arborist/CLAUDE.md

Agent Blackops — arborist

This repo is operated by agent blackops for fox/timehexon on the unsandbox / unturf / permacomputer platform.

Identity shard: ~/git/unsandbox.com/blackops/BLACKOPS.md.

What arborist is

A content-addressed, Merkle-committed document store. Implements the runtime spec from Merkle Providence Reverse RAG (April 2026 whitepaper) scaled up to the Merkle-AGI v9.8 admissibility ledger.

Three layers stacked on one SQLite file:

1. Surface — ingested documents (Wikipedia dumps, HTML pages, anything with a URI). Chunked, Merkle-rooted, FTS5-indexed.
2. Core — distilled documents Merkle-bound back to surfaces via per-chunk inclusion proofs in derivations.proof_blob. Recursive.
3. Providence cache — Q&A records keyed on the v9.8 8-dim invariant. Each record carries audit_mode (STRICT / HYBRID / UNGROUNDED) decided by the verifier in arborist/qa/verify.py. For claim-lattice modes the renderer maps that token to a four- rung ladder (POINTER-LINKED → ANCHOR-WARRANTED → EVIDENCE-WARRANTED; ENTAILMENT-VERIFIED reserved); UNGROUNDED below all rungs. The schema column stays unchanged — programmatic callers see the trichotomy, human-facing surfaces see the ladder. See arborist/cli.py:_render_audit_label.

Source papers

• ~/git/unfirehose-nextjs-logger/whitepaper/merkle-providence-reverse-rag-whitepaper.rst — canonical whitepaper (rst → PDF). Edit here, not the PDF.
• ~/Downloads/merkle-agi-dag_v7.txt — formal substrate (TLV encoding A1, public quantization A2, collision-resistant hash A3, theorems T1–T5).
• ~/git/proxy.unturf.com/pkg/verified/merkle.go — fox's existing Go merkle. Arborist Python ports its conventions exactly.
• ~/git/proxy.unturf.com/docs/merkle-tree.md — convention reference.

Architecture

arborist/
├── merkle.py        # Python port of proxy.unturf.com Go conventions
├── store.py         # v9.8 SQLite schema, audit-chain helpers
├── document.py      # Document, Edge, Chunker (TokenChunker default)
├── source.py        # Source ABC: iter_documents()
├── ingest.py        # batched normalize → chunk → merkle → upsert
├── evict.py         # hot↔cold + rehydrate (v9.8 falsification on drift)
├── compress.py      # chunk pack/unpack (zstd dictionary trained on corpus)
├── snapshot.py      # snapshot.db creation + load
├── journal.py       # NDJSON unfirehose journal sink
├── wikitext.py      # to_base(): wikitext → plain prose (BASE_VERSION-pinned)
├── search/          # SearchBackend ABC + AuditMode + FTS5
├── sources/         # one file per corpus (wikipedia, html_page,
│                   # claim_pack, textbook_tex, grok, vcs, …)
├── distill/         # surface → core distillation (tfidf, first_sentence)
├── qa/              # Q&A: 8-dim cache_key + Merkle-bound answers
│   ├── client.py    #   ChatClient + StubClient + OpenAICompat
│   ├── keys.py      #   cache_key, question_hash (pure functions)
│   ├── verify.py    #   layered verifier + claim-lattice verifiers (G0)
│   ├── evidence.py  #   EvidenceObject + spotlight excerpt
│   ├── inspect.py   #   read-only sidecars (deflection, title-relevance)
│   ├── dag.py       #   per-run Merkle-DAG (7/8 quote · 9/10 CTI · 3 reject)
│   ├── quantifier.py        # broad-quantifier classifier (#000008 P1)
│   ├── model_profiles.py    # per-model claim-cap profiles (#000008 P2)
│   ├── quantifier_reminder.py # broad-query reminder text (#000008 P3)
│   ├── canonical_cache.py   # canonical-projection persistence (#000027)
│   ├── witness.py           # multi-witness fan-out (#000028)
│   ├── warrant_resolver.py  # claim-pack warrant chain resolution (#000031)
│   ├── crosslang.py         # cross-lang guard: signal + es stoppack (#000001 §7 P0)
│   ├── mt/                  # Operation Sandwich MT edges (#000056)
│   │                        # opus-mt es/fr/ru↔en + entity_mask
│   └── runner.py    #   ask(): cache → infer → verify → write
├── concepts/        # corpus-derived synonym + rivalry layer (#000018 sib.)
├── pi_star/         # canonical projection π* registry (#000015)
│                    # arithmetic@v1, logic-kernel@v1, algebra-symbolic@v1, …
├── memory/          # MemoryRoot lifelong-learning audit chain (#000017)
├── selfmodel/       # SelfModel snapshot + falsify (#000014)
├── capital/         # 8-form capital ledger (#000020)
├── substrate/       # Merkle-AGI substrate primitives (paper-spec'd)
│   ├── anchor_prg.py       #   φ_PRG HMAC-SHA-512 (v7 §9.10; #000035)
│   ├── fork_score.py       #   ScoredFork decision fn (v8; #000012 P1a)
│   └── weights.py          #   ForkScore weight set
├── world/           # v7-W spatial-temporal substrate reservation (#000013;
│                   # namespace stub; future kernels under world/pi_star/,
│                   # world/frontier/, world/adapters/)
├── mesh/            # mesh wire format + group-key state machine
└── cli.py           # ingest / search / verify / stats / distill /
                     # evict / rehydrate / ask / providence / emergent /
                     # reclassify / inspect / analyze / canon / sweep /
                     # warrant-resolve / alias / capital / selfmodel

Dir naming convention. Topic-named, never version-prefixed. The substrate-paper version (v7 plastic-training, v8 selection/consensus, v9 falsification controller, …) and the live SQLite schema version (v9.8) are two unrelated numbering schemes that share decimals; version-prefixed dirs (arborist/v7/, arborist/v8/) were tried 2026-05-10 and retired the same day because readers asked "is this schema-v7 or paper-v7?". Substrate-paper-spec'd primitives now live under arborist/substrate/; topic dirs (capital/, memory/, selfmodel/, concepts/, pi_star/) hold cross-version mechanisms.

Build, test, run

Every workflow is a make target. Bare python is not the user interface. See the Makefile for the full list.

make bootstrap           # venv + editable install with [dev] extras
make test                # pytest -q
make all                 # bootstrap + fetch-cur + ingest-cur + verify + stats
make verify-shards       # round-trip Merkle proofs (cross-shard sample)
make analyze-shards      # cross-shard compression + audit integrity
make chain-check-shards  # audit-chain break count per shard (0 = intact)
make query Q="..." [JSON=1 BURN=1 K="extra retrieval keywords" ANSWER_MODE=…]
make bench-qa            # QA-quality sweep (live LLM)

# Textbook surface-ingest layer (#000031 — PD/open-licensed
# math/logic/CS textbooks for warrant promotion):
make textbooks-summary   # license + URL counts per manifest entry
make crawl-textbooks     # BFS-crawl every entry with crawl_url
                         #   → ~/.arborist/crawl/textbook_<id>.db
make textbooks-tex       # PG LaTeX-source ingest (Hilbert, Boole)
make textbook ID=<id>    # ingest one textbook by manifest id
                         # (idempotent at DB layer)
make textbooks-base-knowledge  # bulk: Cantor + De Morgan + Russell
                               # IMP + Judson (the four 2026-05-09
                               # base-knowledge additions)

# Per-textbook convenience targets (one per active manifest id):
make textbook-bogart textbook-keller-trotter textbook-levin
make textbook-aristotle-prior textbook-aristotle-posterior
make textbook-newton textbook-morin
make textbook-judson textbook-cantor textbook-demorgan
make textbook-russell-imp textbook-russell-pom
make textbook-laplace textbook-pm textbook-grinstead-snell
make textbook-hilbert textbook-boole textbook-peano textbook-dedekind
make textbook-plfa textbook-sf-lf

# Claim-pack warrant-chain resolver (#000031 Phase 2 + 2.5):
make sweep TARGET=warrants  # warrant-resolve --use-aliases --write
arborist warrant-resolve --use-aliases --write
                         # → 92 / 92 (100%) coverage as of 2026-05-10:
                         # 18 textbook substrates + 74 fox-decided
                         # citation-aliases + 13 term-aliases.
                         # (alias counts grow as fox adds substitutions
                         # — `arborist alias citation list | jq length`
                         # for live count). Per-pillar 13/13 · 10/10 ·
                         # 13/13 · 18/18 · 5/5 · 5/5 · 14/14 · 14/14.
arborist alias citation list  # see substitutions
arborist alias term list      # see vocabulary aliases

Hygiene after any state-changing op (rebuild, reclassify, hash bump, mass falsify): make chain-check-shards first (every shard should report 0), then make analyze-shards for the spectrum + chain audit. Chain breaks are the loudest possible signal.

Schema invariants (do not break)

• v9.8 8-dim cache_key: source_root | question_hash | model_profile_hash | conversation_hash | governance_policy_hash | schema_version | canonicalization_version | chunking_version. Bumping any one invalidates prior records on lookup.
• falsification_state ∈ {live, failed, stale, quarantined}. Cache lookups filter on state='live'. Drift → stale.
• Audit chain: every state-changing op writes one row in audit_events with event_hash = sha256(prev || canonical(body)). Verified by make chain-check-shards. Use arborist.store.append_audit — never insert into audit_events directly.
• Cores never evict. evict_to_cold only touches kind='surface'.
• Idempotent re-ingest. Same content → same document_root → no-op insert. Same URI + different content → new doc + supersedes edge (lossless history).

Conventions (do not silently change)

Each rule below has full rationale in the named source file. Don't revert without reading why. When in doubt, walk the Five-step algorithm first.

• Merkle conventions: non-commutative HashCombine prefix 0x03, leaves 0x00, odd-element rule = self-duplicate (NOT zero-pad). MerkleProof.siblings carries is_left flag — never sort lexically. See arborist/merkle.py.
• Versioned defaults: tok-512-v1 (chunker), norm-v1 (canonicalization), v9.8.0 (schema), wikitext-base-v1 (prose). Changing any default stales every prior cache record. Add a new name instead.
• question_hash is dedup-mode-aware (strict | equivalence_class); folds into governance_policy_hash. JIT fidelity parameter on query()/ask() decouples lookup tolerance from write policy. See arborist/qa/keys.py.
• audit_mode is decided by the verifier, never asserted. Four layered strategies tried in order, first to find evidence classifies: quote (sequential pair-matching, NOT regex — prevents phantom inter-pair captures), span (verbatim line match), entity (proximity-clustered proper nouns; entity_policy ∈ {strict, hybrid, drop, proximity}), paraphrase (token-coverage, prose-shaped only; verifier_method='paraphrase'). Trichotomy: STRICT = every unit verifies, HYBRID = mixed, UNGROUNDED = none. Never overclaim. See arborist/qa/verify.py.
• Verifier stays binary; falsifications carry soft signal. No per-quote diagnosis fields on hard verifier output. Sidecars (arborist.qa.inspect.diagnose_*, arborist inspect --cache-key X) classify unverified spans, deflection, title-relevance — never write to providence_cache or audit_events.
• Trailing-citation strip: _strip_trailing_citation peels one trailing parenthetical at end-of-span (gated on a citation cue or URL) before substring testing. See arborist/qa/verify.py.
• Soft hash vs hard hash: hard = SHA-256 (commitments, proofs, cache_key); soft = embeddings/TF-IDF/similarity (training, ranking, distillation). Soft never enters proof path.
• Cross-language = the sandwich, MT on the edges only: translate query in (retrieval + LLM prompt) → English answer → the byte-identical verifier grounds English-vs-English → translate the verified answer out as display-only (display_*, banner- labelled, zero grounding — the _render_audit_label projection discipline). Translation NEVER re-enters the verifier (that's the #000049 model-in-proof-path cage). Invariants: question_hash = the user's original question (untranslated); verifier_policy_hash unchanged; MT engine identity binds into RetrievalPlan.mt_* (run- DAG), NOT governance_policy_hash (the flag moves it like any policy flag — correct cache partition — but that hash covers the whole policy, keys.py:182; don't mistake "no new governance field" for "governance untouched"). Engine = local [mt] opus-mt, hash-pinned, never Hermes-for-translation, never an API. Default OFF (crosslang_guard_enabled P0; crosslang_translate_enabled Sandwich; crosslang_entity_mask default-OFF — measured net-negative, kept only behind the flag). Measured net win over the real "nothing" baseline (raw es → noise/UNGROUNDED): es ≈0 % → 71 % grounded; the −14 pp vs English is the cost of a new capability, not a regression (compare to no-cross-lang, never to native English). The recall lever (entity-preservation) is still open — entity_mask v1 failed at bench scale; corpus-title anchoring is the untried idea. See arborist/qa/crosslang.py, arborist/qa/mt/, #000001 §7, #000056 §9.
• Three answer modes: policy["answer_mode"] ∈ {"quote", "claim_lattice_pointer", "claim_lattice"}, default "quote". Bench 2026-05-02T15:07Z on Hermes-3-8B (post-Sprint-1b/2, n=3 × 71 questions, sample-shuffled @ c=4): quote 0.54 strict-rate, pointer 0.20, JSON 0.42. Quote leads on raw lexical grounding; JSON leads among lattice modes. Per-mode peak buckets: quote 8-16KB (0.58), pointer 16-32KB (0.20), JSON 32-64KB (0.48) — these drive max_context_chars_by_mode. 99% directive coverage (D2/D3/D4/D6/D7) on lattice modes. Both lattice modes share verifier_method="claim_lattice"; answer_mode on the run-DAG
  • json_fixups disambiguate. Each mode folds into governance_policy_hash. See arborist/qa/verify.py, docs/qa-modes-bench.md.
• Four-rung ladder (lattice-mode display layer): POINTER-LINKED → ANCHOR-WARRANTED → EVIDENCE-WARRANTED → (ENTAILMENT-VERIFIED reserved); UNGROUNDED below. Each rung names a strictly stronger property the lexical verifier could confirm. WARRANT_MISSING drops to POINTER-LINKED; soft-demote violations (LAZY_ANCHOR_DEMOTED, POINTER_OVERFLOW_TRIMMED, TOO_MANY_CLAIMS, BARE_NAME_CLAIM, TITLE_MISMATCH) cap at ANCHOR-WARRANTED. Quote / span / entity / paraphrase modes keep their original audit_mode tokens (those verify against pinned spans, not synthesis). Pure render-layer — cache_key, governance_policy_hash, & all programmatic callers see the underlying audit_mode unchanged. See arborist/cli.py:_render_audit_label.
• Claim-lattice-pointer mode (G0 / CTI): runtime mints pointer_id (E1, E2, … — what the model sees) and content-addressed evidence_id (what the cache & run-DAG store). Renderer interpolates literal source spans via _spotlight_excerpt. Synthetic-elision-by- construction-impossible — model never types the quote string. 9-stage run-DAG. See arborist/qa/evidence.py, docs/cti-architecture.md.
• Claim-count ceiling (TOO_MANY_CLAIMS): default 12 per answer. Catches "tell me all there is to know about X" runaway. Demotes STRICT → HYBRID without truncating. Folds into governance_policy_hash. See arborist/qa/verify.py.
• Broad-quantifier preflight guard (Ticket #000008): pure lexical classifier (arborist/qa/quantifier.py) maps a question onto a 10-rung intensity ladder (ABSENT → SINGULAR → … → ALL → COMPREHENSIVE). Per-model arborist/qa/model_profiles.py PROFILES dict picks a per-call claim-cap from the (intensity, model) pair; arborist/qa/quantifier_reminder.py synthesizes a one-line user-turn reminder for broad questions. Six-level disable hierarchy (per-test, per-call CLI, per-phase policy, per-mode, per-model, master-via-governance-hash). 7 policy fields fold into governance_policy_hash so flipping any of them invalidates prior records. Defaults preserve dry-run discipline: quantifier_guard_apply_caps=False, quantifier_reminder_enabled=False, quantifier_reject_broad=False. CLI flags on arborist query: --no-quantifier-guard, --allow-broad, --reject-broad, --apply-quantifier-caps. Bench A/B (2026-05-03, n=3 × 9 broad questions × 3 modes): reminder default-on supported (FORMAT_COLLAPSED −100%, NO_EVIDENCE_POINTER −33%, JSON UNGROUNDED −22pp); cap default-on for JSON only (+14pp STRICT-rate, no gain on pointer); cap+reminder best on pointer mean ratio (0.684) but not strictly best on JSON STRICT-rate. See docs/tickets/ticket-000008-broad-quantifier-preflight-guard.md §12 for the four-cell A/B data.
• Wikitext base prose: arborist/wikitext.py:to_base() runs before the LLM call AND inside verify_quotes so model and verifier see the same prose. Optional dep — graceful fallback when mwparserfromhell is missing.
• Deflection sidecar: diagnose_deflection(question, answer) detects topic-shift via subject-anchor heuristic (LAST content token in question must appear in answer). Suppressed for date / count / cause shapes ("when", "why", "how many"). See arborist/qa/inspect.py.
• Coherence sidecar (#000052 §3.1): diagnose_coherence(answer) in arborist/qa/inspect.py — per-sentence lexical shape check, no model. Emits kind ∈ {phrase_component_reuse, circular, vacuous, ok, empty}: circular = subject content-tokens ⊆ predicate's and the predicate leads with a subject token ("X is X"); phrase_component_reuse = the subject quotes a phrase & the predicate reuses one of that phrase's own tokens as a bare the/a/an <token> referent (the 2026-05-12 "the phrase 'Zionist entity' is used as the entity" field case — a token collision the verifier + deflection + title-relevance all pass and NLI returns neutral on); vacuous = predicate is only placeholder hypernyms
  • filler ("X is a thing"). Surfaced in inspect_cache_key + the arborist inspect human view (· incoherent: <kind>). Advisory sidecar — never writes providence_cache / audit_events / run_dag_root; a demote-only verifier hook is possible but deliberately not wired (would fold into governance_policy_hash).
• Title-relevance hard check (Rule 8): _claim_title_overlap in arborist/qa/verify.py. For each claim that resolved, at least one cited evidence's source title must share ≥1 stemmed content token with the claim text. When NO cited title overlaps, record a TITLE_MISMATCH violation & demote STRICT → HYBRID. Catches retrieval-driven hallucinations where the cited chunk's SOURCE is structurally unrelated to the claim's subject (2026-05-02 spin-glass case: claim about spin glass cited to a chunk from Quantum chromodynamics; span had incidental physics-vocab overlap, but the source title shared zero stems with the claim). Renderer surfaces a · title mismatch tail on the audit-line label alongside · warrant missing.
• Title-relevance sidecar (legacy diagnostic): diagnose_title_relevance(claim, cited_titles) in arborist.qa.inspect returns the same signal in dict form for per-cache-key inspection. Sidecar; never enters proof path. Pre-dates the Rule 8 promotion (2026-05-02).

Live endpoints

• LLM: https://hermes.ai.unturf.com/v1 (Hermes-3 Llama-3.1-8B-FP8- Dynamic on vLLM, 82K ctx, no auth). uncloseai.com is marketing only. Override via --endpoint or ARBORIST_LLM_ENDPOINT.
• Wikipedia dumps: https://dumps.wikimedia.org/archive/2003/2003-05-16/en/. robots.txt returned 404 → no rules.

Retrieval pipeline (arborist/qa/query.py)

Multi-stage. Each stage exists because something earlier wasn't enough; revert at your peril. Order:

1. Four parallel FTS5 search routes per shard, merged — body BM25, title-LIKE, core-keyword (TF-IDF cores), and phrase-pattern (verbatim n=5/n=6 sequences from the question). Phrase route closes the allusion gap (Orwell case: "always been at war" verbatim matches the 1984 article whose title shares zero tokens with the query).
2. Body-coverage sqrt rerank — counters BM25's short-doc bias.
3. Title-token boost — boost × overlap on title-token-matching hits.
4. _filter_by_title_relevance — four accept paths: title-token overlap, TF-IDF core match, body density, phrase match (accept-path 4 lets phrase-route hits with no title overlap survive).
5. Rivalry exclusion + synonym expansion (arborist/concepts/) — Intel-titled docs drop from AMD queries; reverse holds. Backed by the per-shard concept_relations SQLite table (corpus-derived, not hand-curated). 1.6% storage tax measured at backfill on 6 GB wiki — kept flat, no further compaction. (Storage choice rationale lived in docs/concept-relations-design.md prior to its deletion in bb6a89c — the design is now documented inline in arborist/concepts/extract.py + the data is self-describing.)
6. Stem-aware token matching — possessive / plural collapse (superman's → supermans → superman).
7. Per-source context cap — max_context_chars / top_k. Prevents one huge doc from monopolizing the budget.
8. Wikitext base prose runs on assembled context BEFORE the LLM.
9. Template-phrase stopwords (_FTS5_STOPWORDS and _TITLE_STOPWORDS must stay in sync) — strips tell show describe explain summarize say give list find make please all there know everything anything something so "tell me all there is to know about X" doesn't dilute query tokens.

--retrieval-keywords (CLI: K="...") lets an operator augment retrieval-side tokens without changing what the LLM sees as its question. Provenance gap on this is tracked in Ticket #000001.

Hot path / gotchas

• Hand-rolled wikitext parser (arborist/sources/wikipedia.py): char-position state machine, escape-aware, 4× faster than char- by-char loops via str.find + slicing. cProfile any change.
• PRAGMA synchronous=NORMAL per-connection in store.connect(). Safe under WAL. Don't downgrade without measured reason (~5× cost).
• HTML source has optional deps: pip install '.[html]' for selectolax. CLI surfaces --source html only if import succeeds.
• Background ingest/distill processes: stdout is buffered. Use export PYTHONUNBUFFERED=1 or python -u.
• Disk pressure: full cur ~2 GB; full old ~5–8 GB. df -h first.

Operational rules

• I propose, fox decides. Unsure = ask. Can't ask = stop.
• Never offer "stop here / take a break / fresh eyes tomorrow" as a next-action option. Fox keeps going. When proposing options at a decision point, list the actual moves (high-payoff vs low-risk cleanup, etc.); don't pad the menu with a no-op "we're done for today" choice. If a task is genuinely complete, say so flat — don't dress it up as a third option.
• No autonomous destructive ops (clean-data, clean-db, force-push, DB drops) without explicit instruction.
• Never add Co-Authored-By or "Generated with Claude" lines to commits. Code speaks for itself.
• Python only in arborist. No Rust, C, JS, or other languages inside this repo. arborist is the source-of-truth implementation; forks and downstream clients/servers in any language follow our schemas, canonical encodings, and audit protocols. Optional toolchains for ZK/world-model/etc. live in sibling repos (arborist-zk-bench, arborist-world, etc.) so a fresh checkout needs only python3.12 + venv + sqlite3.
• Always export PYTHONUNBUFFERED=1 for long-running processes.
• AUTOCOUNT discipline on numeric claims in docs/. Any numeric claim added to a doc (test count, fixture-row count, SQLite row count, filtered-row count) should be wrapped in an AUTOCOUNT tag at write time so future drift fires the regression test. Format: <!--AUTOCOUNT:metric:path-->N<!--/AUTOCOUNT--> where metric is one of tests / fixture-rows / db-rows / db-where. Tags are invisible in rendered markdown (GitHub strips HTML comments). Tags inside ``` fenced code blocks are auto-skipped (illustrative examples, not live claims). Closed-ticket "N tests pass" snapshots stay UNtagged (they're point-in-time historical records). Full discipline + 4-metric reference + future-metric recipe in docs/tickets/ticket-000044-autocount-doc-drift-discipline.md; harness at tests/test_doc_counts.py.
• Fail-closed. Cleanup crew, not demolition.
• DRY in context — single source of truth, no sprawl.
• Never say "AI" — always say "machine learning."
• Prefer "defect" over "bug."
• Check robots.txt before any web fetch the user didn't authorize.

Five-step algorithm

When proposing or evaluating change, walk these in sequence. Skipping a step makes the next ones expensive and the system worse.

1. Make the requirements less dumb. Every requirement gets a person's name, not a department. If you can't name who asked or which defect closed, the requirement is suspect.
2. Delete the part or the process. If you aren't putting back at least 10% of what you delete, you aren't deleting hard enough. Verifier-stays-binary and "no soft signals in hard chain" are deletion-first guardrails.
3. Simplify and optimize. Only after 1 + 2. Don't optimize a process that shouldn't exist.
4. Accelerate cycle time. Only after simplifying. "Don't dig the grave faster."
5. Automate. Last, not first. Hand-rolled before scripted, scripted before declarative, declarative before generated.

When in doubt, ask "have we tried deleting it?" before reaching for steps 3-5.

Bench-maxing — measure deltas, not opinions

Full discipline + worked examples in docs/bench-maxing.md. Headlines:

• Bench before AND after every change (n=3, signal under 5pp is noise).
• When a lever's failure class is sub-noise-floor, fix the instrument, not just the lever. The curated n=3 audit_mode bench can't resolve a class that is ≤3-5 of 75 q (four 2026-05-18 hypotheses died there — round-trip, entity-mask, disambiguation). Mine ground-truth-carrying questions from corpus titles (bench/mine_questions.py) and grade by deterministic retrieval recall@k via query --dry-run (bench/recall_at_k.py): no LLM, no verifier, no n=3 noise, no 5pp floor, scalable to the corpus. Resolves a single lever to ±1 question (worked example: numeral- fold recall 55→75%, a3ac653). Caveats: measures retrieval surfacing — necessary-not-sufficient for STRICT; mined fixtures are answerable-by-construction so they complement, never replace, the curated adversarial set (the verifier-honesty/trap gate). Conflating the two is itself a bench-maxing error.
• Report recall@1/@3/@k, not one lenient k — a coarse k hides a rank-only lift. recall_at_k.py returns the target's rank, so recall at every k is free from one retrieval. Measured 2026-05-18: accent-fold looked inert at recall@8 (95→98, noise) but the OFF baseline was recall@1 55% vs @8 95% — a too-lenient k flattered it to a near-ceiling and nearly got a real lever wrongly reverted. recall@1/@3 is the resolution that matters (primary-source selection keys on rank, not mere top-k presence). Prevalence ≠ miss-rate either: the corpus survey ranked accent #1 at 8.1% of titles, but the measured miss-rate (recall@1) is what decides — measure headroom, never rank candidates by raw prevalence.
• Fan out independent measurements; serial-by-caution is halting in disguise. Mined recall is deterministic per query (read-only, no LLM, no shared state) — concurrency cannot change which sources rank; the only risk is the per-probe timeout, and dry-run retrieval (~2 s) has huge margin under the 120 s cap. Run the whole fold-search backlog (accent / hyphen / honorific / …) as parallel background sweeps; only same-fixture A/B that toggles query.py state needs git-stash serialisation.
• Avoid negation in prompts (Hermes-3-8B inverts under attention).
• Bench is the scoreboard; live fixtures are the gates.
• Self-heal beats retry (preserve partial output, never fabricate).
• Honest verdicts beat optimistic ones (false-positive STRICT is corruption).
• Old maps vs runtime maps — when the model and runtime disagree, the runtime wins. Pointer IDs, runtime-interpolated spans, evidence maps, policy hashes, hard verifier checks all move authority OUT of the model's prior and INTO runtime artifacts.
• When researching a config/model decision: sweep wide, sweep deep, on real-traffic-sized samples — don't guess, and don't trust small denominators. Three times running (ticket #000049 §7 #18→#24) the answer flipped: a clean synthetic eval said one thing, bench-qa pipeline output said another; the default config said one thing, a {model × hyperparams} grid sweep said another; an n=1 FP sample said "this config passes", an n=3 FP sample said "that one collapses — a different config passes". So: (1) gate numbers come from bench-qa pipeline output, never contrived fixtures; (2) the sweep must be wide enough — multiple checkpoints, the full hyperparam grid; there is no "bigger model is better" law, the specific checkpoint and the score-shape (e.g. a single-threshold margin vs a two-threshold rule) dominate — wide enough to include the config that survives; (3) every gate number is provisional until the denominators are big enough — bump BENCH_QA_N and re-confirm before promoting anything. The GPU box (ai, the 4090 — make bootstrap-nli-only, then ARBORIST_NLI_DEVICE auto-detects cuda) makes this cheap: a 7-model × full-grid × ~300-record sweep is ~3.5 min. That electricity is well spent — burn it; a guessed config that ships is far more expensive than a sweep that doesn't. Heavy off-device passes (NLI sweeps, embedding backfills) run on the GPU producer, never in arborist's python+sqlite3 core (cf. the #000051 vecpack pattern).

Docs index

North-star:

• docs/seven-point-program.md — the architectural directive distilled 2026-05-01. Every new ticket / feature / prompt edit walks past this. Bench harness reports per-mode directive coverage.
• docs/warrant-substrate-cookbook.md — per-pillar map of the 18 open-licensed textbook substrates that back every claim-pack record. Ingest patterns (HTML / textbook_tex / PDF→localhost), alias discipline, cascade tuning, honest tier breakdown. Read first when extending substrate coverage.

Architecture / ongoing work:

• docs/cti-architecture.md — CTI Clause Tree Intelligence.
• docs/mesh.md — mesh wire format + group-key state machine.
• docs/benchmarks.md — orientation: harnesses, fixtures, signal floor, make targets, bench-row schema, addenda index. Read first when running a bench.
• docs/qa-modes-bench.md — bench journal (rolling addenda). Headlines + cross-references to per-ticket bench data.
• docs/bench-maxing.md — bench discipline (5pp signal floor etc).
• docs/soft-hash-channel-analysis.md — #000018 closure; analysis of soft-hash covert channel risks under M0/M1/M2 threat models.
• docs/soft-hash-channel-t3-bound.md — #000036 closed-form per-window budget bound; pairs with t3_bound_calculator.
• docs/calculator-test-patterns.md — sister discipline for testing calculator-style code (12 patterns; see #000044 for numeric drift discipline).
• docs/spec-methodology.md — #000019 specification methodology for π* canonical projections.
• docs/v7w-frontier-catalog.md — #000013 v7-W spatial-temporal frontier catalog (4 ε-frontiers).
• docs/v8-fork-score.md — #000012 ForkScore Phase 1a reference.
• docs/pi-star-composition.md — π* cross-domain composition reference (#000015 deliverable).

Tickets: docs/TICKETS.md is the authoritative index with Next ID. Closed tickets stay in place as the design log. New tickets bump Next ID atomically. Close tickets when the work lands — flip Status to closed · landed in commit <sha> (or closed · YYYY-MM-DD) in the ticket file AND in the index row, in the same commit as the implementation. An open ticket whose code already shipped is a stale map.

Default: extend an existing ticket. Don't proliferate. When follow-up work surfaces during implementation (a consumer-side fix the new code needs to actually take effect; a bench-gated tuning step; a small downstream tweak), add it to the open / most-related ticket: reopen closed → in progress if the prior closure was premature, extend the scope, list the new sub-items, keep the design log linear. fox tracks the index by skimming a small set of threads; spawning #000054.1 / #000055 / #000056 / … for every follow-up turns the index into a wall of micro-tickets he can't easily keep state on, and dilutes the design log instead of concentrating it. Split only when the new piece needs a distinct audience — specifically, when it's a self-contained design decision a Dav1d de-novo review (the external code-review thread that runs against one ticket at a time) needs to read independently. Architectural inflection points, large scope changes, or fundamental discipline questions (NLI may touch audit_mode? — #000049 split from #000048) cross that bar. "The extractor needs a small consumer-side tweak" does not. When in doubt, ask before opening a new ticket — a two-sentence "fold into #X or spawn a sibling?" is cheap; an unwanted ticket is friction. Closing on the implementation commit is good practice only when the work is actually complete end-to-end; pattern-match closure on the user-visible outcome, not on the commit.

Orientation protocol

date -u
pwd
git log --oneline -5
git status
make test
make chain-check-shards   # 0 per shard = intact
.venv/bin/arborist --shards-dir ~/.arborist/shards stats
.venv/bin/arborist --shards-dir ~/.arborist/shards analyze --gravity-top 5

Then ask fox what the mission is.