs (0 = no cap, depth is the only bound)
FAST=1	unset	flip the verbatim AsyncWebFetcher into fast_mode: 5s timeouts, CPU×3 parallel page workers, ignore robots.txt crawl-delay. Disallow is still honored. Use only against sites where aggressive fetching is acceptable.
CRAWL_SHARD=	derived from URL	override the destination shard path

Feeds and sitemaps (atom, rss, sitemap.xml, wp-rss2.xml, etc.) are skipped at ingest — they're discovery infrastructure, not knowledge. ETag and Last-Modified per page are captured so a future probe can ask "does this need recrawling?" without re-fetching bodies:

make recrawl-check DOMAIN=russell.ballestrini.net

Conditional If-None-Match / If-Modified-Since HEAD requests classify each ingested doc as fresh (304), stale (200), gone (404/410), or unreachable. One tiny round trip per URL with no body transfer when content's unchanged.

The crawler is a verbatim lift from ~/git/agents.ai.unturf.com/core/ (provenance documented in arborist/sources/crawler/__init__.py); arborist-side changes drop the chat-bot fetch triggers and skip web_cache_manager.py (arborist has its own content-addressed cache). Run make test-crawler for the lift's own tests.

Asking the corpus

make query Q="What is the philosophy of stoicism?"
make query Q="tell me about permacomputer ?"
make query Q="…" QUERY_TOP_K=12        # widen the source set
make query Q="…" JSON=1                # raw record (cache_key, merkle_proof, timings, full sources)
make query-dry Q="…"                    # assemble context but skip the LLM call

Default render is human-readable: question, audit-mode summary line, answer, sources list, unverified spans, short cache_key. JSON=1 gives the full record. Same trailing-question-mark question deduplicates (question_hash strips trailing .?!,;:？！。、… after lowercasing).

The query path:

1. Search — FTS5 (body) + SQL LIKE (title) + JOIN over derivations (TF-IDF core keywords) across every shard. Three accept paths to the relevance filter.
2. Concept overlay — per-shard concept_relations SQLite table (corpus-derived, not hand-curated). Synonyms widen retrieval; rivalries narrow it unless the query uses comparative phrasing ("compare X vs Y"). Built-in extractor link_reciprocity_synonym reads the existing edges table for reciprocal A↔B link pairs and emits synonym edges between their title-tokens. ~1.6% storage tax measured on 6 GB Wikipedia.
3. Context assembly — top-K sources concatenated up to a 60 KB budget. Wikitext is stripped to plain prose via arborist.wikitext.to_base() (the corpus stores raw [[wikilinks]] so the link graph is recoverable on demand; the LLM and verifier both see clean prose).
4. LLM — Hermes-3 with strict attribution rules in the system prompt + a user-turn grounding reminder.
5. Verifier — every claim runs through a layered lexical check; the result rolls up into the v9.8 trichotomy (audit_mode ∈ STRICT / HYBRID / UNGROUNDED) at the schema layer AND a four-rung display ladder at render time (POINTER-LINKED → ANCHOR-WARRANTED → EVIDENCE-WARRANTED → UNGROUNDED). See below.
6. Cache — the v9.8 8-dim cache_key (source_root | question_hash | model_profile | conversation | governance_policy | schema | canonicalization | chunking) keys the answer in qa.db. Cache hits replay in ~100 ms. Per-phase timings in every result.

LLM endpoint defaults to https://hermes.ai.unturf.com/v1 (Hermes-3 Llama-3.1-8B, 82K context, no auth). Override:

export ARBORIST_LLM_ENDPOINT="https://your-vllm.example/v1"
export ARBORIST_LLM_MODEL="meta-llama/Llama-3.1-70B-Instruct"
export ARBORIST_LLM_API_KEY="..."

Verifying answers (audit modes & label ladder)

Two layers of labels stack on every answer:

Schema layer — v9.8 trichotomy (audit_mode column, persisted, drives cache lookups & audit chain):

mode	meaning
STRICT	every evidence unit verifies against context
HYBRID	some claims source-grounded, some emerged from training
UNGROUNDED	no evidence, or none verifies — purely emergent

Display layer — four-rung ladder (claim-lattice modes only; renderer-only transformation, schema unchanged):

rung	what's actually proved
EVIDENCE-WARRANTED	pointer verified + warrant ran & passed + no soft demotes
ANCHOR-WARRANTED	pointer-linked + warrant passed; soft-demote violations present
POINTER-LINKED	pointer/source/chunk verified, but warrant either didn't apply or failed for some claim
UNGROUNDED	no verified pairs

HYBRID gets a -PARTIAL suffix on whichever rung applies. The point of the display ladder: STRICT in claim_lattice mode is NOT "the answer is correct" — it's "every pointer resolved to a valid evidence object whose source_role is allowed AND citation-coverage passed." The display label spells out the actual property so users don't read STRICT as full semantic entailment.

Verifier strategies run in sequence; first to find evidence classifies. Each is lexical (substring or token-coverage), never embeddings — soft signals stay out of the proof path.

#	strategy	what triggers it
1	quote	model wraps claims in "..."; each verbatim-substring tested
2	span	bullet/sentence units substring-tested as fallback
3	entity	multi-word proper nouns tested with proximity-cluster gating
4	paraphrase	inside the span path: ≥85% token coverage on prose-shaped spans
5	claim_lattice	model emits claim text. [E1,E2] pointer-line OR {"claims":[{"text":..., "evidence_ids":[...]}]} JSON; verifier resolves pointers to runtime-built evidence objects & runs seven hard checks

The claim_lattice path runs seven deterministic hard checks: parser succeeded, evidence_id resolves, source_role allowed, claim text non-empty, citation coverage threshold, pointer count cap, anchor-class warrant. Anchor-class warrant composes five lexical anchor classes — proper-noun, date, entity-list, count (with digit↔word equivalence), and why-cause — each gated on either question shape, claim content, or both. See whitepaper §13.9 for the architecture.

Trailing (Source: https://...) parentheticals the model appends to verbatim source sentences are stripped before substring testing, so verbatim-with-citation no longer flags HYBRID.

unverified_quotes on each record is the corpus-growth signal — model output that didn't ground anywhere. arborist emergent --aggregate ranks them by frequency (the worklist of "things to ingest more sources for"). arborist reclassify re-runs the verifier against existing live records after corpus growth without any LLM call; HYBRID promotes to STRICT, UNGROUNDED to HYBRID, and one providence_reclassify audit event per change.

To dig into a specific record's unverified spans:

make inspect KEY=<cache_key>

Read-only sidecar diagnostic: pulls source chunks, classifies each unverified span as verbatim_in_base / trailing_artifact / paraphrase / partial_paraphrase / no_overlap. Writes nothing — sidecars never enter the v9.8 hard chain (that invariant is what keeps audit_mode a binary classification rather than a soft score).

Mesh / federation (off by default)

Optional gossip layer for peer-to-peer corpus sync. Default off — no code path touches the network unless mesh.enabled is set. See docs/mesh.md for protocol contract and the arborist mesh CLI subcommands.

Inspecting

make stats-shards                    # totals across shards
make analyze-shards                  # compression spectrum, depth histogram, audit chain integrity
make verify-shards                   # round-trip Merkle proofs on a random sample
make activity                        # recent Q&A + ingests + derives + falsifications (agent timeline)
make activity ACTIVITY_LIMIT=20

activity is JSON; pipe through jq to drill in.

Architecture

Generated API reference (every module, class, and function from docstrings): build with make docs-api (output at docs/_source/_build/html/) or browse the published Read the Docs site. Pipeline diagrams live in docs/diagrams/; render with make docs.

Tests & bench

make test                # 2000+ tests, default suite, ~60s sequential / ~10s with pytest -n auto
make test-crawler        # opt-in: tests for the verbatim crawler lift
make test-live           # gated: live QA quality fixtures against Hermes (~1 min)
make bench               # ETL throughput across configs (serial / shared-WAL / attached)
make bench-qa            # full QA-quality sweep, sample-shuffled, 71q × 3m × 3n at c=4
make bench-qa-smoke      # 5-question smoke fixture for prompt-iteration loops (~1-3 min)

The default suite never hits the network. The crawler suite is gated behind make bootstrap-crawler (installs the [crawler] extras).

make bench-qa writes JSONL + markdown into bench/qa_results/<utc-stamp>.{jsonl,md} (gitignored); design-log entries live in docs/qa-modes-bench-<date>.md. The bench is stop/start-able via --resume <jsonl-path> (same --seed required for shuffled-task-order alignment). BENCH_QA_CONCURRENCY=N Makefile variable overrides the default 4. See docs/bench-maxing.md for the full speed playbook.

License

License: AGPL-3.0-only · This algorithm, its implementation, & all associated code carry the GNU Affero General Public License v3.0 (only). You may use, modify, & distribute under those terms. No proprietary relicensing exists.

Permacomputer Preamble

This is free software for the public good of a permacomputer hosted at permacomputer.com, an always-on computer by the people, for the people. Durable, easy to repair, & distributed like tap water for machine learning intelligence.

Our permacomputer is community-owned infrastructure optimized around four values:

• TRUTH — First principles, math & science, open source code freely distributed.
• FREEDOM — Voluntary partnerships, freedom from tyranny & corporate control.
• HARMONY — Minimal waste, self-renewing systems with diverse thriving connections.
• LOVE — Be yourself without hurting others, cooperation through natural law.

NO WARRANTY. Software is provided "AS IS" without warranty of any kind. Full text in LICENSE.

(Verbatim from the Merkle Providence Reverse RAG whitepaper, April 2026.)